By the end of this chapter, the reader will be able to: 1. Distinguish tracking from measurement in operational terms a finance partner would accept. 2. Explain the counterfactual and why no platform dashboard can observe it. 3. Diagnose four failure modes that inflate platform-reported results. 4. Analyze a reported cost per acquisition and restate it as a cost per incremental acquisition. 5. Design a simple geo holdout or audience holdout with a stated decision rule. 6. Build a reporting standard that reconciles platform numbers to the order system of record.
\nHere is a question most brand managers answer wrongly. Your dashboard says a campaign returned four dollars of revenue for every dollar spent. How many of those sales would have happened if you had never bought a single ad?
\nMost people say \"fewer, but I couldn't tell you how many.\" That answer sounds humble. It is actually the problem, because the budget was set as if the number were known. Somebody approved a larger spend because the screen said four to one, and nobody on the call could say what the screen was counting.
\nDigital advertising is the most observable form of marketing ever built. Every impression is logged, every click is timestamped, every conversion is tied to a campaign name. That observability creates a feeling of knowledge, and the feeling is stronger than the knowledge underneath it. A stadium camera records every fan who walks through a gate. It cannot tell you which came because of the highway billboard.
\nNothing in this chapter says digital advertising does not work. Much of it works very well. The claim is narrower and more useful: the numbers most teams use to prove it works are records of activity, and records of activity cannot separate sales you caused from sales you merely sat next to. Until a team can draw that line, it is allocating money by habit and calling it analysis. The rest of this book builds on the line drawn here.
\nStart with plain definitions. Tracking records that an event occurred: an ad was shown, a person clicked, a page loaded, a purchase completed. Measurement answers a different question: how much of the outcome happened because of the ad?
\nA click is a tracked event. A sale that would not have occurred without the ad is a measured effect. Those two things overlap sometimes, and sometimes they do not overlap at all. Game film shows every play. It cannot show the score with a different defense on the field.
\nThe word for the missing piece is the counterfactual: the outcome in the world where the ad never ran. Platforms observe the world where it did. They have no view into the other one, and no amount of engineering inside the platform changes that, because the information does not exist in the data they hold. A person who was never shown the ad and bought anyway looks, to the platform, like nobody at all.
\nTracking is cheap and automatic. Measurement takes a design, a control group, and patience. Teams that confuse the two end up paying measurement prices for tracking quality.
\nHere is the thesis in the form I will point back to throughout the book. Digital is trackable, not automatically measurable. Click data describes activity. Only a controlled experiment describes cause.
\nHere is a worked example. The brand is fictional and the figures are illustrative.
\nSuppose Alder Quay Outfitters, a regional outdoor retailer, spends 13,300 dollars in a month on paid search for its own brand name. The platform reports 950 orders. Divide spend by orders and the reported cost per acquisition is 14 dollars. That looks wonderful, and the team asks to double the budget.
\nNow suppose Alder Quay pauses that brand campaign in 10 matched markets for two weeks while leaving it running in 10 others. Matching means the markets have similar size, history and seasonal pattern. When the analyst compares total orders from all sources, the paused markets lost only a fifth of the orders the platform would have credited. Everyone else typed the brand name, saw the free organic listing in the first result position, and bought anyway.
\nSo of the 950 reported orders, about 190 were incremental. The cost per incremental order is 13,300 divided by 190, which is 70 dollars. The reported figure was 14 dollars. The real figure was five times higher.
\nThe platform did not lie. Every one of the 950 orders really did follow a click on a Alder Quay ad. The report was accurate and the interpretation was wrong. The dashboard is correct about what it counts and silent about what you needed to know.
\nEvery platform lets you set an attribution window, which is the period after an ad interaction during which a conversion is credited to that ad. Common choices include one day, seven days or 30 days after a click, and shorter windows after a view.
\nA window measures nearness, not influence. If a person clicks an ad on Monday, ignores it, and returns through a bookmarked link on Thursday, a seven-day window credits the ad. The platform cannot know the ad played no role. It only knows the sequence of events that fit inside the rule you gave it.
\nWiden the window and reported conversions rise, even though nothing about the real world changed. Platform defaults usually lean wide, because a wider default makes the product look good in the first review. Your job is to set the window from your actual buying cycle. A shirt purchased on impulse has a window of hours or days. A kitchen remodel has a window of months. If your own order history shows most buyers convert within three days of first visit, a 30-day window is mostly counting coincidence.
\nDocument the window you chose and why. Then keep it fixed. A window that quietly changes between reporting periods makes trends meaningless.
\nEach platform grades its own homework. Search credits itself under one rule, social under another, video under a third, and email under a fourth. None of them knows the others exist. A single purchase can be claimed by every system that touched the buyer in the preceding window.
\nSuppose your order system recorded 400 orders last month. Search claims 270. Social claims 220. Video claims 150. Email claims 120. Add those together and you get 760 claimed orders against 400 real ones. The gap of 360 is not an error in anyone's code. It is four honest systems each reporting a share of the same sales as their own.
\nThe gap is also a free diagnostic. Divide claimed orders by actual orders and you get an overlap ratio, here 1.9. Track it monthly. A rising ratio means more channels are claiming the same buyers, which usually means more spend is landing on people who were already converting. A falling ratio after a budget change tells you something real shifted.
\nSome customers arrive with no help from anyone. They know your brand, they type your name, they walk into your store, they open your app. That flow is your organic baseline, and it is the quiet engine of most businesses.
\nPaid media often stands in front of that flow and collects credit for it. Branded search is the classic case. When someone searches for your company by name, a paid ad appears above the free listing. The person clicks whichever is first. The platform logs a paid conversion. The organic channel logs nothing. Over a year the organic line in your report shrinks and the paid line grows, and it looks like paid is winning share when it is really relabeling the same visits.
\nRetargeting works the same way: ads follow people who already added items to a cart, and many were headed to checkout anyway.
\nThe correction is a holdout, covered in a later section. Before then, one caution about instinct: this does not mean branded search is worthless. It protects against competitors bidding on your name, and in some markets that matters a great deal. The point is that its value should be measured as protection and defended on those grounds, not reported as a growth engine because the dashboard says so.
\nA view-through conversion credits an ad that a person saw but never clicked. An engagement-based conversion credits any interaction at all, including a short video play that began automatically as someone scrolled.
\nBoth metrics have a legitimate purpose. People do see a video and later search the brand name, and a click-only report would miss that. The trouble is that a person actively shopping sees dozens of ads in a week. The platform that served the last impression before purchase can claim the sale, whether or not the impression changed anything.
\nPicture a ballpark video board showing a ticket ad to a fan already in her seat. Counting that as a conversion is the same error.
\nTreat view-through numbers as a hypothesis to test rather than revenue to bank. A channel whose results lean heavily on view-through credit should be tested first.
\nFinance teams are not hostile to marketing. They are trained to ask where each number comes from and whether it ties to the books. Marketing dashboards fail that test in three predictable ways.
\nFirst, the totals do not reconcile. When channel reports sum to more orders than the system of record shows, as in the 760 versus 400 example, the finance partner correctly concludes that the dashboard is not a ledger. Once that conclusion forms, every later number inherits the doubt.
\nSecond, the definitions move. Windows change, conversion events are added, attribution rules are updated by the platform without notice. A trend line that crosses one of those changes is two different measurements stitched together.
\nThird, no one can state the counterfactual. The finance lead asks what happens if the budget is cut by a third. The honest answer from a dashboard is \"unknown,\" and a team that cannot answer that question has no claim on the next quarter's money.
\nReconcile platform orders to the order system monthly. Freeze definitions and log every change. Then bring one experiment to the budget meeting. A finance partner will accept an experiment with a stated interval of uncertainty far faster than a dashboard with a confident number, because the experiment shows its work. Legal, tax and accounting treatment of any of these figures is a matter for your counsel and accountants. What I describe here is operating practice.
\nA holdout is an experiment that withholds advertising from a comparable group. The principle is simple: two groups that are alike in every way except one, and a clear comparison of their outcomes. Whatever differs afterward is the effect of the one thing you changed.
\n| Design | How it works | Main weakness | | --- | --- | --- | | Geo holdout | Pause ads in matched markets, keep them on in others, compare total sales | Needs enough markets and similar sales patterns | | Audience holdout | Suppress ads for a random slice of a target list, compare conversion rates | Needs a platform that supports true suppression | | Time-based holdout | Pause for a matched period and compare to adjacent periods | External events can distort the comparison |
\nRandomness is what gives the audience design its strength. If the platform assigns people to the ad group or the control group by chance, nothing about the people themselves can explain a difference in results. Geo tests rely on matching instead, which is weaker, so pair them with a pre-period check: confirm the paired markets moved together before the test began.
\nWhichever design you choose, write the decision rule before the test starts. For instance: if incremental cost per order comes in under 40 dollars, hold spend; if it lands between 40 and 60 dollars, trim by a quarter; if it exceeds 60 dollars, move half the budget to a channel that has not been tested. Pre-committing prevents the quiet temptation to reinterpret an unwelcome result.
\nWhat the experiment gives you is a number a dashboard never will: the sales that disappeared when the ad did. You also get a range around that number, and the range matters. Reporting \"somewhere between 150 and 230 incremental orders\" is more credible than reporting \"190,\" because it shows you understand what you do not know. Run the analysis with open statistical tools, report the interval, and present it beside the platform number so the room sees both.
\nExperiments are not free. Pausing a channel can cost sales if the channel really is working, so test a small share of markets first.
\nA measurement standard has to be small enough that a busy team will keep it. Mine has four parts.
\nFirst, a first-party record. Keep your own event data on your own server so the order system, not the ad platform, remains the source of truth. Platform numbers then become a cross-reference.
\nSecond, a monthly reconciliation of claimed to actual orders, with the overlap ratio logged.
\nThird, a quarterly test on the largest line in the budget. Over a year you will have four tested channels, which is more than most teams ever have.
\nFourth, a reporting format that shows reported and incremental figures side by side, with the interval on the incremental one. The gap between the two columns becomes your most honest indicator of waste.
\nWhere audiences include children or sensitive categories, set a conservative default and get legal review of any data collection that touches minors.
\nTracking: the recording of ad events and user actions, such as impressions, clicks and purchases, without any claim about what caused them. Counterfactual: the outcome that would have occurred if the advertising had not run, which no platform can observe directly. Incrementality: the portion of an outcome that exists because of the advertising and would not have occurred otherwise. Attribution window: the period after an ad interaction during which a conversion is credited to that ad. Holdout: a deliberately unexposed group, whether a set of markets or a slice of an audience, used as the comparison for measuring incrementality.
\n1. Treating every platform-reported conversion as caused by the ad, when the platform can only report what followed the ad. 2. Approving budget increases from reported return on ad spend alone, without a control group to show what the revenue would have been anyway. 3. Letting platform default windows and conversion definitions stand without choosing them against the real buying cycle. 4. Presenting platform totals to finance without first reconciling them to the order system of record. 5. Running a pause test without a written decision rule, then explaining away any result that disagrees with the original plan.
\n1. Keep a first-party event record that you control, so platform reports serve as a cross-check and never as the only source of truth. 2. Test the largest channel first, because that is where an error costs the most and where a correction pays back fastest. 3. Report demand creation and demand capture separately, so branded search and retargeting are not credited for growth that other channels began. 4. Write the decision rule, the test length and the matched groups down before launch and share them with finance. 5. Report incremental cost per acquisition with its interval beside the reported cost per acquisition, in the same table every month.
\n1. Pausing branded search for two weeks in matched markets is among the cheapest tests available, and it usually settles an argument that has run for years. 2. Do the pre-period check on geo tests: if your matched markets did not move together before the test, the result will not hold up under questioning. 3. Run tests outside promotional peaks, because a holiday sale or a product launch swamps the signal you are trying to isolate. 4. Keep a change log for every window, conversion definition and platform setting, so a trend break can be traced to its cause. 5. Ask the platform representative which conversions in a report are view-through, and what share of the total they represent, before you accept the headline number.
\nEvery tool here is free to use and can be run on your own infrastructure. Matomo (GPL) and Umami (MIT) provide first-party web analytics that keep a record independent of any ad platform. CausalImpact (Apache 2.0) estimates the effect of an intervention from a time series and a set of comparison series, which suits geo and time-based holdouts. statsmodels (BSD) supplies the regression and statistical testing routines for analyzing audience tests and building intervals. Metabase (AGPL) displays platform-reported and incremental results side by side for finance. No open source option replaces the experiment itself; the design work is still yours.
\nDigital advertising is trackable, and tracking is not measurement. Platforms record activity that follows an ad, but they cannot observe the counterfactual, so their reports mix caused sales with captured ones. Four failure modes widen the gap: attribution windows that reward proximity, double counting across platforms that each grade their own work, paid placements that sit in front of organic demand, and view-through or engagement credit that rewards an ad for being nearby. Finance teams distrust dashboards because totals fail to reconcile, definitions drift and nobody can state what happens if spend changes. A holdout, whether by geography, audience or time, produces the answer a dashboard cannot, and a small standard of first-party records, monthly reconciliation, quarterly tests and side-by-side reporting makes that answer routine.
\n1. Tracking records that something happened; measurement estimates what the ad caused, and the two should never share a column heading. 2. Platforms cannot see the counterfactual, so a high reported return can sit on top of a very low incremental one. 3. An overlap ratio of claimed orders to actual orders gives a monthly warning on double counting. 4. Attribution windows should come from your real buying cycle, be documented, and stay fixed across periods. 5. A holdout with a pre-written decision rule and a reported interval is the form of evidence finance will accept.
\nThis quarter, pick your largest channel, pause it in a matched set of markets for two weeks, and put the incremental cost per order next to the platform's number in front of finance.
\n1. Brodersen, K.H., Gallusser, F., Koehler, J., Remy, N., Scott, S.L. Inferring causal impact using Bayesian structural time-series models. Annals of Applied Statistics. Published methodology behind the CausalImpact software. 2. Seabold, S. and Perktold, J. statsmodels: Econometric and statistical modeling with Python. Proceedings of the Python in Science Conference. Reference for the statsmodels library. 3. Matomo. Matomo analytics documentation, including data ownership and conversion tracking. Open source web analytics reference. 4. Media Rating Council. Viewable Impression Measurement Guidelines. Standards for impression validity and viewability. 5. Media Rating Council. Outcomes and Data Quality Standard. Industry reference on measurement quality for outcomes data.
"},{"id":"chapter-2-how-money-moves-through-the-online-ad-economy","title":"Chapter 2: How Money Moves Through the Online Ad Economy","html":"By the end of this chapter, the reader will be able to: 1. Trace a single budget from the advertiser's bank account to the publisher's revenue line and name every hop where a fee is taken. 2. Distinguish the walled garden economy from the open web economy and explain why the same budget behaves differently in each. 3. Calculate the publisher share of a campaign from a set of fee quotes using simple arithmetic. 4. Interpret ads.txt, sellers.json and the supply chain object to confirm who is authorized to sell an impression. 5. Evaluate two competing buying paths by cost per customer instead of cost per thousand impressions. 6. Design a short supply path review that a media buyer can run every quarter.
\nMost brand managers have never opened a seller file, and most of them sleep fine. Somebody will say: who cares where the money goes? The dashboard shows what I paid and what I got. If cost per customer looks fine, the plumbing is somebody else's problem.
\nIt is a fair objection. Plenty of good operators run profitable campaigns without ever reading a seller file.
\nHere is where the objection breaks. The dashboard reports what the platform chose to count, and the platform is paid out of the same budget it reports on. A result that looks fine tells you nothing about the result you would have gotten on a cleaner path. Two campaigns can post the same cost per customer while one leaves a third of its budget in the pipes, and its owner finds out only when a competitor undercuts it.
\nThink of a stadium concession. The fan pays eight dollars for a drink. The team does not receive eight dollars. A distributor, a stadium operator and a payment processor each took something. The owner who knows the split can renegotiate it. The owner who does not sells a lot of drinks and wonders why the margin is thin.
\nAdvertising money moves the same way, with more hops and less disclosure. The budgets that disappear are rarely stolen. They are absorbed, slowly and legally, by layers nobody asked about. This chapter follows a budget through those layers so that you can ask. Everything here is operational guidance. Contract language and disclosure obligations belong with your counsel.
\nThe dashboard draws a short version: you fund a campaign, an auction runs, an ad appears, a report arrives. It is clean and mostly fictional as a map of where money went.
\nThe real path has stations. At the front sits the advertiser, who owns the objective, the budget and the creative. Next may come an agency or an in-house trading desk, which plans, buys and reports. Then the buying platform, usually a demand-side platform (software that lets a buyer bid on impressions across many sites). Then an exchange, which is the marketplace where bids meet inventory. Then a supply-side platform (software publishers use to offer their ad space to many buyers at once). Finally the publisher, who owns the page and the audience.
\nAround that spine sit side vendors: verification companies that measure viewability and screen out invalid traffic, data providers that sell audience segments, and resellers that buy inventory and resell it. Each is paid from the same pool.
\nNot every campaign touches every station. A search ad bought directly in a search platform skips most of them. A programmatic display campaign can touch all of them. You should know which road yours takes.
\nArithmetic makes this concrete. The figures below are illustrative and round. They are not benchmarks, and your actual quotes will differ. Suppose a brand commits 10,000 dollars to a programmatic display campaign managed by an agency.
\nQuote each fee as a share of the original 10,000 dollars, because that is the only way to compare them cleanly. Suppose the agency charges 10 percent for planning and management. The buying platform takes 12 percent. A data provider takes five percent for audience targeting. Verification takes three percent. Between the exchange and the supply-side platform, suppose another 20 percent is taken in combined fees and resale margins.
\nAdd it up: 50 percent in fees. The remaining half, five thousand dollars, reaches the publisher. The publisher share is 50 percent.
\nWatch what the dashboard says. If the publisher cleared those impressions at five dollars per thousand, the publisher revenue bought one million impressions, and the brand paid an effective 10 dollars per thousand. That number looks ordinary. It hides the fact that half the budget never touched a publisher.
\nNone of this proves misconduct. Every one of those fees may buy something real. The question is not whether a fee exists. It is whether you can see it, chose it, and need what it buys.
\nBrand managers hear the phrase working media constantly. It describes the part of a budget that pays for the placement of the ad, as opposed to the part consumed by fees.
\nThe trouble is that the phrase has no single definition. Some buyers count only the publisher's revenue. Some count everything paid to the exchange and the platform, treating verification and data as part of the media. Some count agency fees as non-working, and some bury them inside a blended rate. An agency that reports an 80 percent working media share and an agency that reports 50 percent may be describing the same path with different definitions.
\nDisclosure is thin for ordinary reasons. Fees are negotiated one buyer at a time and often bundled into a single rate. Some intermediaries buy inventory on their own account and resell it at a markup, which makes the fee invisible because it never appears as a line item. That is not improper on its face. What matters is whether your contract says it is happening.
\nSo ask for the definition in writing before you ask for the number. Does the report show publisher revenue, or a spend figure that includes fees? A vendor that answers cleanly is worth keeping. A vendor that cannot answer is telling you something too.
\nSpending in a large social or search platform is a different economy from spending on the open web.
\nA walled garden is a platform that owns the audience, the auction, the measurement and the reporting. You pay one party. That party sells you the impression, decides who wins, counts the result and shows you the report. The fee is built into the auction price. You trade transparency for simplicity and scale. There is no path to inspect, only the outcome the platform chooses to publish.
\nThe open web works the other way. Many independent parties sit between you and a publisher. Standards exist to expose the path, you can choose which sellers to buy through, and you can negotiate each layer. You also carry the work of doing so, and the risk of a chain nobody checked.
\nNeither economy is better. They fail differently. A walled garden fails by being opaque: you cannot tell whether a good result came from your creative, your audience or the platform's own tuning. The open web fails by being crowded: a good-looking rate can sit on top of a long chain that left little for the publisher. Treat them as separate budget lines. In the garden, ask what you can verify outside the platform. On the open web, ask which path each dollar took.
\nMost buyers compare vendors by price. Who offers the lowest cost per thousand impressions? Who quotes the lowest platform fee? It feels disciplined. The problem is that a rate describes one hop, and your outcome depends on the whole path. Consider two paths for the same 10,000 dollar budget, again with illustrative figures.
\nPath A is the long, cheap-looking path. It runs through several resellers, and the fee stack leaves half the budget for publishers. It delivers one million impressions, a rate of 10 dollars per thousand to the buyer. Suppose a good share of that inventory sits on thin sites and low-attention placements, and the campaign produces 125 customers. Cost per customer: 80 dollars.
\nPath B is a direct deal with a known publisher through a private marketplace, which is an invitation-only auction where the publisher chooses who may bid. Fewer hops mean fewer fees. Suppose the agency takes 10 percent, the platform 12 percent, verification three percent, and the seller-side platform 11 percent. That is 36 percent in fees and 64 percent to the publisher. At a clearing price of eight dollars per thousand, that buys 800,000 impressions, an effective 12.50 dollars per thousand to the buyer. Suppose the better placements and a real audience produce 250 customers. Cost per customer: 40 dollars.
\nPath B looks 25 percent more expensive per thousand impressions and costs half as much per customer. Rate shopping would have picked Path A every time.
\nThe cheapest impression is rarely the cheapest customer, because the dollars wasted in the supply chain are the same dollars that would have bought attention. Path control beats rate shopping. You win by deciding which road the money takes, not by haggling over the toll at one booth.
\nThose conversion figures are invented to show a mechanism. In a real account a direct path will not always win, and it will often reach fewer people. Run a test before you believe any story, including mine.
\nThe industry built three public records to let buyers check the path. They are imperfect and underused, which is exactly why they are worth learning.
\nThe first is ads.txt, a plain text file a publisher posts at the root of its site. It lists the companies authorized to sell that publisher's inventory. Each line names the seller, the account ID, and whether the relationship is direct or reseller (the seller passes along inventory it does not own). Anyone can fetch it. If a seller offers a site's inventory and does not appear in that site's file, you have found either an error or a problem.
\nThe second is sellers.json, a file posted by a supply-side platform or exchange. It lists the sellers on that platform, whether each owns the inventory or is an intermediary, and often the business name. Where ads.txt says who the publisher authorized, sellers.json says who the platform says they are.
\nThe third is the supply chain object, usually shortened to schain. It rides inside the bid request, the message that offers an impression to buyers. Each party that handles the request adds a node naming itself. By the time the request reaches you, the object shows the sequence of parties that touched it, and a flag says whether the chain is complete.
\nUsed together, the three make a simple check. Take the seller on the bid request, find it in the publisher's ads.txt, find its account in the platform's sellers.json, and read the schain nodes to see whether the path matches the story. A mismatch is a reason to pause that route until someone explains it.
\nTwo limits apply. These files are self-reported, so a bad actor can post false data, and a clean record is evidence rather than proof. And many requests arrive with incomplete or missing chains. Treat a missing chain as a risk signal, then decide how strict to be based on how much traffic you can afford to lose.
\nA fee is only a problem when it buys nothing you value. The agency sells judgment and labor. The buying platform sells bidding software. The exchange runs the auction, and the supply-side platform gives the publisher access to many buyers. Verification vendors sell assurance that the impression was viewable and human. Data providers sell the claim that a user belongs to a segment.
\nEach is a real service, and each can be overpaid or duplicated. Buy an audience segment on top of a platform that already targets well and you may pay twice for one signal.
\nBefore a budget moves, ask every vendor four things. What is your fee, and what is it a percent of? Is any part of the price a markup on inventory you bought for your own account? Which exchanges and sellers do you route through, and can I restrict the list? Can you show publisher revenue as a share of my spend? A platform that will not discuss its fee is a candidate for a small test, not for most of your money.
\nSome companies operate more than one layer. One firm can run a buying platform, an exchange and a selling platform at once. That can help, since systems built together may be faster and the path shorter. It can also create a conflict. When one company stands on both sides of an auction, the rules decide whose interest wins, and you are rarely in the room when the rules are written.
\nYou do not need to condemn integration. You need to know who owns what, and whether routing favors the platform's own exchange. All of this fits into a path review you can run each quarter in an afternoon. 1. Pull the list of sellers that received your spend, from your platform's reporting. 2. Check each against the relevant ads.txt and sellers.json files. 3. Calculate publisher share for your largest campaigns using vendor quotes. 4. Rank sellers by cost per customer, not cost per thousand. 5. Cut the lowest performers, shift money to the strongest direct paths, and record the changes.
\nThe review will not be perfect. It beats trusting a dashboard paid from the budget it describes.
\nWorking media: The portion of a budget that pays for the placement of the ad itself, as distinct from fees. Definitions vary by vendor, so always ask for the one in use.
\nSupply path optimization: The practice of choosing the shortest, most direct and most transparent routes to publishers, rather than buying through every available intermediary.
\nPrivate marketplace: An invitation-only auction in which a publisher selects the buyers who may bid, usually offering better inventory and clearer terms.
\nReseller: A seller that passes along inventory it does not own, adding a hop and usually a margin between buyer and publisher.
\nPublisher share: The fraction of an advertiser's spend that ultimately reaches the publisher as revenue after every fee and margin is taken.
\n1. Treating the dashboard spend figure as proof of what the media cost. The figure includes fees the report never itemizes, so true publisher share stays unknown. 2. Choosing between vendors on the quoted cost per thousand impressions alone. A cheap rate over a long path often yields the highest cost per customer. 3. Assuming an exchange is a neutral referee. Some exchanges are owned by the same company that runs the buying or selling platform on the other side of your bid. 4. Never opening an ads.txt file. The file is public and takes a minute to read, and skipping it means you trust the seller to describe its own authorization. 5. Accepting the phrase working media without a written definition. Two vendors can report very different shares for the identical path.
\n1. Ask each vendor for fees in writing, including what each fee is a percentage of, before you commit a budget. 2. Set a minimum seller standard for open web buying, such as a requirement that sellers appear in the publisher's ads.txt file. 3. Keep a short list of direct or private marketplace partners for your highest-value placements, and compare them to open buying on cost per customer. 4. Run the quarterly path review and keep the results in one document, so changes can be traced back to evidence. 5. Test any path change on a small slice of budget first, and read the outcome after enough customers have accumulated to mean something.
\n1. When you move from percentage fees to fixed fees as spend grows, compare the two on your actual volume, since the cheaper structure changes with scale. 2. Ask whether your agency buys inventory on its own account. If it does, ask that the arrangement be described in the contract and reported to you. 3. A chain marked incomplete is not necessarily bad, but a high volume of incomplete chains from one seller deserves a conversation. 4. Cap the number of verification and data vendors per campaign. Extra layers add fees faster than they add safety. 5. Keep walled garden and open web spend on separate lines in the plan, so that you can see which economy is helping and which is only absorbing budget.
\nThe Prebid documentation (Apache 2.0) is the most useful free reference for understanding how a bid request is built. It explains how the supply chain object is carried, which fields identify the seller, and how publishers declare their relationships. Reading it does not require running any software. Compare what it describes with the requests your own platform reports, and an unclean path becomes easy to spot. No open source product matches a full scale demand-side platform for buying. For supply path reading, though, the documentation is free, public and sufficient.
\nA budget does not travel straight from brand to audience. It moves through an agency, a buying platform, an exchange, a selling platform and a publisher, with side vendors taking fees along the way. The dashboard reports spend, not publisher share, so it cannot tell you how much of your money bought attention. Walled gardens and the open web are two economies with different failure modes: one is opaque, the other is crowded. Rate shopping compares a single hop, while outcomes depend on the whole path, which is why a cheap impression on a long chain can cost more per customer than a pricier impression on a short one. Ads.txt, sellers.json and the supply chain object give buyers a way to check the path, within honest limits. A quarterly review turns that checking into a habit.
\n1. Every hop between your budget and a publisher takes a fee, and the dashboard reports spend rather than what reached the publisher. 2. Working media has no single definition, so ask for the vendor's definition in writing before you compare numbers. 3. Walled gardens and the open web are separate economies and deserve separate lines in the plan and separate questions. 4. Compare buying paths by cost per customer, since a cheap impression on a long path is often the expensive customer. 5. Ads.txt, sellers.json and the supply chain object let you verify a path, though all three are self-reported and a clean record is evidence, not proof.
\nThis week, ask your largest vendor for its fee structure and its written definition of working media, then calculate your publisher share on one campaign. If you cannot, you have found your first negotiating lever.
\n1. Interactive Advertising Bureau Tech Lab. ads.txt specification, Authorized Digital Sellers standard, current version. 2. Interactive Advertising Bureau Tech Lab. sellers.json specification, seller identity disclosure standard for exchanges and supply-side platforms, current version. 3. Interactive Advertising Bureau Tech Lab. OpenRTB specification and the SupplyChain object extension, current version. 4. Prebid. Prebid documentation, sections on the supply chain object and on seller identification (Apache 2.0). 5. Organisation for Economic Co-operation and Development, Competition Committee. Competition in Digital Advertising Markets, roundtable background note.
"},{"id":"chapter-3-auction-mechanics-first-price-second-price-and-floors","title":"Chapter 3: Auction Mechanics: First Price, Second Price, and Floors","html":"By the end of this chapter, the reader will be able to: 1. Explain how a first-price auction and a second-price auction each set the price a winner pays. 2. Calculate the outcome of a small auction under both rules using illustrative bids. 3. Apply bid shading logic to a first-price buy and recognize when the winner's curse is eating margin. 4. Analyze how floors and reserve prices change who enters an auction and what the clearing price becomes. 5. Compare rank-based auctions in walled gardens with pure bid-based exchange auctions. 6. Evaluate which parts of a media cost a buyer controls and which parts the auction decides for them.
\nPicture a regional boat dealership I will call Harbor Line Marine (an illustrative composite, not a real brand). Its media manager read that the exchange her team buys through had been a second-price auction for years, so she did what the textbooks said. She bid her true value for a boat-show audience, 14 dollars per thousand impressions, on the theory that she would only ever pay a hair above the next bidder. It was a sensible decision. It was also wrong.
\nBetween her planning meeting and her campaign launch, the exchange had moved to first price. Every winning impression now cost her the full 14 dollars. Her reports showed delivery on schedule and the average cost nearly double what she had forecast. Nothing was broken. She had simply bid under a rule that no longer existed.
\nHere is the assumption this chapter challenges: that the price you see on a report is a price you set. It is not. It is the output of a rule, and the rule belongs to someone else. A maximum bid is a statement of intent. The auction decides what it means.
\nThat matters because much digital media is sold through some kind of auction. The mechanics decide whether your bid strategy is rational, whether an intermediary is pocketing the gap between what you bid and what you pay, and whether a cheap CPM (cost per thousand impressions) is a bargain or a warning sign. Ask one question before any bid: under what rule will this be priced?
\nIn a first-price auction, the highest bidder wins and pays exactly what it bid. In a second-price auction, the highest bidder wins and pays a price based on the runner-up's bid, usually a small increment above it. That single difference changes how a rational buyer should behave.
\nUnder second price, honesty pays. If your true value is 10 dollars, bid 10. If the next bid is six, you pay about six. If it is 11, you lose, rightly. Your bid never sets your price, so shading it down only costs you wins you wanted.
\nUnder first price, honesty is expensive. Bid 10 dollars when your value is 10 dollars and you win with zero surplus, no matter how far ahead of the field you were. Think of a silent charity auction for a signed jersey. If you privately value it at 500 dollars and you write 500 on the sheet, you will be proud of winning and have gained nothing. A smart bidder writes something lower, guessing what it takes to beat the room.
\nNeither format is morally better. Sellers move between them for reasons of revenue, transparency and fraud control, often with little notice. Confirm the rule at the start of every campaign and whenever a platform announces a change.
\nNumbers make this concrete. Suppose four advertisers bid on one impression in an open exchange. The bids, in dollars per thousand, are 12, nine, seven and four. The seller has set no floor. These figures are illustrative.
\nUnder second price, Bidder A wins with the 12 dollar bid and pays about nine dollars (plus a small increment, which we will ignore). The seller collects nine dollars. Bidder A keeps three dollars of surplus compared with its stated value.
\nUnder first price, Bidder A wins with the same 12 dollar bid and pays 12 dollars. The seller collects 12 dollars. Bidder A keeps nothing, assuming 12 was its true value.
\nNow add intelligence. Suppose Bidder A has seen enough past auctions to estimate that the winning price in this inventory usually lands near 10 dollars. Under first price it might bid 10.50 instead. If the field stays where it was, A still wins, pays 10.50, and keeps a dollar and 50 cents of surplus. If a competitor bid 11, A loses that impression. That is the whole trade in one example: lower bids raise your margin on wins and reduce how often you win.
\nNotice the seller's side: nine dollars under second price, 12 under naive first price, 10.50 once bidders shaded. The gap between naive and shaded is where informed buyers beat uninformed ones.
\nBid shading is the practice of bidding below your true value in a first-price auction to keep some of the gap for yourself. Every serious programmatic bidder does it, either by hand or through software that estimates a clearing price from past results. It is not cheating. It is the rational response to the rule.
\nThe question is how much. Shade too little and you pay more than you needed to. Shade too much and your win rate collapses, and your campaign underdelivers. A reasonable method starts with your own logs. Pull the winning bids and your losing bids for a placement over a few weeks, find the price at which you win about half the time, and test bids a little above and below that point. Watch the combination of win rate and cost. Pick the bid where an extra cent buys little additional delivery.
\nThen there is the winner's curse. When many buyers estimate the value of the same impression with some error, the one who wins tends to be the one who overestimated the most. Winning is itself a piece of bad news. If the true value is five dollars, some bidders will guess three and some seven. The winner is likely the one who guessed seven.
\nThe curse shows up in real campaigns as the bid that \"wins everything.\" If a new line item is delivering far faster than you forecast, do not celebrate. Check whether you are the highest bidder on inventory that everyone else has judged to be worth less. The usual causes are a bid set too high for the audience quality, an audience list that other buyers have already screened out, or a seller reselling the same impression through several paths. Slow your bids and compare outcomes on winners against the ones you lost.
\nA reserve price is the minimum a seller will accept. A floor is the same idea applied to a particular piece of inventory, such as a placement, a deal or a user segment. Bids below the floor are rejected, no matter how many of them there are.
\nFloors matter for three reasons. First, they remove the cheap bottom of the auction. Suppose a seller sets a floor of five dollars on a premium video placement. Our bidder at four dollars never enters. Second, in a second-price auction, a floor can become the price. If the highest bid is 12 and the second bid is four, a five dollar floor means the winner pays five instead of about four. Third, in a first-price auction a floor changes bidder behavior, because every bidder now knows the lowest price that can possibly clear.
\nSellers also set floors dynamically, by user, device or time of day, and a buyer usually cannot see them. You bid six dollars and win nothing, and cannot tell whether the competition outbid you or the floor was seven.
\nTo find a floor, raise the bid in a controlled test. If delivery jumps once you cross a threshold, you have found something like one. In private deals, negotiate the floor as seriously as the volume, and write the answer into the deal documents.
\nThis is operational guidance only. Contract terms on floors, guaranteed delivery and make-goods should be reviewed by counsel.
\nSearch engines and social platforms (often called walled gardens, because the platform controls the inventory, the data and the auction) do not simply award the top slot to the highest bid. They rank ads by combining your bid with an estimate of how likely the user is to take the action you want, plus a measure of ad and landing page quality. The exact formulas are proprietary and change, but the structure is public enough to reason about.
\nThe simplest way to picture it: effective score equals bid multiplied by estimated action rate, adjusted for quality. An advertiser who bids two dollars per click with a high estimated click rate can outrank one who bids three dollars with a low one.
\nHere is an illustrative example. Advertiser One bids four dollars per click and the platform estimates a two percent click rate. Advertiser Two bids three dollars and the platform estimates a three percent click rate. Per thousand impressions, two percent is 20 clicks, times four dollars is 80 dollars for Advertiser One. Advertiser Two gets 30 clicks times three dollars, 90 dollars. Advertiser Two ranks higher despite the lower bid.
\nThe price is also not simply the bid. In many of these auctions the winner pays just enough to hold its position against the next-ranked competitor, which means the price depends on the runner-up's bid and quality. A better ad lowers what you pay for the same position. This is why the creative and landing page work in later chapters is not a branding exercise. It is a bid substitute.
\nOne trap: the estimated action rate is a prediction, not a fact. If your tracking is poor or conversion events are thin, predictions get worse, and that shows up as a higher price.
\nA budget is not a number the auction knows about. It is a constraint the buying system enforces by choosing which auctions to enter. That choice changes your clearing price, often more than the bid does.
\nPacing is how a system spreads a daily budget across the day. Even pacing enters a share of auctions at each moment. Accelerated pacing spends as fast as eligible auctions allow, so a daily budget of 900 dollars may be gone by mid-morning, spent on whatever came first at whatever price.
\nA low budget with a high bid wins early auctions at the top of your range and then stops. A high budget with a low bid wins less often but keeps entering. The same spend can reach very different audiences.
\nCompetition adds pressure. At a major holiday or a playoff run, many advertisers raise budgets at once, and clearing prices rise though nothing about your campaign changed. The room got crowded.
\nWatch spend and cost by hour for a week. If you hit the cap before your best audience is online, change the pacing or move budget to a later window.
\nA CPM looks like a price. Reports present it like one, and rate cards print it like one. In an auction it is an average of outcomes. It is the total spend divided by impressions delivered, and it moves with everything above: the format, the floors, the competition, the pacing, the quality of your ad, and the mix of inventory you won.
\nSuppose two campaigns both report a six dollar CPM. One won thin late-evening app inventory. The other won dense midday premium articles. Same CPM, very different value.
\nThe reverse happens too. A CPM can rise from six dollars to nine because the buyer began winning better placements, and a manager who reads that as failure may cut the traffic producing results.
\nTreat CPM as a symptom to be explained rather than a target to be hit. When it moves, ask what changed in the mix: the placements, the audience, the time of day, the format, the floor, the pace. Then compare cost per useful outcome, not cost per thousand views. A low CPM with weak outcomes is expensive. A high CPM with strong outcomes may be cheap.
\nStart with what is out of your hands. You cannot control the auction format. You cannot see competitors' bids. You cannot see most floors. You cannot set the platform's quality formula, and you cannot stop a competitor from raising a budget during your best week.
\nWhat you can control is more than people think. You can control the bid and the logic behind it, including how much you shade in a first-price environment. You can control budgets and pacing. You can control the creative and landing page quality that feed rank-based auctions. You can control which placements, deals and audiences you enter at all, which is often the largest lever. You can control measurement: whether you log win rates, clearing prices and delivery by hour so that your next decision rests on evidence.
\nYou can also control the questions you ask. Ask each seller for the auction format and floor rules. Ask agencies whether you are billed at the clearing price or a marked up rate. A vendor that cannot answer in plain language is telling you something. You do not set the price, but you set the conditions under which a price is possible.
\nFirst-price auction: An auction in which the winning bidder pays exactly the amount it bid, which makes bid shading a rational strategy.
\nSecond-price auction: An auction in which the winner pays a price set by the next-highest bid, usually plus a small increment, so bidding true value is safe.
\nBid shading: Lowering a bid below your true value in a first-price auction to keep surplus while still winning often enough.
\nFloor price: The minimum bid a seller will accept for an impression, placement or deal, below which no bid can clear.
\nPacing: The method a buying system uses to spread a budget across time, which affects which auctions you enter and what you pay.
\n1. Bidding true value in a first-price auction, which guarantees zero surplus on every win and often means paying far more than the competition required. 2. Assuming the format never changes, then discovering months later that a seller moved from second price to first price and that last quarter's bidding rules no longer apply. 3. Treating a high win rate as proof of efficiency, when it often signals a bid set too high and the early signs of the winner's curse. 4. Reading a CPM as a price list and cutting any placement whose average cost rose, without checking whether the mix or the outcomes improved. 5. Diagnosing a failed bid without considering floors, so that a placement is blamed for weak competition when the seller's minimum was simply above the bid.
\n1. Confirm the auction format in writing for every seller and platform before launch, and recheck it whenever the platform announces a product or policy update. 2. Keep your own log of bids, wins, losses and clearing prices by placement so shading decisions rest on your data rather than on a vendor's summary. 3. Test bid levels in small controlled steps and stop when an additional cent of bid buys little additional delivery. 4. Ask for floor terms, floor update frequency and any dynamic floor rules in private deals, and record the answers in the deal documents. 5. Judge media on cost per useful outcome and watch CPM only as an explanatory signal for changes in the inventory mix.
\n1. When a line item delivers far faster than forecast, slow it and inspect the winners before you celebrate, because fast delivery is a common sign of overpaying. 2. Run an hourly spend and cost report for a full week at the start of any new campaign, since pacing problems hide inside daily totals. 3. Before big seasonal events, assume clearing prices will rise and plan budget and bid ceilings in advance, not in the middle of the rush. 4. In rank-based platforms, fix landing page speed and message match before you raise bids, because quality improvements often lower your cost for the same position. 5. Ask agencies whether your invoice reflects the clearing price or a fixed rate, and ask them to describe in plain words how any difference is handled.
\nPrebid Server (Apache 2.0) is an open source server that runs auctions between demand partners on behalf of a publisher. A brand manager does not need to deploy it, but a technical teammate can read its documentation and code to see how bid requests are structured, how bids are compared, how floors are applied and how a winner is chosen. That view of exchange-side behavior makes vendor claims easier to test. Be plain about the limit: Prebid Server shows how auctions work from the seller side, and it is not a replacement for a full-scale demand-side platform, for which no open source option reaches parity today.
\nDigital ad prices are produced by auction rules, not by buyer intent. First price makes the winner pay its bid and rewards bid shading, while second price sets the price from the runner-up and makes honest bidding safe. Floors remove low bids and can lift the clearing price, and walled gardens rank ads by combining bids with predicted action rates and quality. Budgets and pacing change which auctions you enter, which means they change your price as much as your bid does. A CPM is the average of those outcomes and should be explained rather than chased. Buyers control their bids, creative, placements and measurement, and they should ask sellers plain questions about format and floors before spending.
\n1. The rule of the auction, not the bid alone, decides what a winning impression costs, so confirm the rule first. 2. Bid shading is the rational response in first-price auctions, and the work is in deciding how much to shade using your own logs. 3. Floors can lift prices and silently block bids, so test for them when a placement underdelivers. 4. In rank-based platforms, ad and landing page quality work like a discount on your bid. 5. A CPM is an average outcome shaped by mix, pacing and competition, so judge media by cost per useful outcome.
\nIf you have not asked every seller in writing what the auction format is and what floors apply, do that before the campaign launches. Then log your bids, wins and clearing prices for two weeks so you shade from evidence instead of guesswork.
\n1. Prebid.org. Prebid Server documentation, describing auction handling and bid selection in the open source server. 2. Google Ads Help. About Ad Rank and how the Google Ads auction works, describing the use of bid, quality and context in ranking. 3. Krishna, V. Auction Theory. Academic textbook covering first-price and second-price formats, reserve prices and the winner's curse. 4. Edelman, B., Ostrovsky, M. and Schwarz, M. Internet Advertising and the Generalized Second-Price Auction. Journal paper on search ad auction design. 5. IAB Tech Lab. OpenRTB specification, describing bid request and bid response fields, including floor price signaling.
"},{"id":"chapter-4-the-ad-stack-supply-demand-exchanges-and-verification","title":"Chapter 4: The Ad Stack: Supply, Demand, Exchanges, and Verification","html":"By the end of this chapter, the reader will be able to: 1. Name the six layers of the advertising technology stack and describe what each layer does to a single impression. 2. Explain why each layer's financial incentive differs from the advertiser's incentive. 3. Identify what OpenRTB, ads.txt, sellers.json, the SupplyChain object, VAST, and the Open Measurement SDK each prove and what each leaves unproven. 4. Calculate the share of a media dollar that reaches the publisher in a worked fee-map example. 5. Evaluate whether a verification report supports a claim a finance partner would accept. 6. Decide which parts of the stack an open source tool can replace and which parts it cannot.
\nPicture a Monday review. The dashboard says the campaign delivered four million impressions, a viewability rate the team likes, and a cost per thousand impressions well under plan. Then the finance partner asks a plain question: how much of the money we wired out ended up paying for a page a person actually looked at? Silence. Everyone in the room knows the dashboard cannot answer that, because the dashboard was built by one layer of the stack to report on that layer's own work.
\nThat is the contradiction this chapter starts from. Advertising technology produces more reporting than any other part of marketing, and almost none of it is the kind of evidence an auditor would sign. A dashboard counts what the system that sold the impression says happened. A finance partner wants to know what was paid, to whom, and what came back.
\nMany brand managers assume the stack is a pipe. Money goes in one end, attention comes out the other, and the pipe's only job is to carry it. It is wrong. The stack is a chain of separate companies, each paid for its own step, each producing its own report, and each rewarded when more money flows through it rather than when your result improves. Nobody is lying to you. Each is measuring its own lane.
\nThis chapter walks one impression through the stack, names the standards that make parts of it checkable, and tells you plainly where open source tools help and where they do not. Operational guidance only; counsel decides contract questions.
\nA person opens a news page on a phone. In the time it takes the page to load, a short chain of events runs. The publisher's page asks its ad server whether it has a direct deal to fill the slot. If not, the page offers the slot to a supply-side platform (SSP), which is software that sells a publisher's inventory to many buyers at once. The SSP packages details about the page and the visitor into a bid request and sends it to one or more exchanges, which are the automated marketplaces where the auction runs. Buyers' demand-side platforms (DSPs), the software advertisers use to bid, look at the request, check it against their targeting rules, and respond with a price and an ad. The highest bid wins. The ad loads. A verification script measures whether it was viewable. A measurement tool records the click or the later purchase.
\nSix layers, then: publisher ad server, SSP, exchange, DSP, a data and identity layer that supplies audience information, and verification and measurement. All of it takes a fraction of a second.
\nHere is what matters for you. Each hop is a separate business with a separate contract, a fee, and its own idea of a successful impression. The auction is fast, but the accounting is slow, and the accounting is where your money disappears or survives.
\nThe publisher's ad server is the first stop. It holds the direct deals, the sponsorships, and the house ads, and it decides which of those gets priority over the open auction. A publisher with a healthy direct sales team fills its best slots here and sends only leftover space to programmatic buyers. That detail changes how you read quality: inventory that reaches an auction is often what the publisher could not sell directly.
\nThe SSP exists to maximize what the publisher earns per impression. It connects the publisher to many exchanges and buyers, applies floor prices (the minimum bid the publisher will accept), and takes a fee from the winning bid. Its fee grows with price and volume, so it is happy to send the same slot through every route it owns.
\nSome companies run both an SSP and an exchange. That is not a scandal, but the SSP may prefer to route your bid through its own marketplace even when a different path would clear cheaper or deliver a better page. Path control, which I covered in Chapter 2, is the buyer's answer to this, and the standards later in this chapter are the tools that make it possible.
\nThe exchange is the meeting room. It receives requests from sellers and bids from buyers, runs the auction by a published rule, and announces the winner. The rule matters more than most buyers realize. In a second-price auction, the winner pays a hair above the second-highest bid. In a first-price auction, the winner pays exactly what it bid. Most of the industry has moved toward first-price, which sounds simple until you notice that it makes bidding strategy, and the quality of the DSP's bid-shading logic, a direct driver of what you pay.
\nBid shading means the DSP deliberately bids below its computed value to avoid overpaying.
\nExchanges also hold interests. An exchange owned by a company that also runs a DSP can, in principle, write rules that favor its own buyer. An exchange attached to an SSP can favor its own seller. A brand manager who cannot answer \"who owns the marketplace my bid just entered?\" cannot judge whose interests the rules protect. Put it to your media partner in writing; a vague answer is information too.
\nA private marketplace (PMP) is an invitation-only auction where a publisher offers selected inventory to selected buyers. A programmatic guaranteed deal fixes volume and price in advance and skips the auction. Both shorten the path.
\nThe DSP is where you, or the people buying for you, spend the budget. It takes the budget, audience rules, and creative, then bids in thousands of auctions per second. The DSP's fee is commonly a percentage of media spend or a charge per thousand impressions, which means its revenue rises when your spend rises. A DSP's optimization system is built to deliver against the goal you set, but the goal you set may be clicks or impressions, and neither is the same as an incremental sale.
\nNow the plain statement. No open source demand-side platform reaches parity with the commercial ones. Not close. A functioning DSP needs live connections to dozens of exchanges, a bid engine that answers in roughly a tenth of a second, fraud filtering, identity matching, and a team maintaining all of it daily. Free code supplies none of that, and a small business assembling its own bidder would spend more on engineers than on platform fees.
\nSo for the demand side, the honest path is to buy through an experienced operator. For readers who want an in-house execution path, Sports Media and RAMMww plan and place programmatic buys through managed platforms so that you receive the reporting and fee detail without running a bidder yourself.
\nA DSP cannot target \"people who like trail running\" unless somebody supplies that list. The data layer is where audience segments come from: website behavior, app activity, purchase records, surveys, and modeled guesses. The identity layer links the person across devices and sites so a segment can follow them. In the past this mostly meant third-party cookies, small files a browser stored. Browsers have been restricting them, and the identity layer is now a patchwork of logged-in identifiers, hashed emails, and contextual signals, meaning targeting by the content of the page rather than the person reading it.
\nData providers earn money by selling segments, so their incentive is to sell many segments to many buyers. A segment sold to a hundred advertisers is no secret edge. A segment built on old behavior may describe someone who changed jobs, moved, or lost interest months ago.
\nBefore you buy any segment, ask four questions in writing. Where was the underlying data collected? How is membership decided? How often is it refreshed? How large is the segment, and how many other buyers have purchased it? A vendor that cannot answer all four is selling you a label, not an audience. If the paid segment does not beat a contextual buy of equal budget, the data fee is a donation.
\nIAB Tech Lab, the nonprofit technology standards body of the advertising industry, publishes the protocols that let the layers talk to each other. You need to know what each one proves, because when a partner says \"we are compliant\" you should be able to ask \"with what?\"
\nOpenRTB is the language of the auction itself: the format of a bid request and a bid response. It proves that two systems can exchange offers. It does not prove the bid was fair or that the page was good.
\nads.txt is a plain file a publisher posts on its own domain listing which companies are authorized to sell its inventory. It proves the publisher named those sellers. It does not prove the seller followed through honestly.
\nsellers.json is the matching file that each ad platform publishes, listing the accounts it represents and whether each is the owner of the inventory or an intermediary. It proves who is behind a seller identifier.
\nThe SupplyChain object travels with the bid request and lists every hop the impression passed through. It proves the path, if each party reports honestly, and it lets a buyer count the intermediaries.
\nVAST (Video Ad Serving Template) is the format that tells a video player which ad to play and how to report the events. It proves a video ad was delivered in a standard wrapper. It says nothing about whether the person watched.
\nThe Open Measurement SDK is a software kit that lets a single measurement script run inside apps and players without each vendor inserting its own code. It proves the measurement came from a consistent method.
\nUsed together, ads.txt, sellers.json, and the SupplyChain object let a buyer check a path. Check the publisher's ads.txt for the seller, check the seller's sellers.json to see whether the account is the publisher or a reseller, then read the SupplyChain object to count hops. A path with four resellers and no findable publisher account tells you something no dashboard will. These checks reduce risk; they do not remove it, since the files only record what each party chooses to declare.
\nVerification vendors sit near the end of the line. They run code on the page or in the app to judge three things: whether the ad was viewable, whether the surrounding content was safe for the brand, and whether the traffic was real people and not automated scripts, which is called invalid traffic.
\nThe Media Rating Council (MRC) publishes measurement standards and accredits measurement services for viewability and invalid traffic. A common viewability standard asks that most of the ad's pixels be on screen for a minimum time, with a longer threshold for video. Accreditation means an outside body audited the vendor's method. A vendor with no accreditation may still be accurate, but you have no outside party saying so.
\nHere is the trap. A viewability rate is a measurement under a definition. A seventy percent rate tells you that seventy percent of measured impressions met the standard. It does not tell you the other thirty percent were wasted, it does not say a human noticed the ad, and it does not say how many impressions the vendor could not measure at all.
\nMeasurement is different from verification. Verification asks whether the impression was real and seen. Measurement asks what changed because of it: a purchase, a signup, a store visit. That is why no single report could answer the finance partner at the Monday review. The honest answer comes from stitching together the invoice, the verification report, and a result you measured outside the buying platform.
\nThe cleanest way to bring your finance partner into the conversation is a fee map. It is a one-page sheet listing every company that touches your spend and what each takes. Here is an illustrative example, with round numbers, not a benchmark.
\nSuppose you spend 500 dollars of media through an agency. The agency keeps 50. The DSP takes 60 from the remainder. A data segment costs 10 more, and a verification vendor is charged at 10. The exchange and SSP together keep 70 of what remains, and an added reseller hop takes another 40. The publisher receives roughly 260 of your original 500 dollars. Your dashboard will report on the whole 500, as impressions delivered, and none of it will say 260.
\n| Layer | Illustrative take | Running balance | |---|---|---| | Agency | 50 | 450 | | DSP | 60 | 390 | | Data segment | 10 | 380 | | Verification | 10 | 370 | | Exchange and SSP | 70 | 300 | | Added reseller hop | 40 | 260 |
\nNone of these numbers describes a real company. They show a method. Fill in your own figures from invoices, then ask what each layer does that you could not do more cheaply. If the path is long, do not buy more verification to patch it. Remove an intermediary instead.
\nCutting two hops from a path often does more for the publisher's share than negotiating a lower agency rate. Review the map every quarter, because a path clean in spring may have an extra hop by autumn.
\nSupply-side platform (SSP): Software a publisher uses to offer ad space to many buyers at once, earning a fee on each winning bid. Demand-side platform (DSP): Software an advertiser or agency uses to bid on impressions across many exchanges against targeting rules and a budget. Private marketplace (PMP): An invitation-only auction in which a publisher offers selected inventory to chosen buyers, often with fewer intermediaries. Invalid traffic: Ad requests or views produced by automated scripts, hijacked devices, or other non-human sources . Bid shading: A DSP technique that lowers a bid below the computed value of an impression in a first-price auction, aiming to avoid paying more than needed.
\n1. Treating the DSP dashboard as proof of results, when it is the selling system's report on its own activity and should be checked against invoices and outside measurement. 2. Never asking who owns the exchange or SSP carrying the bids, which leaves the buyer unable to judge whose interests the auction rules serve. 3. Reading a viewability or brand safety percentage without asking which standard, which vendor method, and how many impressions went unmeasured. 4. Buying audience segments without written answers on collection method, refresh rate, and segment size, then crediting the segment for results the creative produced. 5. Adding a verification vendor or data provider on top of a long path instead of removing an intermediary, so cost rises while control stays flat.
\n1. Build a one-page fee map listing every company that touches your spend and what each takes, and refresh it each quarter from invoices. 2. Require your media partner to check ads.txt, sellers.json, and the SupplyChain object on high-spend placements, and to report the hop count in plain language. 3. Favor private marketplaces and guaranteed deals for your most important inventory, so fewer hands sit between your budget and the page. 4. Choose verification vendors whose viewability and invalid traffic methods are accredited by the Media Rating Council, and ask for the measured rate alongside the unmeasured share. 5. Compare any paid audience segment against an equal-budget contextual buy before renewing it.
\n1. Ask for line-item reporting once a quarter, since a line-item sheet shows which domains took your money. 2. When a partner says \"we are fully compliant with the standards,\" reply with \"which ones, and what does that show?\" The answer is usually shorter than the claim. 3. A low reported cost per thousand impressions is not a win until you know the share of spend that reached the publisher. 4. In a first-price auction, ask your DSP in writing whether it shades bids and how. Even a short answer tells you whether anyone is watching that variable. 5. Keep a short blocklist and a short allowlist of domains from your own review. Your own list reflects your own judgment.
\nOpen source belongs on the supply side, not the buy side. Prebid.js (Apache 2.0) is header bidding code that lets a publisher's page offer its slots to many bidders in one auction, and Prebid Server (Apache 2.0) runs that auction on a server to keep pages fast. Revive Adserver (GPL) is a full ad server for owned sites, handling direct deals, house ads, and delivery reporting. A league, venue, or publisher with an owned audience can run all three on infrastructure it controls and see every fee and every bid. As stated earlier, no open source demand-side platform matches commercial buying systems, so do not plan to build one.
\nThe advertising stack is six layers: publisher ad server, supply-side platform, exchange, demand-side platform, a data and identity layer, and verification and measurement. Each is a separate business with its own fee and its own definition of success, and none is paid for your incremental result. IAB Tech Lab standards make parts of the path checkable, but each proves only what its owner declares. Verification reports measure under a definition, and a fee map built from invoices shows what share of spend reaches the publisher. Open source fits the publisher side; buying through a DSP means working with an experienced operator.
\n1. Every layer of the stack reports on its own lane, so a dashboard is evidence of activity and not proof of value. 2. Each layer earns more when volume or spend rises, which is not the same as when your result improves. 3. ads.txt, sellers.json, and the SupplyChain object together let you inspect a path, but they record declarations, not guarantees. 4. A verification percentage is a measurement under a stated standard, so ask for the standard and the unmeasured share. 5. A fee map from real invoices is the fastest way to give a finance partner an answer and to find the layer you can remove.
\nEvery campaign needs a one-page fee map, built from your invoices and brought to the review, so that the first question about where the money went has an answer on paper.
\n1. IAB Tech Lab, OpenRTB specification, current published version, iabtechlab.com. 2. IAB Tech Lab, ads.txt and app-ads.txt specifications, iabtechlab.com. 3. IAB Tech Lab, sellers.json and SupplyChain object specifications, iabtechlab.com. 4. IAB Tech Lab, VAST (Video Ad Serving Template) and Open Measurement SDK documentation, iabtechlab.com. 5. Media Rating Council, Viewable Ad Impression Measurement Guidelines and Invalid Traffic Detection and Filtration Guidelines, mediaratingcouncil.org. 6. Prebid.org and Revive Adserver project documentation, prebid.org and revive-adserver.com.
"},{"id":"chapter-5-funnel-architecture-for-brand-managers","title":"Chapter 5: Funnel Architecture for Brand Managers","html":"By the end of this chapter, the reader will be able to: 1. Distinguish demand creation from demand capture and name which channels usually do which. 2. Explain why a loop describes real buying behavior better than a straight funnel. 3. Assign each channel a single role and a stage-appropriate KPI (key performance indicator). 4. Evaluate a proposed upper-funnel cut by estimating the delayed rise in search cost. 5. Design a holdout test that separates caused sales from captured sales. 6. Choose between a capture-only plan and a blended plan based on budget size and existing brand pull.
\nSearch advertising is mostly a collection booth. It rarely makes the crowd. It stands near the exit and takes the ticket of whoever already decided to come in.
\nNow the strongest counterargument, because it deserves a fair hearing. Some brands should run nothing but search, and they are right to. A plumber with a two thousand dollar monthly budget cannot buy a billboard presence in her town. People with a burst pipe do not need persuading. They need a phone number at the moment of panic. For her, capture is the whole job, and a funnel diagram is a distraction. The same holds for a brand that already carries heavy pull, where shoppers type its name unprompted and the only task is to be the first result.
\nSo the claim is not that capture is bad. The claim is narrower and more expensive to ignore: capture is a harvest, and a harvest needs a field somebody planted. Search platforms report their own results from the point of the click. They cannot report who put the idea in the buyer's head three weeks earlier on a video, a podcast, a creator post or a sponsorship. When those earlier efforts get cut because they look weak on a last-click report, the dashboard improves first and the business weakens later. The delay is what makes this so hard to catch. Nobody connects a cost rise in March to a budget decision made the previous autumn.
\nI have sat in the room where this happened. The finance lead is happy. The brand manager looks efficient. Two quarters on, everyone is asking why cost per sale is climbing in every channel at once.
\nThis chapter gives you a way to design the whole system on purpose, so each channel has a job, each job has a fair measure, and a cut is tested before it is made.
\nThe funnel is a fine teaching picture and a poor description of behavior. It suggests that people enter at the top, narrow down, and exit as customers. Real buyers wander. A shopper sees a video, ignores it, searches for something unrelated, sees a display ad, forgets it, gets an email, opens it, leaves the site, and returns days later by typing your brand name. The last click belongs to branded search. The earlier touches did the persuading.
\nA loop fits better. After purchase, the customer talks, reviews, reorders and refers, and those actions become the next round of awareness for other people. A good product feeds its own top of funnel. A mediocre one forces you to buy every new customer from scratch, and that is the difference between a business that compounds and one that rents its growth.
\nThink of a football season. The Saturday result is what shows up in the standings, but it was built by recruiting, conditioning and film study done months before. Nobody credits the scoreboard operator. Yet that is how last-click reporting works: it credits the person nearest the final whistle.
\nYou do not need to abandon the funnel vocabulary. Use it as a filing system for jobs, and stay aware that real people skip stages, repeat them, and move backward. Plan the structure, then measure honestly knowing the structure is a simplification.
\nTwo jobs sit under every plan. Creation puts a need, a problem or a brand name into a head that did not hold it before. Capture collects the person who now holds it and is acting on it.
\nCreation channels typically include online video, connected television, podcasts, social reach campaigns, creator partnerships, sponsorships and display. Capture channels typically include search ads, branded search, retargeting, email to existing leads, and marketplace listings. Typically is the operative word. A search ad for a problem phrase can create awareness of a solution the reader never knew existed. A video can close a sale. The channel does not set the job. The message and the audience do.
\nThe practical test is this: would this sale have happened without this ad? For a captured sale, usually yes, through some other path. For a created sale, usually no. Most dollars sit somewhere between, which is exactly why you test instead of assuming.
\nHere is the mechanism that makes this thesis worth a chapter. Search volume for a category or a brand is not a fixed pool. It is a stock that other activity refills. Every video view, podcast mention and friend recommendation can become a later search query. Cut the activity and the stock drains. Search then has fewer people to capture, the auction for the shrinking pool gets no cheaper, and your cost per sale rises. You pay for the cut in search, which is the one line that never appears in the meeting where the cut was approved.
\nGive every channel one job and one scorecard. Mixed jobs produce muddy reports.
\n| Role | Typical channels | Fair KPI | Misleading KPI | |---|---|---|---| | Create | Video, podcast, creators, reach social | Reach, frequency, branded search lift in exposed areas | Direct conversion rate | | Shape | Non-branded search, content, reviews | Assisted conversions, share of qualified visits | Last-click cost per sale | | Capture | Branded search, retargeting, email to leads | Incremental conversions, cost per incremental sale | Platform-reported return on ad spend | | Retain | Email, owned community, referral | Repeat rate by cohort, referral rate | Open rate alone |
\nThe right column deserves attention. Every one of those misleading KPIs is attractive because the platform hands it to you for free. Direct conversion rate makes a video look like a failure. Platform-reported return on ad spend makes retargeting look like a miracle, since retargeting talks to people already leaning in.
\nThat is why a single KPI across the whole funnel misleads. Blend creation and capture into one cost per acquisition and the capture channels flatter the average while the creation channels drag it. A manager who optimizes the blend will cut the drag. The average improves right up until the pool empties.
\nSuppose a regional outdoor retailer, Cedar Hollow Supply (a composite, illustrative brand), reports a blended cost per sale of 38 dollars. Splitting by role shows retargeting at 14 dollars and video at 90 dollars. The instinct says move the video money to retargeting. But retargeting can only spend as far as the audience video built. Double its budget and the cost climbs fast, because the audience does not grow. Two numbers measured in different units were being compared as if one could replace the other.
\nThe delay is the hardest part to teach, because the evidence arrives late and looks like something else.
\nSuppose Cedar Hollow cuts 12,000 dollars a month from video and podcast spending. In month one, reported efficiency rises, since the cut removed the channels that look worst on direct reports. In month two, branded search volume starts to soften, but only slightly, and seasonality offers a convenient excuse. By month four, qualified site visits are down, retargeting audiences are thinner, and cost per sale is rising everywhere. The team reacts by trimming the top again. That is the loop running in reverse.
\nWhy the lag? Awareness decays slowly. People who already heard of you keep searching for you for weeks. The pool drains gradually, and the cost shows up only when the old stock runs low. How long it takes depends on your purchase cycle. A pizza shop may feel it in weeks. A company selling kitchen remodels may take two quarters or more.
\nThe leading indicators move first. Watch three: branded search volume (use your own search console and analytics data, which cost nothing), direct and returning visits, and new-visitor share of site traffic. If those turn down while paid capture efficiency still looks healthy, you are seeing the harvest of an old crop.
\nBe honest about the limit here. Branded search also moves with seasons, press, price changes and competitor activity. A softening line is a flag to investigate, not a verdict. The point of a leading indicator is to start the question early, then answer it with a test.
\nOpinions about this topic tend to harden into camps. A test settles it for your brand. The cleanest version is a geographic holdout.
\nPick matched regions, similar in size, season and sales history. Keep upper-funnel spending running in most of them. Pause it, or cut it by a stated amount, in the rest. Run long enough to span at least one full purchase cycle, with an agreed stopping rule written before the test begins. Then compare branded search volume, total sales and new customers between groups, not platform-reported conversions.
\nTwo cautions. First, sample size: a handful of small regions produces noise that looks like signal, so use as many matched areas as you can and be wary of a result from only a few. Second, spillover: people travel and streaming does not respect county lines, so choose regions with some separation.
\nWhen you cannot run a split, use a time-series method. CausalImpact, an open source package built on Bayesian structural time-series, estimates what a metric would have done without the change, using related series as controls. It gives a range, not a single number, and you should report the range.
\nA branded search holdout belongs in the same toolkit. Pause branded search ads in a few matched regions and watch whether total orders drop. Often much of the traffic simply moves to the free listing underneath. In some accounts it does not, because competitors bid on your name. Either answer changes how you spend, and neither can be known without trying.
\nPerformance skeptics are partly right, and a fair chapter says where.
\nCapture-only works when the budget is too small to create anything. Below a certain level, a reach campaign touches too few people too few times to register. Your money is better spent on a clean website, a well-built search account and strong reviews.
\nIt works when the brand already has pull. A company with a loyal base, steady word of mouth or a physical presence that does the creation for free can harvest and protect that demand without paying to generate it.
\nIt works when the need is urgent and local: locksmiths, towing, emergency repair. The buyer arrives already decided.
\nAnd it works as a starting position for a young brand with a short cycle that is still proving the product. Prove the offer converts first. Funding awareness for a product people do not want is expensive theater.
\nA rough guide by situation:
\nI will not give you a magic percentage split, because any single figure would be invented. Let the holdout results, your cycle length and your margin set it.
\nYou do not need an expensive suite to see this. You need four things: consistent event tracking, a stage definition everyone shares, a place to store rows, and a weekly view.
\nDefine stages by what the buyer is doing: has not yet heard of you, is aware of the problem, is comparing options, is ready to act, has bought. Write the definitions down, and use the same ones in every platform so a \"consideration\" audience means the same thing everywhere.
\nThen track how many people reach each stage and how many move to the next. Matomo funnels let you define steps on your own site and see where visitors drop. Store stage counts and cohort rows in PostgreSQL or DuckDB, and a handful of queries gives you the weekly stage report, including repeat purchase by acquisition source.
\nKeep the report plain. One table per week: reach, branded search volume, assisted conversions, incremental conversions where tested, repeat rate. Few numbers, read consistently, beat a dashboard nobody trusts. Keep family-first defaults in mind for audiences: set conservative targeting around children and sensitive categories, and let your counsel decide on consent and disclosure questions, since this is operational guidance only.
\nDemand creation: Advertising that puts a need, problem or brand name into the mind of a person who did not hold it before. Demand capture: Advertising that collects people who already want something and are actively looking for it. Branded search: Searches that include your company or product name, which usually signal existing awareness. Holdout: A test in which some regions or audiences are deliberately left unexposed so you can compare them to exposed ones. Incrementality: The share of results that happened because of the advertising, as opposed to results that would have happened anyway.
\n1. Judging a creation channel by direct conversions, which guarantees it looks like waste while the search channel collects its work. 2. Trimming the top of the funnel to improve a blended cost per acquisition, then blaming the search team when costs rise two quarters later. 3. Moving money from video to retargeting because the retargeting report looks better, without noticing that retargeting can only spend as far as its audience reaches. 4. Treating a dip in branded search as proof of failure without checking season, price, press and competitor bidding first. 5. Running a holdout in too few regions or for less than one purchase cycle, then trusting a noisy result.
\n1. Give each channel one role and one scorecard, and write both down where the whole team can see them. 2. Track branded search volume, returning visits and new-visitor share as standing leading indicators. 3. Test before cutting: run a matched-region holdout or a time-series estimate on any upper-funnel reduction of meaningful size. 4. Report cost per incremental sale where you have test results, and label platform numbers as platform numbers. 5. Reset stage definitions once a year, and keep them identical across every platform account.
\n1. Ask every platform rep one question: how much of this reported conversion would have happened without the ad? A straight answer is rare, and the pause tells you plenty. 2. Match test length to purchase cycle. A holdout shorter than your typical time from first touch to purchase tests nothing. 3. Protect small creation budgets from quarterly cuts by labeling them as a reserve with its own scorecard. 4. Post-purchase email and referral asks cost little and feed the loop, so give them real attention before buying more reach. 5. When a cut is unavoidable, cut by region first and watch for four to eight weeks before cutting everywhere.
\nMatomo funnels (GPL) let you define the steps of a buying path on your own site and see where visitors drop away, with the data staying on servers you control. For stage reporting, store counts and cohort rows in PostgreSQL (PostgreSQL License), a reliable database for ongoing reports, or DuckDB (MIT License), an embedded analytical database that runs queries on exported files from a laptop. Between them you can produce the weekly stage table described above with a few queries and no paid reporting software. For time-series holdout estimates, the CausalImpact package named in the references is open source under the Apache 2.0 license.
\nSearch is the most visible channel and the least independent one. It collects demand that video, audio, creators, sponsorships and word of mouth help generate, and it reports as if the credit were its own. A funnel is a useful filing system for channel roles, but real buyers loop, skip and double back, so each channel needs one job and a fair KPI. Cutting creation spend flatters reported efficiency now and raises capture cost later, after a lag set by your purchase cycle. Leading indicators give early warning, and a matched-region holdout gives the answer. Capture-only is correct for tiny budgets, urgent local needs and brands with strong existing pull. Everyone else owes the top of the system a protected, tested budget.
\n1. Search captures demand that other activity creates, so cutting creation eventually raises search cost. 2. Assign each channel one role and judge it by a KPI that fits the role. 3. A single blended KPI rewards harvesting and punishes planting. 4. The cost of an upper-funnel cut arrives late, so watch leading indicators and test before cutting. 5. Pure capture is a sound plan for tiny budgets, urgent local needs and brands with strong existing pull.
\nPick one creation channel you are tempted to cut ahead of the budget review and run a matched-region holdout on it for one full purchase cycle. Decide with the test result, not with the last-click report.
\n1. Brodersen, K.H., Gallusser, F., Koehler, J., Remy, N., and Scott, S.L. Inferring causal impact using Bayesian structural time-series models. Annals of Applied Statistics. Methodology behind the CausalImpact open source package. 2. Matomo. Funnels feature documentation. Reference for defining and reading funnel steps in a self-hosted analytics installation. 3. PostgreSQL Global Development Group. PostgreSQL documentation. Reference for storing and querying stage and cohort data. 4. DuckDB Foundation. DuckDB documentation. Reference for local analytical queries over exported event files.
"},{"id":"chapter-6-setting-objectives-that-survive-finance-review","title":"Chapter 6: Setting Objectives That Survive Finance Review","html":"By the end of this chapter, the reader will be able to: 1. Define the chain that links a business goal to a marketing objective and then to a single channel KPI. 2. Calculate contribution margin, payback period and the customer acquisition cost to lifetime value ratio from a simple order model. 3. Apply an incrementality adjustment to a platform-reported result and restate the true return. 4. Diagnose an optimization-event mismatch in an ad account and choose a better event to optimize toward. 5. Write an objective statement with all required parts, including guardrail metrics and a pre-agreed measurement method. 6. Evaluate a draft campaign plan for the weak points a finance reviewer will find first.
\nSuppose a campaign spends 20,000 dollars and the ad platform reports 500 purchases. Each order is 100 dollars, so the dashboard shows 50,000 dollars in revenue. The cost per purchase is 40 dollars. Your product costs, shipping and payment fees leave 45 dollars of contribution margin (the money left from an order after the direct costs of fulfilling it). So each order appears to earn five dollars after ad cost. A tidy profit of 2,500 dollars. Everyone smiles.
\nNow suppose a clean holdout test shows that half of those buyers would have purchased anyway. The campaign created 250 orders, not 500. Those 250 orders produced 11,250 dollars of contribution against 20,000 dollars of spend. The campaign lost 8,750 dollars while its dashboard said it made money.
\nHere is the assumption I want to challenge: that a good objective is one with a number in it. \"Hit a cost per purchase of 40 dollars\" is a number. It is specific, measurable and dated. It is also a trap, because it can be achieved perfectly while the business loses cash. Finance people have seen this movie. They have approved budgets against numbers like that, watched the bank balance disagree, and learned to distrust anything a marketing team calls a win.
\nI spent years around lenders, credit files and balance sheets before I spent them around ad accounts, and the lesson transfers cleanly. A lender does not care how impressive the application looks. The lender cares whether the money comes back, how fast, and with what margin of safety. Your finance reviewer reads your plan the same way.
\nThis chapter is about writing objectives in that language, so the plan survives the meeting where someone asks, \"And how do we know this works?\" Operational guidance only, and your own finance team decides what counts as acceptable evidence in your business.
\nEvery objective sits on a ladder with three rungs. At the top is a business objective: grow contribution margin dollars, reduce dependence on one product line, enter a new region. In the middle is a marketing objective: acquire a defined number of new customers at a defined cost, or lift repeat orders from existing buyers. At the bottom is a channel KPI: the single metric a media buyer can steer day to day.
\nThe rule is simple and almost nobody follows it. Each rung must be derivable from the one above it, and a reviewer should be able to walk down or up without a gap. If the business wants 300,000 more dollars of contribution margin, and a new customer is worth 60 dollars of contribution in the first year, you need roughly 5,000 new customers. If your sales team or email program will deliver two thousand of them, paid media owns the remaining 3000. Now the channel KPI writes itself: new-customer orders at a stated cost ceiling.
\nNotice what is missing from that chain. Impressions, clicks and click-through rate appear nowhere. They still matter, but as diagnostics, the way a coach watches film. The scoreboard is the top rung.
\nOne more discipline helps. Write the ladder on one page, before choosing channels. When the channel comes first, people reverse-engineer a business case around a plan they already wanted, and finance can smell it.
\nFour terms carry most of the weight in a budget conversation.
\nContribution margin is revenue minus the variable costs of the sale: product cost, shipping, payment processing, returns allowance. It is not gross margin on a spreadsheet tab and it is not profit. It is the pool the ad budget has to be paid from.
\nPayback period is how long it takes the contribution from an acquired customer to repay what you spent to acquire them. A business with strong cash reserves may accept a payback of 12 months. A small shop paying suppliers every month may need 60 days. Ask finance which it is before you plan, because the answer decides which campaigns are even eligible.
\nCustomer acquisition cost, or CAC, is the total cost to win one new customer, including creative production and agency fees, not only media. Lifetime value, or LTV, is the contribution expected from that customer over a defined period. A ratio of lifetime value to CAC of three to one is a common rule of thumb, but treat it as a conversation starter, not a law.
\nIncrementality is covered in the next section because it deserves room.
\nA small table makes the logic concrete. These are illustrative figures for a fictional subscription snack brand.
\n| Item | Illustrative value | |---|---| | First order revenue | 40 dollars | | Contribution margin on first order | 14 dollars | | Expected contribution over 12 months | 56 dollars | | Fully loaded CAC | 28 dollars | | Twelve-month LTV to CAC | two to one | | Payback period | about five months |
\nWhether that table passes depends on finance's cash position, not on how it feels to marketing.
\nAn incremental conversion is one that happened because of the ad and would not have happened otherwise. Platforms count every conversion they can connect to an ad exposure. That includes loyal customers who were already on their way to your site, and people who saw an ad while their purchase was already decided.
\nThe adjustment is plain arithmetic. Take the reported conversions. Multiply by the incremental share your test measured. Divide spend by that smaller number. The result is your incremental CAC, and it is the only CAC finance should see in a profitability argument.
\nWhere does the incremental share come from? From a test, never from a guess. The common designs are a geographic holdout (certain regions get no ads and serve as the control), a user-level holdout where the platform supports it, or a time-based on and off test. Each has weaknesses. Geographic tests are clean but need enough regions and enough volume. User-level tests depend on the platform's own randomization. On and off tests are easy to confuse with seasonality.
\nDifferent campaign types carry different incremental shares. Branded search, which captures people already looking for you, often shows a low incremental share. Prospecting to new audiences often shows a higher one, though it also converts later and at lower volume. I will not hand you benchmark numbers because they vary too much by account. Measure yours.
\nIf you cannot test before the first budget request, say so out loud and propose a reduced budget with a test built in. A finance director forgives uncertainty. Disguised uncertainty is what gets people replaced.
\nAd platforms learn from the event you tell them to optimize toward. Tell them to find people who add to cart, and they will find the people who add to cart most cheaply. Those people are not necessarily the people who buy.
\nThis is optimization-event mismatch: the account is steered toward a cheap event whose link to profit is weak. Picture a coach rewarding a striker for shots on goal rather than goals. The shots go up. The scoreboard does not.
\nConsider a fictional online furniture retailer. Its campaign optimizes to lead form submissions, which cost 12 dollars each. The sales team closes one lead in 10, so each sale costs 120 dollars in media. Average contribution per sale is 90 dollars. The campaign looks efficient at the top of the funnel and loses money at the bottom. The platform did exactly what it was told.
\nThe fix is to move the optimization event as deep into the funnel as volume allows. Most platforms need a minimum flow of events each week to learn well, so a purchase event may be too sparse for a small account. In that case, optimize to a mid-funnel event that you have shown predicts purchase, such as a qualified lead marked by sales, and feed that signal back. Document the reasoning in the objective so finance understands why the optimized event differs from the reported outcome.
\nCheck the pairing quarterly. Customer behavior changes, and an event that once predicted revenue can drift away from it.
\nA guardrail metric is a number that must not get worse while you chase your main objective. It exists because any single target can be hit in ways that damage the business.
\nPush hard on CAC and you may pull in buyers who return products, cancel early or never reorder. Push hard on volume and the mix tilts toward discounted items. Push hard on cheap leads and sales burns hours on people who never answer the phone.
\nChoose three or four guardrails and set limits before spend begins. Good candidates include refund or return rate, share of orders on a discount code, new-customer share of total orders, 60 day repeat purchase rate, sales acceptance rate of leads, and brand search volume trend. Pick the ones tied to how this particular business loses money.
\nWrite the guardrail with a trigger and an owner: \"If return rate exceeds the agreed ceiling for two consecutive weeks, the media lead pauses the affected campaign and meets finance within three working days.\" A guardrail without an owner is decoration.
\nFor businesses serving families or sensitive categories, add a conservative guardrail on audience eligibility and creative review. A fast CAC is never worth an audience you should not be targeting.
\nHere is the template I use. It forces every part finance will ask about into a few sentences.
\n\"Over [time period], we will acquire [number] new customers in [market or segment] at an incremental CAC no higher than [amount], measured by [method], so that payback occurs within [period] and contribution margin from new customers reaches [amount]. We will optimize toward [event], justified because [evidence]. We will hold [guardrail one], [guardrail two] and [guardrail three] within [limits]. The budget is [amount], released in [stages], with continuation conditional on [decision rule].\"
\nRead it aloud. Does it name the business outcome? The time? The measurement? The evidence for the optimization event? The guardrails? The staged release?
\nStaged release matters more than people expect. Asking for the full amount in one lump says \"trust me.\" Asking for the first slice, with the second released when a pre-agreed readout hits a pre-agreed threshold, says \"check me.\" Finance usually prefers the second, and approval comes faster.
\nMost budget fights are not about results. They are about how results should be counted, argued after the numbers arrive. The cure is a one-page measurement agreement signed by marketing and finance before the first dollar moves.
\nIt should answer six questions. Which data source is the system of record for orders: the store back end, the accounting system or the ad platform? What counts as a new customer? Which attribution window applies, and is it the same for every channel? What test design measures incrementality, and what result would count as success, failure or inconclusive? Who runs the analysis? And what happens to budget under each outcome?
\nPre-agreeing the \"inconclusive\" outcome is the part that feels awkward and pays off most. Tests often land in a gray zone. If you have already agreed that a gray result means \"continue at the same budget for one more cycle,\" nobody gets to relitigate it with a hunch.
\nA free toolchain can carry this load. Export orders and ad spend to files, join them in DuckDB, and keep the agreement and the scoreboard together in a spreadsheet that finance can open.
\nConsider a fictional regional home-fitness brand, Wrenfield Rowing. Its business objective is 240,000 dollars of new-customer contribution margin over two quarters. A new customer buys one machine at 900 dollars, with contribution margin of 270 dollars after cost of goods, delivery, payment fees and a returns allowance. Customers buy accessories later, adding an expected 30 dollars of contribution within 12 months, so LTV is 300 dollars.
\nTo hit the goal, the brand needs about 800 new customers. Email and referrals are expected to deliver 200. Paid media owns 600.
\nFinance sets a payback ceiling of 90 days and says CAC must stay under one third of first-order contribution plus accessories, which means a ceiling near 100 dollars per incremental customer. Media budget requested: 90,000 dollars, released in two halves.
\nThe platform's reporting suggests a cost per purchase of 75 dollars. But a regional holdout in a prior test found that about 60 percent of reported purchases were incremental. Dividing 75 dollars by the 60 percent incremental share gives an incremental CAC of 125 dollars. That fails the ceiling. So the plan changes: shift budget from retargeting, where the incremental share is typically lowest, toward prospecting, and rerun the holdout at the first readout. The objective statement now says \"incremental CAC no higher than 100 dollars, verified at the midpoint, with the second half of budget released only if the midpoint test shows 110 dollars or lower.\"
\nGuardrails: return rate under the agreed ceiling, discount-code orders under the agreed share, and repeat accessory purchases holding steady.
\nThat plan can be rejected, but it cannot be dismissed as marketing hopes. Every claim has an owner, a number and a test.
\nContribution margin: Revenue from a sale minus its variable costs, such as product cost, shipping, payment fees and returns. It is the money available to pay for advertising and overhead. Payback period: The time it takes for contribution from an acquired customer to repay the cost of acquiring them. Incrementality: The portion of results caused by advertising, as opposed to results that would have happened anyway. It is measured with a holdout or similar controlled test. Guardrail metric: A secondary number with a pre-agreed limit that must not deteriorate while the main objective is pursued. Optimization event: The specific conversion action an ad platform is told to maximize when it decides whom to show ads to.
\n1. Reporting platform conversions as if they were all caused by the ads, which overstates return and exposes the plan the moment finance compares it to bank deposits. 2. Setting a cost per acquisition target without checking it against contribution margin, so a campaign can hit its goal and still lose money on every order. 3. Optimizing the ad account toward a cheap early event, such as page views or form starts, because the volume looks healthy while paying customers stay scarce. 4. Defining success after results arrive, which turns every ambiguous number into an argument about method rather than a decision about money. 5. Requesting the entire budget up front with no stage gates, so finance has no way to approve cautiously and defaults to saying no.
\n1. Start every plan with the contribution margin per order and the payback ceiling finance will accept, then work backward to a media cost limit. 2. Report incremental CAC beside platform CAC in every readout, and label which one the budget decision rests on. 3. Choose the deepest optimization event your weekly volume can support, and document the evidence that it predicts revenue. 4. Set three or four guardrail metrics with named owners, limits and a written response before spend begins. 5. Sign a one-page measurement agreement with finance that covers data source, attribution window, test design and what each outcome means for budget.
\n1. Ask finance for the cost of capital or cash constraint behind the payback ceiling, because it often reveals more room, or less, than the stated rule. 2. Include creative production, agency fees and tools in CAC. Finance will add them eventually, and it is better that you did it first. 3. When you cannot test beforehand, request a smaller initial budget with a test built in rather than defending an untested assumption. 4. Rerun your incrementality test after major changes in offer, audience or channel mix, since an old result describes an old account. 5. Keep a running file of past objectives and what happened. Your track record becomes the most persuasive page in your next request.
\nTwo free tools cover most of the work in this chapter. LibreOffice Calc (MPL license) is a spreadsheet where you can build the contribution margin model, the payback calculation, the incrementality adjustment and the signed measurement agreement on separate sheets that finance can open without any special software. DuckDB (MIT license) is a fast analytical database that runs on a laptop. Use it to join order exports, ad spend files and test results with ordinary queries, then export clean tables back into Calc for the readout. Neither tool needs a server, an account or a subscription, and finance can inspect every formula and query.
\nAn objective survives finance review when it descends from a business goal, speaks in contribution margin, payback and CAC against lifetime value, and restates platform results through an incrementality adjustment. It names an optimization event that actually predicts revenue, sets guardrail metrics with owners, and follows a template that forces evidence, timing and staged budget release into plain view. The measurement method is agreed in writing before spend begins, including what an inconclusive result means. The worked example shows that a plan with these parts can be challenged on its numbers but not dismissed as hope.
\n1. A target that can be hit while the business loses money is not an objective, so check every cost target against contribution margin. 2. Platform CAC is a claim. Incremental CAC, backed by a holdout, is evidence. 3. The optimization event is a spending instruction, so pair it with proof that it predicts paying customers. 4. Guardrails with limits and owners protect against hitting one number by damaging another. 5. Agreeing the measurement method, including the inconclusive case, before spend ends most budget arguments early.
\nNever ask for a budget increase without one page showing contribution margin per order, the payback ceiling finance accepts, and your incremental CAC from a real holdout. Then ask for the first half of the money, with the second half tied to a readout both teams signed in advance.
\n1. Brodersen, K.H., Gallusser, F., Koehler, J., Remy, N., Scott, S.L. Inferring causal impact using Bayesian structural time-series models. Annals of Applied Statistics. Published methodology behind the CausalImpact approach to measuring ad lift. 2. Kohavi, R., Tang, D., Xu, Y. Trustworthy Online Controlled Experiments: A Practical Guide to A/B Testing. Cambridge University Press. Reference for experiment design and interpretation. 3. Federal Trade Commission. Advertising and Marketing guidance on truthful claims and substantiation. Operational context only, counsel decides. 4. DuckDB Documentation. SQL reference and data import guides for the MIT licensed analytical database. 5. The Document Foundation. LibreOffice Calc Guide. Reference for spreadsheet modeling in the MPL licensed office suite.
"}]},{"title":"Part II: Search and Intent","items":[{"id":"chapter-7-paid-search-structure-and-intent-tiers","title":"Chapter 7: Paid Search Structure and Intent Tiers","html":"By the end of this chapter, the reader will be able to: 1. Name the five intent tiers of a paid search account and describe the job each one does. 2. Organize campaigns and ad groups around intent tiers instead of product themes. 3. Compare the control and automation tradeoffs of common campaign types in general terms. 4. Design a negative keyword architecture that keeps each tier from stealing traffic from the others. 5. Calculate tier budget caps from a business objective and a worked example. 6. Evaluate a search term report by tier and decide which tier deserves more money, less money, or a test.
\nHere is the assumption I hear most often from brand managers who inherit a search account: \"Search is the channel we can trust, because it shows its return right there in the dashboard.\" The dashboard does show a return. That is exactly the problem.
\nOpen any account that has run for a year and sort the spend by search term. A large share of the conversions will usually sit on queries that contain the company name. Those searches are cheap, they convert often, and they make the blended numbers look wonderful. Pleased with the blended number, the brand manager asks for more budget and gets it. The extra money chases the same easy queries, the ones where the customer had already decided, and the account gets more expensive without getting any bigger.
\nA search account is a stadium with five entrances, each at a different distance from the buying decision. Someone typing your brand name is at the gate. Someone typing a vague question about a problem is still in the parking lot, and may never walk in. Averaging those two people into one conversion rate is not analysis. It is a way of hiding the part of the account that is working from the part that is only collecting the credit.
\nThis chapter shows how to build an account where the structure itself prevents that confusion. Platform menus change often, so this chapter stays on structure and leaves out button-level steps.
\nIntent is the reason behind the query. A tier is a group of queries that share a reason, and therefore share a job, a measurement method and a spending logic. I use five.
\nOwn brand. The searcher typed your company or product name. They are navigating, not shopping. The job here is defensive: keep the result page clean, keep competitors from sitting on your name, and send people to the right page fast. Conversion rates are high and costs are low, which is precisely why this tier flatters the whole account.
\nHigh-intent category. The searcher named a product type plus a signal that they are ready to act, such as a location, a price word, a comparison, or the words \"near me\" or \"buy\". They have not chosen a brand. This is the tier where paid search earns its keep, because it captures people deciding right now.
\nBroad category. The searcher named the product type with no buying signal. They may be browsing, studying or killing time. Volume is large, quality is mixed, and the temptation to overspend is strong.
\nCompetitor. The searcher typed a rival's name. They may be a loyal customer looking for a login page, or a frustrated one looking for the exit. In many accounts the former is more common, and the search term report will tell you which.
\nResearch. The searcher asked a question or described a problem without naming any product category. \"How do I keep a small crew scheduled across three job sites\" is a research query. These people are earliest in the process. Direct conversions are rare and that is fine, because the job of this tier is introduction, not harvest.
\nThe structure most accounts inherit is organized by product theme. A footwear brand has a running shoes campaign, a trail shoes campaign, a walking shoes campaign. It reads cleanly and it makes ad copy easy to write. It is also useless for a budget conversation.
\nInside the running shoes campaign you will find all five tiers mixed together under one budget, one bidding setting and one reported conversion rate. When the manager asks \"is running shoes working?\", nobody can answer, because the answer is five different answers.
\nThe tier structure turns this around. Campaigns are named for the tier first and the theme second. So you get Own Brand, High-Intent Category, Broad Category, Competitor and Research as the top-level campaign groups, and the product themes become ad groups inside the category tiers. Now every question about money has a clean answer.
\nThere is a cost. More campaigns thin out conversion data, so a small account should collapse to three groups (own brand, category split into two ad groups, everything else) and add finer splits as volume grows.
\nEvery major search platform now offers campaign types that sit along one spectrum. At one end, you choose keywords, write ads and set bids. At the other end, you hand over a goal, a budget, some assets and a landing page, and the system decides which queries to enter. I will describe them generically here. Each platform names and configures these campaign types differently, so I leave out menu-level steps.
\nMore automation means more reach and less visibility. More control means more visibility and more work.
\nWhere should each tier sit? My rules of thumb:
\nThe deeper issue with fully automated campaign types is that they report one blended result, which is what this chapter argues against. Before running one across tiers, ask what controls exist for excluding your brand name and what search term reporting it gives you. Where the answers are thin, keep it small and watch Own Brand for cannibalization.
\nA negative keyword tells the platform not to show your ad for queries containing a given word or phrase. Most managers treat negatives as a cleanup chore: spot a bad query, add it to a list, move on. In a tiered account, negatives do something more important. They are the walls between the tiers.
\nWithout walls, queries fall into whichever campaign the platform judges most eligible, the tiers leak into each other, and the reporting turns to mush.
\nBuild the walls in three layers.
\nLayer one: shared lists across tiers. Create a shared negative list of your brand terms and attach it to every campaign except Own Brand. Create a list of competitor names and attach it to every campaign except Competitor. Now a query containing your name can only be served by the Own Brand campaign, and a query containing a rival's name can only be served by the Competitor campaign.
\nLayer two: cross-tier exclusions inside the category tiers. The high-intent category campaign should exclude the soft words that define the broad tier, such as \"what is\", \"meaning\", \"definition\", \"jobs\", \"free\", \"template\" and \"diy\". The broad category campaign should exclude the buying words that define the high-intent tier, so those queries route to the campaign with the higher bid and the tighter landing page. Each query should have exactly one home.
\nLayer three: true junk. Careers searches, free-download seekers and wrong-geography phrases belong on an account-level list that applies everywhere.
\nReview search terms weekly for the first two months of a new structure, then every two weeks. Ask which queries landed in the wrong tier, which burned money for nothing, and which deserve their own keyword.
\nOne caution. A careless exclusion can quietly cut off valuable traffic. Keep a change log, and when a tier's volume drops without explanation, check the recent negatives first.
\nA tier structure only works if the money respects it. Without caps, the platform will happily pour budget into whichever campaign converts cheapest on paper, which is nearly always Own Brand. The cap is how you tell the account what you want, instead of what is easiest.
\nStart from the job of each tier and the stage of the business, then set a cap as a share of total search spend. Here is an illustrative example. Suppose a regional home services brand, which I will call Stonebrook Services, has a monthly search budget of 100,000 dollars. This is a fictional composite for teaching purposes.
\n| Tier | Monthly cap | Share | Reasoning | |---|---|---|---| | Own brand | 10,000 dollars | 10 percent | Defense only. Cap it near what the brand actually needs to hold its name. | | High-intent category | 45,000 dollars | 45 percent | Core demand capture, where the account earns most of its return. | | Broad category | 20,000 dollars | 20 percent | Reach with strict guardrails and a long negative list. | | Competitor | 10,000 dollars | 10 percent | A controlled test budget until the data says otherwise. | | Research | 15,000 dollars | 15 percent | Introduction spend, judged on different measures. |
\nThose shares show a method, not a recommendation. A young brand nobody searches for by name would shrink Own Brand and push money into research and broad category. A mature brand might run Own Brand heavier and cut research.
\nThree rules keep caps honest. Enforce them with tier-level campaign budgets, not one account total. Move money between tiers only on a schedule and with evidence, not in reaction to last week's dashboard. Treat the Own Brand cap as a ceiling rather than a target. If the tier needs only six thousand dollars to hold its position, spending 10,000 is not defense, it is waste with a nice conversion rate.
\nThe reporting view that matters is one row per tier. Spend, clicks, conversions, cost per conversion, and the share of total conversions each tier produced. Place that table at the top of every monthly report, above the blended number. If a stakeholder wants only one figure, give them the tier table anyway.
\nThen add the measures that fit each tier's job, because they are not the same measure.
\nFor Own Brand, track cost per conversion and, more importantly, schedule a holdout test. Pause the tier in a few matched regions, compare total conversions against regions where it kept running, and see how many sales moved over to organic listings.
\nFor High-Intent Category, judge by cost per conversion against the margin each sale carries. This tier should be the most accountable one in the account.
\nFor Broad Category, watch the share of spend that lands on queries you would never have chosen and the rate at which new converting queries graduate into the high-intent tier. A healthy broad tier feeds the one above it.
\nFor Competitor, report conversions alongside average order value and, where you can track it, repeat purchase. Also track clicks from existing customers hunting for a login, which cost money with no upside.
\nFor Research, direct conversions are the wrong measure. Look instead at changes in brand-name search volume in regions and weeks where the research campaigns ran harder, and at assisted conversions where your tracking can see them. Be honest in the report about what this evidence can and cannot prove. A rise in brand searches is a signal, not a verdict.
\nReturn to Stonebrook with illustrative numbers. After a quarter on the tier structure, Own Brand shows the lowest cost per conversion in the account. High-Intent Category produces the most conversions. Broad Category spends its full cap with few direct conversions, though several queries keep converting. Competitor matches the account's average order value, but a fifth of its clicks look like rival customers hunting for support pages. Research shows almost no direct conversions.
\nA manager reading the blended number sees a healthy account and asks for more budget. A manager reading by tier acts differently. The Own Brand holdout shows most of that tier's sales would have arrived organically, so its cap is cut by 40 percent and the savings move to High-Intent Category. The converting broad queries are promoted into the high-intent tier and added as negatives in the broad campaign. Support phrases join the Competitor negative list. Research stays for one more quarter only if brand-name searches rose where it ran harder.
\nNo heroics. Each change follows from a tier-level number that a blended report would have buried.
\nIntent tier: A group of search queries that share the same reason for searching and therefore the same job, budget logic and measurement method in the account.
\nSearch term report: The platform report listing the actual queries people typed that triggered your ads, as opposed to the keywords you chose to bid on.
\nNegative keyword: A word or phrase that blocks your ad from showing on any query containing it, used to route traffic between campaigns and to remove waste.
\nCannibalization: A situation in which one of your campaigns takes clicks that another of your campaigns, or your organic listing, would have earned at lower cost.
\nHoldout test: A measurement method that pauses a campaign in matched regions or periods and compares results against where it kept running, to estimate how much of its output was truly caused by the ads.
\n1. Reporting one blended conversion rate for the whole account, which lets cheap brand-name conversions hide expensive and unproductive spend elsewhere. 2. Letting an automated campaign type run across all tiers with no exclusion of the brand name, so it quietly absorbs queries the Own Brand tier would have served for less. 3. Treating negative keywords as a one-time cleanup instead of a standing routing system, which lets the tiers leak into each other within weeks. 4. Judging the Research tier by direct conversions and cutting it in the first month, before any lift in brand searches could have appeared. 5. Setting a single account budget and trusting the platform to divide it, which sends money to whichever campaign is cheapest to convert rather than whichever one the business needs.
\n1. Name campaigns for the tier first and the product theme second, so every budget and reporting question maps to a campaign. 2. Attach a shared brand-name negative list to every campaign except Own Brand, and a competitor-name negative list to every campaign except Competitor. 3. Set a monthly cap for each tier, treat the Own Brand cap as a ceiling, and move money between tiers only on a fixed review schedule. 4. Run a holdout test on the Own Brand tier before approving any budget increase for the account. 5. Review the search term report weekly after any structural change, and keep a written log of every negative keyword added.
\n1. When a competitor name appears in your own brand campaign reports, check whether your ad copy mentions it, because loose matching can pull those queries in. 2. If the account is too small to feed automated bidding, simplify the tier structure before adding more campaigns, since thin data produces erratic bids. 3. Add your most common typos and spelling variants of your brand name to the Own Brand tier, so those searchers are not routed into the category campaigns. 4. Give each tier its own landing page where you can: the Own Brand visitor wants the right page fast, the Research visitor wants an answer before a pitch. 5. Check the Competitor tier for ad copy that makes comparison claims, and have counsel review those claims, because comparative advertising rules vary and this book gives operational guidance only.
\nTwo free, open source tools handle most of the tier analysis described here. DuckDB (MIT license) is an embedded analytical database that runs SQL directly against exported CSV files, so a year of search term data can be grouped by tier in seconds without setting up a server. pandas (BSD license) is a Python library for tables of data, useful for tagging each query with its tier through simple rules, joining search term exports to conversion records, and producing the one-row-per-tier report. Both are free to use and run on an ordinary laptop.
\nA paid search account is not one channel but five jobs sitting at different distances from the purchase decision: defending the brand name, capturing high-intent category demand, reaching broad category audiences, contesting competitor searches, and introducing the brand to people still researching. Structuring campaigns by tier, building negative keyword walls between them, capping the budget per tier, and reporting each tier by its own measure turns a blended dashboard into a set of specific decisions. Automation has a place, but it should sit where it can be observed, and the tier table should always sit above the blended number.
\n1. The five tiers are own brand, high-intent category, broad category, competitor and research, and each has its own job and its own measure. 2. Theme-based structure hides the answer to the question that matters, which is where the money is actually producing something new. 3. Negative keywords, used in layers, act as the walls that keep each query in exactly one campaign. 4. Tier budget caps should come from the job of each tier and the stage of the business, and be moved only on a schedule and with evidence. 5. Report the tier table before the blended number, and use a holdout for Own Brand and brand-search movement for Research.
\nOne rule for the budget meeting: rebuild the monthly search report as one row per tier, run a holdout on your Own Brand campaign, and shift any dollars that test shows were never needed into the tier where the account captures real demand.
\n1. DuckDB Foundation. DuckDB documentation. Reference for SQL analysis of CSV and Parquet files with an embedded open source database. 2. pandas development team. pandas documentation. Reference for data manipulation and tabulation in Python. 3. Edelman, B., Ostrovsky, M. and Schwarz, M. Internet Advertising and the Generalized Second-Price Auction. American Economic Review. Analysis of how search ads are priced and ranked. 4. Varian, H. Position Auctions. International Journal of Industrial Organization. Foundation for understanding rank and price in search auctions.
"},{"id":"chapter-8-keyword-strategy-and-match-control","title":"Chapter 8: Keyword Strategy and Match Control","html":"By the end of this chapter, the reader will be able to: 1. Describe how exact, phrase and broad matching decide which searches trigger an ad, and where each one drifts. 2. Mine a search term report to separate queries worth buying from queries that only look relevant. 3. Build a negative keyword structure in layers, so one blunt exclusion does not silence good demand. 4. Calculate whether a long tail query group earns its management cost. 5. Classify queries by intent and assign each class a match approach, bid posture and landing page. 6. Design a keyword taxonomy with seasonal and local modifiers that a new teammate can read in 10 minutes.
\nThe account I inherited looked healthy on every dashboard. Click-through rate was strong. Cost per click was low. The previous manager had left a note that said, in effect, \"broad match is working, leave it alone.\" I left it alone for three weeks. Then I opened the search term report, the list of actual searches that triggered the ads, and read it line by line like a box score after a bad loss.
\nA good share of the spend had gone to people looking for jobs at the company, people looking for free versions of the product, students writing papers about the category, and a steady trickle of searches in a language the business did not serve. The clicks were cheap because nobody wanted what we sold. Cheap and useful are different words.
\nThat experience cost me a client conversation I still remember, and it taught me the assumption I now push against hardest: that a keyword list is a list of words you chose. It is not. A keyword list is a set of invitations. The search platform decides who actually walks through the door, and the match type is how much freedom you give it to decide.
\nMost brand managers treat match types as a setting to pick once. The real work happens afterward, in the weekly habit of reading what the platform did with your money and correcting it. The keywords you type are the smaller half of the strategy. The queries you exclude, the groups you build, and the intent you assign to each one are the larger half.
\nThis chapter is about taking that control back. It covers how matching behaves, how it drifts, how to mine your own data, and how to organize all of it so the account stays readable as it grows. Menus and settings differ by platform and change often, so use the platform's current help pages for exact steps.
\nThink of match types as a dial between precision and reach. At one end, exact matching says: show my ad when someone searches this thing, or something that means the same thing. At the other end, broad matching says: show my ad when someone searches anything the system believes relates to this thing. Phrase matching sits between them, asking that the core meaning of your phrase appear in the query, with extra words allowed around it.
\nNotice the language. I wrote \"means the same thing\" and \"believes relates,\" not \"spells the same way.\" Over the years the major platforms have loosened every match type. Exact now covers plural forms, misspellings, reordered words and queries that share the same intent in the system's judgment. The word \"exact\" has become a courtesy title more than a description.
\nThe practical consequence is that the match type tells you what the platform is allowed to do, not what it did. Only the search term report tells you what it did. Every decision in this chapter starts there.
\nA working rule I use: choose the tightest match type that still gives you enough volume to learn from. Tight matching earns its keep on expensive, high-intent terms where one wasted click hurts. Looser matching earns its keep on discovery, where you are paying for information as much as for conversions.
\nDrift is the gap between the query you imagined and the query that triggered your ad. Every match type drifts a little. Broad matching drifts by design, and in recent years the platforms have added automated expansion that reads your landing page, your existing ad copy and your account history to find searches you never listed. Some of those searches are real finds. Some are noise wearing a plausible costume.
\nAutomated expansion adds a second problem: reduced visibility. When a system chooses queries based on signals it does not fully disclose, you often see only a portion of what triggered the ad. The search term report may group low volume queries into an \"other\" bucket. You cannot exclude what you cannot see, so the unseen portion becomes a tax on your account.
\nI am not telling you to avoid broad or automated matching. I am telling you to price it correctly. Suppose you give a broad campaign a small, capped test budget, say 50 dollars a day, with strong negatives already in place and a conversion goal that reflects real business value, not a soft action like a page view. In that setup, drift is the cost of discovery, and you can measure it. Give the same freedom to a campaign with a vague conversion goal and no negatives, and you are paying the platform to teach itself with your money.
\nThe search term report is the closest thing paid search offers to a customer interview. People type exactly what is on their minds, in their own words, at the moment they want something. If you read it with curiosity, it tells you how buyers describe your category, what they fear, what they compare, and what they assume you sell.
\nStart by exporting a month of data. Sort by spend, not by clicks. High spend with no conversions is your first cleaning pass. Then sort by conversions and look at queries you did not list as keywords. Those are your growth pass. Finally, sort by impressions with low click-through rate. Those are queries where your ad shows but the searcher does not feel addressed, which often points to an ad copy mismatch more than a keyword problem.
\nA short illustrative table shows how one slice of a report might be sorted. The brand is fictional: a regional home security installer.
\n| Query | Intent class | Likely action | |---|---|---| | home security installation near me | Ready to buy, local | Exact match, high bid, quote page | | best home security systems | Comparing options | Phrase match, comparison page | | how to bypass a home alarm | Wrong audience, risky | Negative, no exceptions | | diy security camera setup | Self-service buyer | Negative or low bid test | | home security jobs | Job seeker | Negative | | apartment security system rules | Information seeker | Content campaign, low bid |
\nIf match types are the dial, negative keywords are the fence. They are the strongest control a buyer has, because they apply regardless of how the platform expands your reach. A tight keyword with a weak negative list leaks. A loose keyword with a strong negative list can behave better than a tight one.
\nBuild negatives in layers. The first layer is account level: terms that no campaign in the business should ever show for. Job seekers, free seekers, explicit content, and anything that could embarrass the brand belong here. The second layer is campaign level: terms that are wrong for this offer but right for another. A premium line may exclude \"cheap\" while an entry line welcomes it. The third layer is ad group level and does the finest work: it keeps similar groups from competing with each other for the same query.
\nThat third layer solves a quiet problem called keyword cannibalization, which is when two of your own groups bid on the same search and split the data. Suppose a broad group and an exact group both match \"emergency plumber.\" Your broad group may win the auction and bury the exact group's performance. Adding the exact keyword as a negative in the broad group steers the search to the group you designed for it.
\nNegative matching also has match types, and the mistake I see most often is applying them carelessly. A broad negative blocks a query only when every word in it appears, in any order. A phrase negative blocks queries containing the words in sequence. An exact negative blocks only that query. Adding the single word \"free\" as a broad negative is safe. Adding \"software training\" as a broad negative may block \"training for tax software,\" a search you wanted. Read your negatives before you save them. A good exclusion list is precise enough that you could defend each line aloud.
\nThe long tail is the large collection of rare, specific searches that individually bring almost no traffic and together bring a lot. \"Standing desk converter for tall people with a bad back\" will be searched a few times a month. A thousand queries like it can outweigh your headline keyword.
\nWhy do buyers like the long tail? Because specific searches usually carry specific intent, and specific intent often converts at a better rate with less competition. Advertisers fight over \"running shoes.\" Far fewer fight over \"running shoes for wide feet and plantar fasciitis,\" so the clicks cost less and the buyer is closer to a decision.
\nA worked example with round, illustrative figures. Suppose a fictional furniture retailer has 200 long tail queries that each bring about 10 clicks a month, so 200 times 10 gives a couple of thousand clicks. At a three percent conversion rate that is about 60 orders. If each order brings 90 dollars of gross margin, the group earns roughly fifty-four hundred dollars. At a dollar and 20 cents a click, spend is near twenty-four hundred, which leaves a healthy gap. The catch is labor: managing 200 keywords one at a time would eat hours every month, which is why grouping is the answer.
\nThe fix is to cluster long tail queries by theme, not to manage them one by one. Group queries that share a product, a problem or a modifier, give each group a shared bid and ad, and judge performance at the group level. You will not know whether one query converts, but you will know whether the theme does. Statistical clustering helps when the list is too large to sort by eye, and the callout later in this chapter names free tools for it.
\nIntent is what the searcher is trying to accomplish, and it is the most useful way to organize a keyword list. Four classes cover most commercial accounts.
\nNavigational queries seek a particular brand or site: your name, a competitor's name, a product line. Informational queries seek knowledge: \"how does a heat pump work.\" Commercial investigation queries compare options: \"best,\" \"versus,\" \"reviews,\" \"alternatives.\" Transactional queries signal readiness: \"buy,\" \"price,\" \"near me,\" \"quote,\" \"same day,\" \"coupon.\"
\nEach class gets a different treatment. Navigational queries about your own brand deserve exact match and a clean landing page, because losing them to a competitor is an avoidable embarrassment. Navigational queries about competitors are a risk decision: conquesting costs more per click, converts less often, and invites trademark complaints in some situations, so counsel should review your approach. Informational queries rarely belong in a lead generation campaign, though they can feed a low-cost content or audience building effort. Commercial investigation queries are where phrase match and comparison pages do well. Transactional queries are where you spend your best bids.
\nClassification can be done with rules first and models second. Begin with a simple lookup: if the query contains certain words, label it. This catches most volume with little effort. Then review the unlabeled remainder by hand and add rules. Only when rules stop paying off is it worth training a classifier.
\nFamily-first caution: for any category that touches children, health, money stress or personal safety, use conservative defaults. Exclude queries that suggest vulnerability or crisis unless your offer genuinely serves them and your claims have been reviewed.
\nA taxonomy is the naming and grouping system that keeps an account understandable.
\nBuild it in four levels. At the top, the business goal or product line. Next, the intent class. Below that, the theme, such as a problem, feature or audience. At the bottom, the match type. A campaign might be named \"Residential Security | Transactional | Local\" and contain groups named for themes like \"installation,\" \"monitoring,\" and \"emergency.\" Anyone reading the name can see what it does and what it should cost.
\nDocument the taxonomy on one page: naming rule, intent classes, match policy, negative layers and review schedule. That page is the real asset. Staff and agencies change, and a written taxonomy lets a new person understand the account in an afternoon.
\nSeasonal and local modifiers are words that change what a search means. \"Furnace repair\" is a service question in October. In July it may be a planning question. \"Coats\" is a commodity in summer and urgent in the first cold week.
\nTreat seasonality in two parts. The first is demand shape: plot your own search term data week by week and note when volume rises and falls. The second is intent shape: during peak periods, buyers are further along and willing to pay more, so bids and budgets should rise before the peak, not after it. A buyer who raises budgets only after the rush has started pays for the learning period at the worst moment.
\nLocal modifiers work the same way. \"Near me,\" a city name, a neighborhood, a zip code and \"open now\" all signal readiness to act in a specific place. For a business with a service area, local modifiers are often the most profitable keywords in the account, and they deserve exact or tight phrase matching. Pair them with location settings, because a keyword that says \"Denver\" shown to someone in another state wastes money no matter how well it matches.
\nFinally, treat modifiers as combinations. A seasonal word plus a local word plus a transactional word, such as \"same day furnace repair Fort Collins,\" is a long tail query with high intent. These combinations are small individually, and a template of modifiers multiplied across your services creates them at scale. Generate the list, then let the report show which were real.
\nSearch term report: The platform's record of the actual queries that triggered your ads, along with cost and results for each. Close variant: A query the platform treats as equivalent to your keyword, such as a plural, misspelling or reworded version with the same meaning. Negative keyword: A word or phrase you add to block ads from showing on matching searches. Query drift: The gradual movement of triggered searches away from the intent you planned for, common with broad and automated matching. Long tail: The mass of specific, low volume searches that together bring meaningful traffic at typically lower competition.
\n1. Leaving broad matching on without a review date, which turns a discovery test into a permanent, unmeasured expense. 2. Adding a single common word as a broad negative without checking which good queries contain it, which quietly blocks buyers. 3. Reading the search term report by clicks instead of spend, which hides the expensive queries that never convert. 4. Letting two groups bid on the same query, so your own campaigns compete and neither collects clean data. 5. Building seasonal groups after the peak starts, which pays the learning cost during the most expensive weeks.
\n1. Set a weekly 30 minute search term review and log every exclusion with a date and a reason. 2. Layer negatives at account, campaign and group level so each exclusion has a clear purpose. 3. Promote converting queries into their own exact match keywords with a tailored ad and landing page. 4. Label every keyword group with its intent class so budget can be reported by intent. 5. Document the taxonomy on one page, including naming rules, match policy and review schedule.
\n1. Sort the report by spend before anything else, because wasted money hides at the top of that list, not the bottom. 2. Keep brand terms in their own campaign with their own budget so a broad test cannot drain them. 3. Cluster long tail queries by theme and judge the theme, since single queries rarely hold enough data. 4. Check negatives quarterly against blocked traffic, where available, to catch exclusions that outlived their purpose. 5. Start seasonal groups several weeks before the expected peak so learning happens at a low cost.
\nThree free libraries cover the analysis work in this chapter. pandas (BSD license) loads a search term export into tables you can sort, filter and group, which handles the spend sorting and pattern counting described above. scikit-learn (BSD license) provides text vectorization and clustering methods that sort thousands of long tail queries into themes. spaCy (MIT license) handles language processing, including splitting queries into words, reducing them to root forms and tagging terms, which helps with intent rules and modifier detection. None of these replaces a full demand-side platform or the ad platform's own reporting, and they do not need to. They read the data you already export and make it usable.
\nMatch types set how much freedom the platform has, and drift shows how it used that freedom. The search term report is the honest record, and mining it turns spend into learning. Negative keywords, built in layers, are the main control lever, and long tail queries pay off when grouped by theme instead of managed one at a time. Intent classes tell you how to bid, which match type to use and what page to show, while a documented taxonomy keeps the whole account readable. Seasonal and local modifiers add timing and place, and the buyer who prepares for them early pays less to learn.
\n1. A match type states what the platform may do, and only the search term report states what it did. 2. Broad and automated matching are discovery costs: cap them, separate them and give them a review date. 3. Negative keywords applied in layers are the strongest control in a keyword program, and they need regular review. 4. Long tail queries work when clustered by theme and judged at the group level. 5. Intent classes and a written taxonomy let you assign bids, match types and pages with a reason.
\nThis week, export 30 days of search terms, sort by spend, and add every wrong-audience query as a layered negative with a dated note, then give broad match its own capped budget and a review date.
\n1. Brodersen, K. H., et al. Inferring causal impact using Bayesian structural time-series models. Annals of Applied Statistics. Published methodology for judging whether a keyword change moved outcomes. 2. Federal Trade Commission. Disclosures 101 for Social Media Influencers and Dot Com Disclosures guidance on clear advertising claims. 3. pandas project documentation, data analysis library (BSD license). 4. scikit-learn project documentation, text feature extraction and clustering (BSD license). 5. spaCy project documentation, linguistic features and tokenization (MIT license).
"},{"id":"chapter-9-search-copy-and-landing-page-alignment","title":"Chapter 9: Search Copy and Landing Page Alignment","html":"By the end of this chapter, the reader will be able to: 1. Describe the scent trail that connects a search query, an ad, and a landing page. 2. Write search ad copy that qualifies the right buyer and discourages the wrong one. 3. Explain at the concept level how assets and extensions support the promise an ad makes. 4. Calculate the cost of a message mismatch using conversion rate and cost per click. 5. Design a copy test with one variable and a sample large enough to trust. 6. Evaluate landing page speed as a direct ad cost using free, open source measurement tools.
\nWho owns the landing page at your company? Not who built it. Who is accountable for what it says to a person who just clicked a paid ad?
\nIf you hesitated, you are in the majority, and that hesitation is expensive. In most organizations the ad lives in one place and the page lives in another. A media buyer writes the headlines inside the ad platform. A web team or an agency owns the site. Each side does competent work, each side reports its own numbers, and nobody owns the gap between them. The buyer sees a healthy click rate and calls the ad a success. The web team sees traffic and calls the page a success. The business sees a cost per sale that keeps drifting upward and nobody can explain why.
\nThe common assumption is that a good ad is one that gets clicked. It is not. A click is a cost, not a result. An ad that attracts clicks from people the page cannot satisfy is simply a faster way to spend budget. The most dangerous ad in your account may be the one with the best click rate, because it is the one nobody questions.
\nThink of a ballpark with a gorgeous outdoor billboard promising a tailgate party at the north gate. Fans follow the sign, arrive at the north gate, and find a locked chain and a maintenance cart. The billboard did its job. The experience destroyed the trust. Nobody blames the billboard painter, but every fan remembers the locked gate.
\nThis chapter treats the ad and the page as one piece of communication that happens to be split across two screens. I will show you how to keep the promise intact from the first word of the query to the final button, how to write copy that turns away people you cannot serve, how to count what a mismatch costs, and how to test changes without fooling yourself. Operational guidance only. Your legal counsel decides what claims you may make.
\nUsability researchers borrowed a term from foraging animals: information scent. A person hunting for something follows cues that suggest they are getting closer. A fox tracking a rabbit does not stop to read a map. It checks whether the smell is getting stronger. Searchers behave the same way. They typed a few words, and every screen after that either strengthens the scent or breaks it.
\nThe trail has three links. The query is the person's own words. The ad is your first reply. The landing page is your second reply, and it must sound like the same speaker as the first. If a person searched for \"emergency plumber open now\" and your ad says \"24 hour emergency plumbing, call in minutes,\" the scent is strong. If the page opens with a gallery of bathroom remodels and a form for scheduling a consultation next week, the scent is gone in about two seconds.
\nMessage match is the name for how well those links agree. It is not repetition. It is consistency of promise. Two lines can share no words and still match, and two lines can share every word and still contradict.
\nMost copy advice teaches you to attract. Make the headline compelling, add urgency, stack the benefits. That advice treats every click as a win. Your budget does not agree.
\nGood search copy does two jobs. It attracts the person you want, and it quietly discourages the person you do not. Every click from someone who will never buy is a small tax on the people who will. If you sell a premium service, a headline that hints at the price level saves you from paying for clicks from bargain hunters. If you only serve certain areas, naming them in the copy keeps out the rest. If you sell to businesses, saying so keeps out hobbyists.
\nYour headline should reflect what the person typed, in their language, plus one specific benefit. Your description should add one supporting detail and a clear next step, never a second promise that the page does not repeat. Resist the urge to use every available character. A tight ad that says one thing clearly beats a crowded ad that says five things vaguely.
\nA word about honesty, because it is also a conversion tactic. Claims about price, shipping, guarantees, results, and availability must be true and visible on the page. The Federal Trade Commission expects advertisers to be able to substantiate claims and to make material conditions clear. If your ad says \"free shipping\" and the page reveals a minimum order on the third screen, you have a legal question and a conversion problem at the same time. Let counsel decide the exact wording. Operationally, put the condition in the ad when it is short, or put it on the page at the point of the headline.
\nEvery major search platform lets you attach extra pieces to a basic text ad. The names change over time, so I will describe the concepts rather than the menu labels. You will see these ideas under different names in whatever platform you use.
\nThink of a basic ad as a storefront sign. Extra pieces are the window displays. A set of extra links lets you point to specific pages inside your site, so a person looking for pricing can skip straight there. Short callouts let you state plain facts such as a return policy or hours. Structured lists let you name categories, models, or services. Location and call details let people reach you without visiting a page at all. Image and business name assets add recognition and trust.
\nThe rule that governs all of them is the same rule that governs the headline: each extra piece makes a promise, and each promise must be kept somewhere on the destination. Extra links are the most commonly broken. A person clicks a link labeled \"Pricing\" and lands on a page where pricing requires a sales call. A callout claims \"Free returns\" and the returns page buries conditions. Every extra piece is another trail to maintain.
\nA landing page for paid search has one job. It must confirm, within the first screen, that the person is in the right place, and it must make the next step obvious.
\nStart with the headline. It should echo the ad's promise in the person's own words, not copy it letter for letter, but carry the same offer and the same subject. Then provide a single supporting line that makes the offer concrete. Then offer one primary action. One. A page that offers a phone number, a chat widget, a download, a demo request, and a newsletter signup is asking the visitor to design their own path, and most will leave instead.
\nMatch the action to the intent. A person who searched \"how to compare payroll software\" is gathering information. Greeting them with a credit card form breaks the trail. A person who searched \"buy standing desk\" is ready, and a long educational article delays them. Informational queries deserve helpful, scannable content with a gentle next step. Transactional queries deserve a fast path to the purchase or the form.
\nThe following is an illustrative composite. The brand is fictional.
\nCorbel Dental is a single-location practice that wants new patients seeking dental implants. The first version of its search program looked reasonable. The query \"dental implants cost\" triggered this ad: \"Corbel Dental. Quality Care for Your Whole Family. Visit Us Today.\" The ad pointed to the practice homepage, which opened with a rotating banner about teeth cleaning, a welcome message, and a menu with eleven items.
\nSuppose that in one month the ad group received 800 clicks at an average of four dollars each, for 3,200 dollars of spend. Suppose 16 of those visitors booked a consultation, a conversion rate of two percent. That is 200 dollars per booked consultation. Illustrative figures, but the structure is typical of a mismatch.
\nNow the rebuild. The ad group is narrowed to implant-related queries only. The new headline reads: \"Dental Implants: Cost Explained.\" The description reads: \"See typical price ranges, financing options, and what is included. Free implant consultation for adults.\" The phrase \"for adults\" qualifies out clicks the practice cannot serve. The ad uses extra links labeled \"Financing Options,\" \"Before and After Gallery,\" and \"Meet the Implant Dentist,\" and a callout: \"Free Consultation.\"
\nThe landing page is a dedicated page. Its headline echoes the promise: \"Dental Implants: What They Cost and What You Get.\" The first screen offers a short explanation of price ranges in plain language, a note that exact costs depend on the exam, a financing summary, and a single button labeled \"Book Free Consultation.\" Credentials and a review summary sit just below. The form asks for name, phone, and preferred time. Counsel reviewed the price language.
\nSuppose the next month brought 500 clicks, because qualification removed the browsers. Spend at the same four dollars per click is two thousand dollars. Suppose 25 visitors booked, a conversion rate of five percent. Now each booked consultation costs 80 dollars. The practice spent 1,200 dollars less and booked nine more consultations. Notice what moved. Not the bid, not the budget, not the keyword list. The only change was the quality of the conversation between the person and the page.
\nYou can estimate what a mismatch costs without any special software. You need three numbers: your cost per click, your conversion rate, and your target cost per conversion.
\nCost per conversion equals cost per click divided by conversion rate. At four dollars per click and a two percent conversion rate, you pay 200 dollars. Raise the conversion rate to three percent with better alignment and the same click costs 133 dollars per conversion. Raise it to four percent and it costs 100 dollars. Conversion rate is the lever that makes your budget stretch, and message match is one of the strongest forces behind it.
\nA slow page is a hidden surcharge on every click. You pay for the click the moment it happens. If the page takes too long to appear, a share of those people leave before they see a single word. They never register as a bounce in some analytics tools because the page never finished loading, so the loss can be invisible.
\nThe logic is plain. Suppose 1,000 people click and you pay four dollars each, 4,000 dollars. If a slow page causes 10 percent to abandon before the content appears, you paid 400 dollars for nothing and your conversion rate calculation starts from a damaged base. Speed is not a technical vanity project. It is a media efficiency project.
\nThree free, open source tools cover the work. Lighthouse (Apache 2.0) audits a single page and flags the largest delays, including oversized images, scripts that block rendering, and layout jumps. Run it on the mobile setting first. GrowthBook (MIT core) belongs in the testing section below. sitespeed.io (MIT) runs repeated tests on a schedule from your own machine or server, so you can see whether a recent change made the page slower.
\nMost copy tests are not tests. They are coin flips that someone explained afterward.
\nA real test changes one thing. If you change the headline, the description, and the button in the same variant, and the variant wins, you have learned nothing about why. You may have a winner, but you cannot repeat it. Change one element at a time: the headline, or the qualifying phrase, or the action label. Keep everything else fixed.
\nThen wait for enough data. Suppose version A converts at four percent and version B at five percent after 200 clicks each. That is eight conversions against 10. Two conversions of difference is well within the range that luck alone can produce. A team that declares B the winner at that point will be wrong about half the time. As a rough operating rule, wait until each version has collected enough conversions, not just clicks, that a difference of a few of them would not change your conclusion. In practice many accounts want at least 100 conversions per version for a modest lift, and more for a small one. A statistical calculator, or the testing engine in GrowthBook (MIT core), will compute the confidence for you. If your account lacks the volume, test bigger changes, or test on the page where traffic is concentrated.
\nMessage match: The degree to which an ad and its landing page communicate the same promise, offer, and tone.
\nInformation scent: The set of cues that tells a person they are getting closer to what they wanted, so they keep going instead of leaving.
\nQualifying copy: Ad language that discourages clicks from people the business cannot serve, lowering wasted spend.
\nAssets (extensions): Optional pieces added to a text ad, such as extra links, callouts, and structured lists, that add detail and must be honored on the destination.
\nCore web vitals: A set of page experience measures covering load time, responsiveness, and visual stability, used to judge how a page feels to a visitor.
\n1. Sending every ad group to the homepage. A homepage speaks to everyone, so it answers no one's specific query, and visitors must hunt for the page the ad implied. 2. Optimizing for click rate alone. A rising click rate on copy that attracts the wrong audience raises spend without raising sales, and the dashboard rewards the mistake. 3. Changing the ad and the page at different times without a record. When conversion moves, nobody can say which change did it, and the lesson is lost. 4. Calling a test after a handful of conversions. Small samples swing wildly, and an early lead often disappears with more data. 5. Letting extra links and callouts go stale. Renamed, redirected, or retired destinations send paid clicks to dead ends that no summary report flags.
\n1. Write the ad and the landing page in the same working session, so the promise and the fulfillment have one author and one owner. 2. Build a separate page for each distinct intent, and map query themes to pages in a simple sheet that anyone on the team can read. 3. Include at least one qualifying phrase in every ad, such as audience, place, or next step, and check that the page repeats it. 4. Review the search terms report on a set schedule, and compare actual queries to the ad and the page before touching bids. 5. Test one element at a time, set the decision rule in writing before starting, and record every result in a shared log.
\n1. Read your ad aloud, then read the first screen of the page aloud. If they sound like two different companies, fix the page first. 2. Check every landing page on a real phone over a real mobile connection at least once a month, not only in a desktop preview. 3. Put the primary action in the first screen, and make its label match the verb used in the ad. 4. Ask sales or support what prospects say they expected before they called. Their answers expose broken promises faster than any report. 5. When you want faster pages, fix images and remove unowned scripts before buying any new technology. Those two steps usually deliver most of the gain.
\nThree free tools help with this work. Lighthouse (Apache 2.0) audits a page for speed, accessibility, and common layout problems, and it is the quickest way to see what is slowing a landing page. sitespeed.io (MIT) repeats those tests on a schedule, so you can catch a slowdown the week it appears instead of after a quarter of wasted clicks. GrowthBook (MIT core) runs controlled tests on headlines and page variants and calculates whether a difference is real. All three can be self-hosted at no license cost, which keeps testing under your control. No other software is needed for the practices in this chapter.
\nSearch advertising works only when the person's query, the ad, and the landing page tell one consistent story. That consistency, called message match, depends on keeping a scent trail unbroken: the offer, the action, the audience level, and the tone must agree from the first click to the final button. Good copy qualifies as well as attracts, steering away clicks the business cannot serve. Extra links and callouts are additional promises that must be kept. A mismatch has a measurable price, because cost per conversion equals cost per click divided by conversion rate, and a better page can stretch a budget more than a lower bid can. Slow pages add a hidden surcharge to every click. Honest tests change one variable, run to a pre-set sample, and judge by the conversion that pays.
\n1. The ad and the page are one message split across two screens, and someone must own the whole. 2. A click is a cost. Copy that qualifies can lower click rate and still lower cost per customer. 3. Every extra link and callout makes a promise that the destination must keep. 4. Conversion rate moves cost per conversion as strongly as bid changes do, and often with fewer side effects. 5. A test with several changes, a tiny sample, or a stopping point chosen mid-run proves nothing.
\nName one person who owns the path from query to final button, then have that person click your top five ads on a phone this week and fix the first broken promise found.
\n1. Federal Trade Commission. Dot Com Disclosures: Information about Online Advertising. Guidance on clear presentation of claims and conditions in digital advertising. 2. World Wide Web Consortium. Web Content Accessibility Guidelines, version 2.1. Standard for accessible web pages, including landing pages. 3. Google Chrome Developers. Lighthouse documentation. Reference for the open source auditing tool and its performance measures. 4. GrowthBook. GrowthBook documentation. Reference for open source experimentation and statistical analysis. 5. sitespeed.io. sitespeed.io documentation. Reference for open source repeated web performance testing.
"},{"id":"chapter-10-brand-defense-and-competitor-conquesting","title":"Chapter 10: Brand Defense and Competitor Conquesting","html":"By the end of this chapter, the reader will be able to: 1. Explain why paying for clicks on your own name can be either insurance or a tax on customers you already had. 2. Design a holdout test that separates branded clicks that were caused by the ad from clicks that were merely captured by it. 3. Interpret holdout results and decide whether a branded campaign should be kept, trimmed or switched off. 4. Identify the questions a brand manager must put to counsel before using a trademark in ad text or keywords. 5. Evaluate whether competitor conquesting makes economic sense after relevance, cost and retaliation are counted. 6. Apply a decision tree that connects defensive organic work, branded bidding and conquesting into one budget choice.
\nWhy pay an auction to deliver a customer who was already at your front door? Bidding on your own brand name is a waste. Think about who types that name into a search box. Someone who already knows you, already wants you, and is one click from your site. You are paying an auction to deliver a customer who was walking through your front door anyway. Conquesting is no better. You pay premium prices to show an ad to a person who typed in someone else's name, which means they came looking for someone else. Low relevance, high cost, and a rival who can do the same to you tomorrow.
\nMany seasoned media buyers believe exactly this, and they are often right. Here is the part that makes the subject hard: they are sometimes wrong, and the cost of being wrong lands in different places depending on the market. A brand with a crowded category and aggressive rivals can lose real orders when it goes dark on its own name. A brand alone on its results page can cut a large line item and see nothing change but a better margin.
\nSo the honest question is not whether brand defense is good or bad. The honest question is what happens to revenue in your account if you stop. That is an empirical question with an empirical answer, and you can get it in a few weeks for a modest cost.
\nMost branded campaigns look spectacular in reporting. Conversion rates are high, cost per order is low, and the platform dashboard glows. Those numbers describe who clicked, not who was persuaded. A branded campaign can show the best return in the account while adding almost nothing the business would not have earned anyway.
\nThis chapter gives you a way to find out. We will cover the holdout test, how to read it, the trademark questions that belong with your lawyer, the real economics of conquesting, the organic work that makes defense cheaper, and a decision tree you can hand to your team. Operational guidance only. Counsel decides legal questions.
\nStart with what the ad is actually buying. When your organic listing already holds the first position for your name, a branded ad takes a click that was already coming and attaches a fee to it. The fee is small per click. Across a year, it adds up to a line item that can rival your non-branded spend.
\nThere is also a reporting problem. Branded campaigns often absorb credit for conversions that began elsewhere: a podcast mention, a sports sponsorship, a retail shelf, a friend's recommendation. The customer heard your name somewhere, typed it in, and clicked the ad at the end. The ad gets the credit. The earlier work gets nothing.
\nThat distortion can steer an entire budget. If branded search is reported as the most efficient channel, someone will want to move money toward it. The money moves, the brand-name volume does not grow (it cannot, because it depends on awareness created elsewhere), and the total result flattens. I have watched teams cut upper-funnel work to fund a campaign that was only harvesting what that work had planted.
\nSo the default suspicion is reasonable. If nobody is bidding against you, your organic result is strong, and your page loads fast, the burden of proof sits with the ad. Make it earn its place. Do not fund it on faith, and do not fund it because it looks good on a dashboard.
\nNow the other side. The result page is a small piece of real estate, and the ad unit takes up a lot of it on a phone. A rival sitting above your organic listing, with a headline offering a discount, may collect a share of people who typed your name. Some of those people were never firmly committed. Shoppers who search a brand name are often comparing, and a strong offer in the top slot can pull a few away.
\nOther reasons to hold a defensive campaign exist as well. Ad text lets you control the message in ways an organic snippet does not: a current promotion, a store-hours line, a specific product page, or a required disclosure in a regulated category. If your name is shared with another company or has a common meaning, an ad can steer people to the right destination. And a brand with many product lines can send each search to the matching page instead of a homepage.
\nThink of it like a stadium with one concession stand outside the gate. If a competitor opens a cart at your entrance and you have no cart, some fans buy from them without ever looking at your menu. If nobody sets up a cart outside, your own stand inside gets all the business anyway, and a second cart out front just takes cash from the same fans.
\nThat is the argument in one image. The ad matters in proportion to how much someone is trying to intercept your traffic. Which brings us to measurement, because intuition about interception is cheap and often wrong.
\nA holdout test pauses branded ads in some places and keeps them running in others, then compares total results. Total is the important word. You are not comparing branded-campaign conversions, which will certainly fall to zero in paused areas. You are comparing all orders, calls or sign-ups from those areas, from every source, against matched areas where ads stayed on.
\nBuild it this way:
\nGeographic splits matter because you cannot reliably pause ads for half the visitors in a single market. Some platforms offer experiment tools that help, but the cleanest approach is by location.
\nBefore launching, get internal agreement on the rules. Decide in advance what size of drop would justify keeping the campaign and what size would not. Teams that skip this step argue about the result afterward, and the loudest voice wins.
\nOne caution: a holdout answers the question for the market and the period you tested. If a new competitor enters, retest.
\nSuppose Ashgrove Outfitters, an illustrative camping retailer, spends 14,000 dollars a month on branded search across 20 markets. It pauses branded ads in 10 matched markets for six weeks. Using a statistical tool such as CausalImpact, which builds a forecast of what the test markets would have done without the pause, the team compares predicted against actual total orders.
\nThree outcomes are typical, and each leads somewhere different.
\nFirst outcome: total orders in the test markets stay inside the forecast range. Paused branded clicks flowed to the organic listing. The campaign was capturing, not creating. The team trims branded spend to a small maintenance level, or off entirely, and moves the savings to non-branded work, where new customers are found.
\nSecond outcome: total orders fall noticeably and a rival's ad appeared on the branded results page during the test. Some customers were intercepted. The campaign is doing real defensive work, and the cost per saved order is what matters. If that cost is below the margin on an order, keep it.
\nThird outcome: orders fall a little, with no rival present. People who needed an ad to find the right page clicked away, or the organic snippet was weak. Fix the snippet and landing page first, then retest.
\nKeep expectations honest. A holdout gives an estimate with a range, not a precise figure. Treat a result inside the noise as no evidence of harm, not proof of zero effect.
\nThis section lists questions, not answers. Operational guidance only; counsel decides.
\nThree separate things can involve a trademark in search advertising. Each carries different risk, and platforms treat them differently.
\nThe first is bidding on a competitor's name as a keyword. In many places this is permitted by the platforms, though rules vary by country and over time. The second is using a competitor's name inside your ad text. That raises the risk considerably, because it can suggest an affiliation or confuse readers. The third is a rival bidding on your name, and then using your mark in their ad. Platforms often have a complaint process for that, and you should know how to file one.
\nQuestions to bring to counsel before launch:
\nAffiliates deserve special attention. Without written rules, a partner may bid on your brand name and collect commission on customers you already own. Put the limits in the contract, and audit the results page on a schedule.
\nFinally, keep comparative claims truthful and provable. Counsel decides what is permissible, so route every competitor reference in ad text through review.
\nConquesting means advertising to people who searched for a competitor. The logic is simple: some of those shoppers are comparing, and a good offer might move them. The economics are less friendly, and three forces work against you.
\nRelevance first. Platforms score ads on how well they match the search. Your ad does not match a competitor's name, so quality scores tend to be low. Low scores mean you pay more for the same position, and sometimes you cannot win the position at all. The searcher's intent also works against you. Someone typing a specific name usually wants that specific company.
\nCost second. Higher click prices, lower click-through rates and lower conversion rates stack on top of one another. Suppose clicks on a rival's name cost three times what you pay on your generic terms, and the visitor is half as likely to buy. The cost per order is then roughly six times higher before anyone counts the customer's lower loyalty.
\nRetaliation third. Auctions are visible markets, and rivals watch. Start conquesting and a competitor may respond by bidding on your name, which forces you into a defensive campaign you did not previously need. Both of you then spend more for the same customers.
\nConquesting can still make sense in narrow cases. You need a real, provable advantage, a competitor whose customers are actively unhappy, and a lifetime value high enough to absorb the acquisition cost. Test it in a small, capped campaign. Measure with a holdout as well, and judge it on customer value over time, not the first order.
\nBefore paying to defend a name, make the free listing as strong as possible. This is where the cheapest wins live.
\nCheck what appears when you search your own name on a phone. Is your site first? Do sitelinks (the extra links beneath the main result) point to your most-wanted pages? Is your business profile accurate, with correct hours, address and phone number? Do review sites, social profiles and press mentions show up beneath you, or are unfavorable pages ranking higher than they should?
\nStrengthen the page itself. Give it a clear title, a plain description, and fast loading. Use structured data so search engines understand your products, ratings and locations. Publish pages for the questions people ask about you: pricing, returns, comparisons, support. If a rival or an affiliate holds a page ranking for your name, consider whether a better page of your own can outrank it.
\nStrong organic results make the holdout come out in your favor, in the sense that they lower the need for paid defense. They also protect you if the paid campaign is paused, if a budget is cut, or if a platform changes a rule. Owned results do not depend on an auction.
\nUse this sequence, one question at a time.
\n1. Is any rival, affiliate or reseller showing an ad on your branded results? If no, run a holdout before spending. If the holdout shows no loss, keep branded spend at a small maintenance level or off. 2. If yes, is your organic result strong and first? Fix any weakness, then run the holdout. If orders fall when you pause, fund the defense up to the point where cost per saved order stays below order margin. 3. Is the rival using your trademark in their ad text? Send the evidence to counsel and use the platform's complaint process. 4. Is conquesting under consideration? Confirm a provable advantage, review the ad text with counsel, set a small cap, and decide in advance what customer value would justify continuing. 5. Did a rival retaliate? Return to the first question and retest.
\nReview the whole tree every quarter, and also after any competitor launch, major promotion or platform policy change. The right answer in one season is not guaranteed in the next. Treat each result as an estimate that expires.
\nBrand defense: Bidding on your own name so your ad, not a rival's, appears on the results page for it. Conquesting: Bidding on a competitor's name to show your ad to people searching for that company. Holdout test: An experiment that pauses a campaign in matched areas to measure what total results change. Incrementality: The portion of results that happened because of the ad, not merely alongside it. Quality score: A platform rating of ad relevance and expected performance that influences price and position.
\n1. Judging a branded campaign by its own conversion rate, which mostly reflects customers who already chose you. 2. Pausing branded ads for a week and calling the result final, without matched markets or a long enough window. 3. Putting a competitor's name in ad text without counsel's review, which raises legal and platform risk well beyond keyword bidding. 4. Starting a conquesting campaign with no provable advantage, then raising bids when results disappoint. 5. Leaving affiliates without written bidding limits, so a partner collects commission on customers who were already yours.
\n1. Run a geographic holdout on branded search before setting a defense budget, and write the decision threshold down in advance. 2. Measure total outcomes from first-party analytics, not from the ad platform that is being tested. 3. Fix your organic listing, sitelinks and business profile before paying to defend them. 4. Cap conquesting spend, review its ad text with counsel, and judge it on customer value over time. 5. Review the defense and conquest decisions together every quarter and after any competitor change.
\n1. Check your branded results page on a phone, at different hours, and from several locations, since rival ads can come and go. 2. Keep branded ad text tied to current offers and specific pages so the ad does work the organic snippet cannot. 3. Save screenshots with dates whenever a rival uses your mark, because platform complaints move faster with evidence. 4. If a holdout shows no effect, move the savings to non-branded search or awareness work, then watch total orders. 5. Retest after a new competitor enters, since a result that held last quarter can reverse.
\nCausalImpact (Apache 2.0) is an open source package that forecasts what a set of test markets would have done without a change, then compares that forecast to what happened. It suits holdout analysis well, because it gives a range instead of a single number. Metabase (AGPL) is open source dashboard software you can run yourself. Connect it to your own order data and chart total outcomes for test and control markets side by side, with the branded results audit alongside. No fully open source tool tracks competitor ads on results pages at scale, so that check stays manual for now.
\nBranded bidding looks efficient because it counts customers who were already coming, so the real question is what happens to total results when you stop. A geographic holdout answers it, and the answer differs by market: a brand with active interceptors may need defense, while a brand alone on its page may be paying for nothing. Trademark use in keywords and ad text carries risks that belong in front of counsel, and affiliates need written limits. Conquesting faces low relevance, high cost and likely retaliation, so it fits only narrow cases with a provable advantage and strong customer value. Strong organic results lower the need for paid defense, and a quarterly decision tree keeps both programs honest.
\n1. High branded conversion rates describe who clicked, not who the ad persuaded. 2. A matched-market holdout comparing total outcomes is the only reliable way to size the value of branded bidding. 3. Trademark questions about keywords, ad text and affiliates should go to counsel before launch, not after a complaint. 4. Conquesting pays only with a provable advantage, a capped test and a customer value that absorbs a high acquisition cost. 5. Strong organic listings and a quarterly review keep defense spending proportional to real risk.
\nPause branded search in half of your matched markets for six weeks, compare total orders instead of branded conversions, and set the defense budget at the cost per saved order your margin can bear.
\n1. Brodersen, K.H., Gallusser, F., Koehler, J., Remy, N., Scott, S.L. Inferring causal impact using Bayesian structural time-series models. Annals of Applied Statistics. Methodology behind the CausalImpact package. 2. CausalImpact package documentation. Apache 2.0 licensed software for estimating the effect of an intervention on a time series. 3. Metabase. Open source documentation for self-hosted dashboards. AGPL licensed. 4. Federal Trade Commission. Advertising and Marketing on the Internet: Rules of the Road. Guidance on truthful and substantiated advertising claims. 5. United States Code, Title 15, Sections 1114 and 1125. Federal trademark infringement and false designation provisions, for discussion with counsel.
"},{"id":"chapter-11-local-and-shopping-search","title":"Chapter 11: Local and Shopping Search","html":"By the end of this chapter, the reader will be able to: 1. Describe how a local listing and a product feed each act as the targeting layer for their format. 2. Explain why category choice, service-area settings, and address data change which searches a business can win. 3. Apply a feed hygiene checklist to titles, identifiers, availability, and price fields. 4. Evaluate reviews as a paid-media asset and set a conservative policy for requesting them. 5. Construct schema.org structured data that matches a listing or feed record. 6. Judge the limits of store visit measurement and design a holdout that tests real lift.
\nPicture a regional garden supply brand, fictional and illustrative, called Fernlight Outdoor. The marketing lead looks at last quarter and sees that shopping campaigns are working. Sales are up, spend is steady, and the agency report is tidy. So she makes a sensible decision: raise the daily budget to cover spring. Six weeks later, cost per sale has climbed, two bestselling products have stopped showing, and a customer has driven 30 minutes to a store that was closed for a holiday the listing never mentioned.
\nNothing in the ad account was broken. The problem lived upstream. A price update had reached the website but not the feed. A product number had been reused across two sizes of the same planter. The store hours were wrong for one location. More budget simply bought more of the damage.
\nHere is the assumption worth challenging: that local and shopping search are campaigns you manage inside an ad platform. They are not. They are data programs that happen to have an ad platform attached. In a text ad, you choose the keywords and write the words. In these formats, the platform reads your records, decides what you sell and where you sell it, and then decides which searches you deserve. Your bid is a vote. Your data is the ballot.
\nThat shifts where a brand manager should spend attention. The hours you would normally spend on bid strategy are better spent on whether the business name, category, address, price, and availability are right in every system that feeds a listing. This chapter walks through that work: what to fix, in what order, how to describe it to people who do not live in spreadsheets, and how to measure the result without fooling yourself. The account mechanics live in each platform's own documentation. Here, the focus is the layer underneath.
\nWhen someone searches for a plumber or a pair of trail shoes near them, the platform does not read your website the way a person does. It reads structured fields. For a local business, those fields live in a business profile: name, address, phone, hours, categories, attributes, photos. For a retailer, they live in a product feed: one row per item, with a title, price, availability, and identifier.
\nThink of a listing as a player card in a fantasy sports league. The scouting report decides when you get picked. A card that says \"guard\" gets considered for guard slots, whatever the player can actually do. If your records say you are open when you are closed, or in stock when you are not, the platform will keep placing you in those slots and the shopper will pay for the mistake with a wasted trip.
\nThis is why teams that treat the listing as set-and-forget struggle. Hours change, products sell out, prices move, locations open and close. Each change has to travel from the system where it happens (a point of sale, a catalog, a manager's phone) into the system where the platform reads it. Every handoff is a place for drift.
\nAssign a named owner. Not a department, a person. That person checks the listing against reality on a fixed schedule, and has authority to correct it without a committee. In many small organizations this single decision does more for performance than any bid adjustment.
\nCategories are the most underrated field in a local profile. A primary category tells the platform what kind of business you are. Secondary categories tell it what else you do. Pick too broad a primary category and you compete with everyone. Pick one that does not match what customers actually ask for and you vanish from the searches that matter.
\nA simple method: list the 10 phrases customers use when they call you, then pick the primary category that most of those phrases imply. Add secondary categories only for services you genuinely perform at volume. A bakery that sells a few cakes a month should not claim to be a wedding venue.
\nThen comes the storefront question. A storefront business has a public address where customers show up. A service-area business travels to the customer, such as a plumber, a cleaner, or a mobile mechanic. Mixing these up causes real trouble. A service-area business that publishes a home address as a public storefront invites privacy problems and confuses customers who knock on the door. A storefront business that hides its address and lists only a service radius loses the walk-in traffic it was built for.
\nFor businesses with both, such as a flooring shop that also installs, keep the public address for the showroom and define the service area separately. For a service-area business with no public location, hide the address and define the area by the places you will really drive to. Do not claim 30 towns because you once took a job in one of them. Platforms and customers both notice, and conflicting listings across directories weaken trust in all of them.
\nName consistency matters too. The business name should match the signage and the legal name, without stuffed keywords. Adding a city or a service to the name may look clever. It often creates a policy violation that can suspend the listing.
\nA review is a free piece of creative that a stranger wrote about you, and it sits next to your ad in the place where the decision happens. In a close comparison between two nearby businesses, the rating and the review count often carry more weight than your headline ever will.
\nTreat reviews as an asset with three parts: volume, recency, and your responses. Volume gives the rating stability, since one angry customer swings a rating built on 10 reviews far more than one built on 300. Recency tells a shopper that the place is still operating at the quality described. Responses show the next reader how you handle a problem, which matters even more than the problem itself.
\nHere is an illustrative case. A tire shop averages four reviews a month. One bad week produces three one-star reviews about a long wait. With a low base, the average drops visibly and calls slow down. The shop owner who replies to each one, stating plainly what happened and what changed, usually recovers faster than the one who stays silent or argues.
\nAsk for reviews in a way that holds up to scrutiny. Ask every customer, not only the happy ones. Do not offer payment, discounts, or gifts in exchange for a review, and do not write reviews yourself or ask staff and relatives to do so. Rules from the platforms and from the Federal Trade Commission treat fake or incentivized reviews as a serious problem. Operational guidance only here, and counsel decides what your policy should say.
\nFamily-first standard: never solicit reviews from minors, and for any business serving children, route requests to the parent or guardian.
\nIn shopping formats there are no keywords to choose in the way text ads allow. The platform matches a search to a product by reading your title, description, category, brand, and attributes. So the feed is the targeting. If the title says \"Planter\" and the shopper typed \"12 inch glazed ceramic planter with drainage,\" you are guessing at relevance while a competitor with a fuller title is declaring it.
\nSuppose Fernlight sells 400 products and the feed was exported from the website once, in bulk. Titles read like internal labels: \"CER-PL-12-GL.\" The fix is not a bigger budget. It is a title pattern the platform can read. A reliable pattern is brand, product type, defining attributes, size or count. Put the words a shopper would type first, because titles are often truncated on small screens.
\nCategories need the same care. Product category fields map each item to a taxonomy, which is a standard list of product types. A wrong category sends the item to the wrong auction. Attributes such as color, size, material, and gender or age group refine the match, and apparel in particular often requires them. Be accurate, never creative. Mislabeling an adult item as a children's item, or the reverse, is a policy problem as well as a targeting problem, and the conservative default is always the stricter label.
\nBecause the feed carries the targeting, segment your reporting by feed attributes: brand, category, price band, margin tier. Many accounts discover that a handful of products produce most of the profit while the rest absorb spend. You can only see that if the feed contains the fields to slice by. Add a custom label for margin before you need it, not after.
\nHygiene is the unglamorous work that keeps the program alive. Four fields cause most of the damage.
\nTitles. Keep them specific and honest. Do not add promotional text such as \"free shipping\" or \"best deal\" to a title, since it breaks policy and clutters the match. Do not repeat the brand three times. Check for truncation and for internal codes that mean nothing to a shopper.
\nIdentifiers. Each product needs a unique item ID that never changes and is never reused. Where a manufacturer assigned a global trade item number (GTIN), the standard barcode number, include it. Matching to known products depends on it. Two sizes sharing one ID is exactly what happened to Fernlight, and it makes reporting meaningless. If a product has no barcode, such as a handmade item, say so in the identifier fields rather than inventing a number.
\nAvailability. If the feed says in stock and the shelf is empty, you pay for a click that cannot convert and you teach the platform that your listings are unreliable. Match the feed update frequency to how fast inventory moves. A store that sells out of popular items daily should not refresh weekly.
\nPrice. The feed price must match the price on the landing page and at checkout, including after sale periods end. Mismatches commonly lead to disapproved items and, repeated, to account warnings. Fix the source system, not only the feed.
\nRun these checks with a script, not by eye. A simple pandas routine can flag blank GTINs, duplicate IDs, titles under a minimum length, and prices of zero. Run it before every upload. Ten minutes of checking beats a week of suppressed products.
\nSchema.org is an open vocabulary, a shared set of labels that tells machines what the content of a page means. Search engines read it to understand that a number on your page is a price, a phone number is a phone number, and a block of text is an opening hours statement.
\nFor shopping, mark up product pages with Product, Offer, and availability fields so the page tells the same story as the feed. When a crawler finds a price in the markup that differs from the feed, trust drops. The page, the markup, and the feed should be three views of one fact.
\nFor local businesses, LocalBusiness and its subtypes carry name, address, geographic coordinates, opening hours, and a link to the profile. Mark up each location page separately. A single chain page listing 40 addresses in one block is hard for a machine to read and hard for a person to use.
\nThe deeper value is operational. Writing structured data forces you to keep one source of truth for business facts. When the owner updates hours in one place and the website, markup, and listing all follow, you have solved the root problem instead of patching three symptoms. Validate markup after every template change, because a small theme edit can quietly break it.
\nPlatforms offer store visit estimates built from signals such as device location. These are useful directional numbers and should be read with care. They are modeled, not counted. They depend on enough nearby traffic and on enough users sharing location. They tend to be thin for small, rural, or single-location businesses, and they struggle with multi-story buildings and shared parking lots. A visit is also not a sale. Someone who walked into your store because of an ad may have bought nothing.
\nThere is a second trap: overlap. A shopper might see a shopping ad, a text ad, and a local listing within an hour. Each report claims the visit. Add the three reports together and you may count the same person three times.
\nThe cleaner test is a geographic holdout. Pick matched pairs of markets with similar size and sales history. Pause or reduce the format in half of them for a set period, such as four weeks, and compare total sales, calls, and walk-in traffic between the two groups, using point of sale data rather than platform reports. The gap between them is your lift. Run it during a stable period, not around a holiday, and decide before the test what difference would change your budget.
\nFor small businesses with only one location, a simpler habit helps: ask every new customer how they found you and log the answer at the counter. It is imperfect, but it is yours, and it will often contradict a dashboard in useful ways.
\nPrimary category: The single business type a listing is mainly classified under, which strongly influences the searches it can appear in. GTIN: Global trade item number, the barcode identifier a manufacturer assigns to a product, used to match your item to known products. Service-area business: A business that goes to customers rather than receiving them at a public address. Product feed: A structured file listing every product with its identifier, title, price, availability, and attributes. Structured data: Labeled markup on a web page, such as schema.org, that tells machines what each piece of content means.
\n1. Raising budget before auditing the data, which scales a feed's errors and wastes the extra spend. 2. Reusing one item ID across product variants, which breaks reporting and confuses matching. 3. Publishing a home address as a public storefront, or hiding the address of a business that depends on walk-in visits. 4. Asking only satisfied customers for reviews, or offering rewards for them, which invites policy trouble. 5. Adding store visit, call, and online conversion reports together without checking for overlap.
\n1. Name one owner for listing accuracy and review the listing against reality on a fixed schedule. 2. Write titles in the order a shopper would type them, with attributes before decoration. 3. Make price and availability in the feed, the page, the markup, and checkout agree, and fix the source system. 4. Respond to every review within a few days, calmly and with specifics. 5. Test lift with a matched-market holdout using point of sale data before shifting major budget.
\n1. Add a custom label for margin tier so reporting shows profit, not just revenue. 2. Keep a short list of phrases customers use on the phone and compare it with your categories every quarter. 3. Update holiday hours at least two weeks ahead, and again the day before. 4. Run a feed check script before every upload, not after a disapproval notice arrives. 5. Treat a sudden drop in impressions as a data question first and a bidding question second.
\nThree free tools cover most of this work. Python with the pandas library (BSD license) cleans and checks feeds: find blank identifiers, duplicate IDs, short titles, and price mismatches in a few lines. Schema.org (an open vocabulary, free to use) provides the labels for Product, Offer, and LocalBusiness markup, and its published type pages double as a checklist of fields worth filling. Scrapy (BSD license) crawls your own site to audit whether pages, markup, and prices agree with the feed. None of these replaces judgment, but together they turn a vague worry about data quality into a repeatable weekly check.
\nLocal and shopping search reward the businesses whose records are right. A listing and a feed are the real targeting: they tell the platform what you are, where you operate, what you sell, and whether it is available. Categories, service-area settings, titles, identifiers, availability, and price each decide which searches you can enter and what you pay to stay there. Reviews work as creative that customers trust more than your own words, and they must be requested with discipline. Structured markup keeps the page honest and the facts aligned. Store visit numbers are modeled estimates that overlap across reports, so a matched-market holdout is the sturdier test of lift.
\n1. In these formats, your data records decide which searches you can enter, so audit data before adjusting budget. 2. Categories and service-area settings must match how customers actually search and where you actually work. 3. Titles, unique identifiers, availability, and price accuracy account for most feed damage and are checkable by script. 4. Reviews are a trusted ad asset that depends on volume, recency, and honest responses. 5. Store visit reports are modeled and overlap, so confirm lift with a matched-market holdout.
\nThis week, assign one named owner to reconcile every listing and feed field against the real store, shelf, and checkout price, and freeze any budget increase until that owner reports back.
\n1. Schema.org. Product, Offer, and LocalBusiness type documentation. Open vocabulary maintained by a community of search engines and contributors. 2. GS1. General Specifications, global trade item number allocation rules. Standard governing product barcode identifiers. 3. Federal Trade Commission. Guides Concerning the Use of Endorsements and Testimonials in Advertising, 16 CFR Part 255. Guidance relevant to review solicitation. 4. Federal Trade Commission. Trade Regulation Rule on the Use of Consumer Reviews and Testimonials, 16 CFR Part 465. Rule addressing fake and incentivized reviews. 5. pandas development team. pandas documentation. Reference for data cleaning with BSD licensed software. 6. Scrapy project. Scrapy documentation. Reference for crawling and site audits with BSD licensed software.
"},{"id":"chapter-12-organic-search-as-a-hedge-against-paid-cost","title":"Chapter 12: Organic Search as a Hedge Against Paid Cost","html":"By the end of this chapter, the reader will be able to: 1. Explain why organic traffic is an owned asset with real costs, and why the label \"free\" fails a finance review. 2. Design and run a cannibalization test that separates paid clicks that add customers from paid clicks that only replace organic ones. 3. Map a keyword set into coverage tiers and assign each tier a paid response. 4. Perform a technical audit and rank the findings by their effect on crawling, indexing and speed. 5. Evaluate content ideas by whether they earn demand or merely describe the company. 6. Construct a maintenance plan that keeps organic search alive when the budget shrinks.
\nOpen the dashboard on a Monday and you will probably see two lines. Paid search shows conversions, a cost per conversion, and a tidy return. Organic search shows a bigger pile of visits and a cost column that reads zero. The room nods. Organic is free, paid is expensive, so why not do more of the free thing?
\nNow put that same page in front of a finance partner. She will ask three questions. Who wrote the articles, built the templates and fixed the redirects, and what were they paid? If paid ads were switched off tomorrow, how many of those conversions would have happened anyway through the organic listing sitting two inches below? And if organic really is free, why does the traffic drop whenever nobody tends it? Most dashboards cannot answer any of the three, and a number that cannot survive those questions is a story, not a measurement.
\nBoth lines on that dashboard are wrong in ways that cancel into false comfort. Organic is not free. It is paid for up front, in salaries and agency fees and months of waiting, instead of per click. And paid search is often credited with sales it only collected, especially on searches that include the brand name, where the customer was already walking to the door.
\nThis chapter treats organic search as what it is: a capital project with a slow payback and a long life, sitting next to a rented channel with a fast payback and no residual value. You will learn to test whether the two overlap, to audit the foundation, to build content that earns demand instead of chasing it, and to keep the asset alive when money is tight. I will also be honest about the part nobody can promise: how much click volume the answer boxes and AI summaries will take.
\nPaid search is rent. Every visit is bought at an auction price that responds to competitors, seasons and platform decisions you do not control. When the money stops, the visits stop on the same day. Organic search is ownership, with a catch. The page you publish keeps producing visits after the work is done, but only if it stays accurate, fast and findable. Ownership has a maintenance bill, and the assets can lose value if you ignore them.
\nThat is why I describe organic as a hedge instead of a replacement. A hedge reduces exposure to a specific risk, here the risk that your cost per click climbs or that budget gets cut mid-year. A brand with strong organic coverage can pull paid spend back in a bad quarter without going dark. A brand with none pays what the auction demands or disappears.
\nFor the finance conversation, count the costs honestly. Add content production, technical work, tools, link outreach and staff time into one monthly figure, then divide by the conversions organic search actually drives. That is your real cost per organic conversion. It will often be lower than the paid figure over a long horizon and higher during the first months. Both facts belong in the plan.
\nCannibalization means your paid ad takes a click that your organic listing would have won anyway. You pay for a visit that was already yours. The size of that waste is an empirical question, and guessing wrong in either direction is costly. Guess high and you switch off ads that were adding customers. Guess low and you keep paying for traffic you own.
\nThe cleanest method is a matched-market pause. Pick comparable regions, pause paid ads on a defined set of queries in some of them, keep them running in the others, and compare total conversions from search (paid plus organic combined) across the two groups. Total is the key word. If paid conversions fall by a lot but organic conversions rise to fill the gap, the ads were mostly substituting.
\nHere is an illustrative case. Suppose a regional furniture retailer, Juniper Lane, spends nine thousand dollars a month on searches containing its own name. Paid reports 300 conversions from them. The team pauses those ads in four of eight matched regions for four weeks. Paid conversions in the paused regions fall by nearly all of that share, but organic conversions in the same regions rise enough that total search conversions fall by only 15 percent of the paid figure. The retailer learns that roughly 85 percent of the branded paid credit was substitution. It cuts branded spend sharply, keeps a small defensive budget, and moves the savings to category terms where the ads proved incremental.
\nCannibalization is not one question. It is one question per group of searches, and the answer differs by group. Sort your queries into tiers before deciding anything.
\n| Query group | Typical organic position | Paid response | |---|---|---| | Brand name | First, unless a competitor or reseller crowds you | Small defensive budget, tested | | Product and category with purchase intent | Often contested | Keep paid, bid by margin | | Comparison and problem-aware questions | Strong if content exists | Pull paid back where organic holds top spots | | Long-tail specifics | Often unranked | Organic content gap, not a paid problem |
\nThe brand tier is where substitution usually shows up first, but not always. A brand with weak product pages, a hijacked listing from a reseller, or a famous competitor may find the paid ad earns its keep even there. Search your own terms in a clean browser and note who appears. It takes an hour and teaches more than a report.
\nFor purchase-intent terms, a top organic result does not automatically justify turning off ads. Two listings from one brand can crowd out competitors, and whether that extra share is worth the ad cost is what the test measures.
\nThe informational tier is the quiet opportunity. People asking how to choose, compare or fix something are early in their decision. Paid clicks on those terms are often expensive relative to their conversion rate. A helpful organic page, written for that person, reaches them at no per-click cost and plants the brand before they compare vendors. If you are weighing where to spend the next hour of effort, spend it here.
\nA technical audit asks one blunt question: can a search engine find, read and fairly judge your pages? Content quality is irrelevant if the answer is no. Most audits uncover a handful of problems that cost real traffic, and the fixes are usually cheap compared with new content.
\nStart with crawling. A crawler is software that follows links from page to page to discover content. Check that your robots file (a plain text file telling crawlers where they may and may not go) is not blocking important sections, and that your XML sitemap lists the pages you want found. Then check indexing, meaning whether a page is stored in the search engine's database. Look for stray noindex tags, conflicting canonical tags (the tag naming the preferred version of a page), and duplicate pages competing against each other.
\nSpeed and usability come next. Lighthouse scores a page's loading behavior from a browser or command line, and sitespeed.io repeats such tests and records trends, because one fast test proves little. Test on a mid-range phone, closer to what customers hold than a developer laptop. Scrapy walks your own site and exports every URL with its status code, title, headings and canonical tag, which is the raw material of an audit. Crawl only sites you have permission to crawl, at a polite request rate.
\nHere is the audit checklist in the order I work through it: 1. Confirm the robots file allows crawling of every section that should rank. 2. Confirm the XML sitemap is current and contains only live, indexable pages. 3. Find pages returning errors, and pages redirecting through chains of more than one hop. 4. Find duplicate titles, missing titles, and pages with no clear heading. 5. Check canonical and noindex tags against intent, page by page for the top 50 pages. 6. Measure speed and layout stability on mobile for the template types that carry the most traffic. 7. Review internal links so the pages that earn revenue are reachable within a few clicks of the home page. 8. Check structured data (labeled markup that describes products, reviews and articles to machines) for errors.
\nRank findings by blast radius. One bad robots rule can hide an entire section. A slow template affects thousands of pages. A missing title on one old blog post affects almost nothing. Fix in that order.
\nThere are two kinds of content. One describes what the company wants to say. The other answers what someone is already trying to find out. Only the second earns demand, meaning it brings in people who were not looking for you by name.
\nStart from the question, not the product. Gather the real questions customers ask: sales call notes, support tickets, reviews, the on-site search box, the phrases that appear in your paid search term reports. That last source is useful and underused. Your paid reports show which searches people actually type, so they are a free research file you already pay to create.
\nThen match the format to the need: a fair comparison for the shopper, steps for the person fixing a problem, a plain price for the buyer. A short page that answers fully beats a long one that wanders. Be conservative with any claim you cannot back up, especially in health, money, children or safety topics.
\nSubstance is what earns links, and links are one of the strongest signals that other people consider you worth citing. You rarely win them by asking. You win them by publishing something others want to reference: honest original data, a clear explanation of a confusing process, a template people can use. A local hardware retailer that publishes a plain guide to choosing fasteners for decks, with photographs from its own store, may be cited by builders and hobbyists. A page of keyword-stuffed text about \"quality fasteners\" will not.
\nHere I have to be careful, because anyone who gives you a confident number is guessing. Search results pages now often show direct answers, summary boxes and AI-written overviews above or beside the traditional listings. When a searcher gets what they need from the summary, they may never click anything. For some kinds of queries, especially simple factual questions, that clearly reduces visits to publishers. For others, especially those involving a purchase, a comparison or a local choice, people still click through, though what share and for how long is uncertain and changing.
\nWhat can a brand manager do with uncertainty? Plan for a range. Build the forecast three ways: organic traffic flat, moderately lower and substantially lower on informational queries. See which scenarios break the plan. If the plan only works in the flat case, it is too fragile, and that is a reason to keep a paid response ready for those terms instead of a reason to abandon organic.
\nShift effort toward the pages that summaries serve poorly. Original photographs, firsthand experience, local detail, real inventory, current pricing, calculators, and anything that requires a transaction or a relationship stay valuable because a summary cannot complete them. Pages that exist only to state a widely known fact are the most exposed.
\nSmall budgets force honest triage. You cannot do everything, so do the things whose absence would hurt most.
\nFirst, protect the foundation. A quarterly crawl with Scrapy and a Lighthouse pass on your main templates costs hours, not thousands. Fix the issues with the widest reach and ignore the cosmetic ones. A site that stays crawlable and fast is an asset that waits patiently for better days.
\nSecond, choose a narrow content target. Pick the 10 to 15 questions closest to a sale, where you have something real to say, and write them well. Ten strong pages beat 80 weak ones, because weak pages dilute your site and take upkeep.
\nThird, maintain before you create. Refreshing a page that already ranks tenth into the top five is often faster than producing a new page that must start from zero. Schedule one refresh per week and stick to it.
\nFourth, set an expectation with whoever approves the budget: organic results arrive in months, and the evidence is rankings, impressions, branded search growth and assisted conversions before it is revenue. If the budget will be cut, say in advance what gets paused and what must continue, because a half-maintained site decays quietly and is expensive to revive.
\nCrawl budget: The amount of attention a search engine is willing to spend fetching pages from a given site, which matters most on large sites with many near-duplicate pages. Canonical tag: A line of page code naming the preferred version of a page, so that duplicates do not compete with each other. Search intent: The underlying goal behind a query, such as learning, comparing or buying, which determines the kind of page that deserves to rank. Zero-click search: A search where the person gets what they need from the results page itself and never visits a site. Branded query: A search that includes the company, product or person name, usually from someone who already knows the brand.
\n1. Reporting organic search as free in a finance review, which invites the obvious follow-up about salaries, agencies and tools and damages the credibility of the whole channel. 2. Turning off all branded ads on a hunch, without a matched test, and learning too late that a competitor or reseller now sits in the top spot for your own name. 3. Publishing dozens of thin pages around product keywords instead of a few that answer real customer questions, which dilutes the site and creates upkeep nobody budgets for. 4. Auditing for cosmetic problems first, such as a missing description on an old post, while a blocked directory or a slow main template hides far more traffic. 5. Judging organic search on its first quarter, then cancelling the program just before older pages start to compound.
\n1. Run the cannibalization test tier by tier, report the result as a range, and repeat it whenever competitors or your site change meaningfully. 2. Keep one monthly cost figure for organic that includes labor, tools and outside fees, and compare it against organic conversions on a long horizon. 3. Source content ideas from sales calls, support questions, on-site search and paid search term reports before consulting any keyword list. 4. Audit in order of blast radius, and keep a dated log of what you fixed and what happened afterward. 5. Plan traffic in three scenarios for informational queries, and write down in advance what you will do in each.
\n1. Search your own brand terms in a clean browser every month. Reseller listings, outdated sitelinks and competitor ads show up there before they show up in a report. 2. Reuse your paid search term report as an organic research file. Phrases that already convert are the best candidates for a page you own. 3. Give each important page one owner and a review date. Pages without owners decay, and nobody notices until traffic falls. 4. When a site is rebuilt, map every old address to its new one before launch. Lost redirects are a common way for years of earned authority to vanish overnight. 5. Treat the first sentence of a page as the answer, then add the depth. It serves readers, and it gives automated summaries a clear passage to quote.
\nThree free, open source tools cover most of a small team's technical audit. Lighthouse (Apache 2.0) scores loading behavior, accessibility and basic search readiness for a single page and runs from a command line or browser. sitespeed.io (MIT) repeats those tests on a schedule and stores trends, so you can see whether a release made the site slower. Scrapy (BSD) is a Python crawling framework that exports every URL on a site you own with its status code, title, headings and canonical tag. For paid and organic reporting, a spreadsheet and careful notes work. No open source option here replaces a commercial link database, so treat link research as a manual, small-scale task.
\nOrganic search is an owned asset paid for up front in labor instead of per click, and it hedges against rising auction prices and budget cuts only if it is built early and maintained. Its value against paid search depends on how much your ads replace clicks you would have won anyway, which a matched-market pause test measures tier by tier. A technical audit, ranked by how much traffic each problem hides, protects the foundation. Content that answers real questions earns links and demand, while answer layers and AI summaries put an uncertain, changing share of informational clicks at risk and call for scenario planning. On a small budget, protect the foundation, write fewer and better pages, refresh before creating, and set honest expectations about timing.
\n1. Organic search is not free, so count labor, tools and fees in one cost figure before comparing it with paid. 2. A matched-market pause shows how much paid credit is substitution, and the answer differs for brand, category and informational terms. 3. Technical problems should be fixed by blast radius, with crawl blocks and slow templates ahead of cosmetic issues. 4. Content earns demand when it answers a real customer question better than the alternatives and offers something a summary cannot complete. 5. Exposure to answer layers is uncertain, so forecast a range of outcomes and keep a paid response ready for the most exposed queries.
\nBrand-term spend earns its place only when a test says so: pause paid ads on your brand terms in half of a matched set of regions for four weeks, then bring finance the combined paid and organic result and a real monthly cost for your organic program.
\n1. World Wide Web Consortium. Web Content Accessibility Guidelines, version 2.1. Specification for accessible web content. 2. Google Search Central. Search Essentials and technical documentation on crawling, indexing, canonical URLs and structured data. 3. Brodersen, K.H., Gallusser, F., Koehler, J., Remy, N., Scott, S.L. Inferring causal impact using Bayesian structural time-series models. Annals of Applied Statistics. Methodology for estimating the effect of an intervention without a randomized control. 4. Lighthouse project documentation, Apache 2.0 licensed, describing audit categories and scoring.
"},{"id":"chapter-13-search-measurement-and-waste-audits","title":"Chapter 13: Search Measurement and Waste Audits","html":"By the end of this chapter, the reader will be able to: 1. Separate a search account into brand and nonbrand reporting so each is judged by its own standard. 2. Perform a step-by-step waste audit across search terms, geographies, devices, hours and audience layers. 3. Interpret impression share as a diagnostic signal rather than a target. 4. Design a matched geo test that estimates how much brand search spend is truly incremental. 5. Reconcile platform-reported conversions to orders recorded in the business's own system. 6. Build a monthly audit routine with open source tools and a written decision log.
\nMost search accounts are wasting money, and the platform's own reports will never tell you so. That claim meets a serious objection: search is the closest thing advertising has to a cash register. A person types a need, you answer, and a sale often follows within minutes. If the reports show conversions at a healthy cost, who am I to say the money is wasted?
\nHere is who I am. I am someone who has watched a brand manager celebrate a low cost per conversion while the business's actual order count sat flat. The objection is right that search works. It is wrong about what the reports prove. A report counts every conversion that touched an ad, including the many that were going to happen anyway: the loyal customer who typed your company name, the person who already had your phone number in a browser tab, the shopper who clicked an ad because it sat above the free listing they planned to click.
\nThink of a stadium with a ticket scanner at every gate. The scanner counts everyone who walks through, including season ticket holders who would have shown up regardless of the promotion you ran. Counting heads at the gate tells you attendance. It does not tell you what the promotion added.
\nSo a waste audit has two jobs. First, find spend that buys nothing: queries nobody should have matched, places you cannot serve, hours when no one buys. Second, find spend that buys something that was already yours. The first job is mechanical and pays quickly. The second takes a test and a little courage, because the answer is sometimes uncomfortable. This chapter walks both, with figures that are illustrative only, so you can run the same audit on your own account without needing anyone's permission but your own.
\nBrand search means queries that contain your company, product or founder name. Nonbrand search means everything else: people describing a problem or a category without naming you. These two behave like different businesses wearing the same account login.
\nBrand queries convert at very high rates and low cost because the searcher already knows you. That makes blended reporting dangerous. A cheap brand campaign sitting next to an expensive nonbrand campaign produces an average that flatters the nonbrand work and hides the question of whether brand spend is needed at all.
\nStart by putting each into separate campaigns with separate budgets, and make sure the brand campaign cannot absorb nonbrand queries. Then add your own brand terms as exact match negatives in the nonbrand campaigns, so a person typing your name always lands in the campaign built for them. Competitor terms deserve a third bucket, since they cost more and convert worse, and you should judge them on their own.
\nReport the three buckets side by side every month. If someone asks for one blended cost per conversion, give it, then immediately show the split. The blended number is a fact. It is rarely a useful one.
\nThe search terms report lists the actual queries that triggered your ads, which is different from the keywords you bid on. Keywords are your guesses. Search terms are the evidence.
\nWork through the report in this order.
\n1. Pull at least 30 days, longer if volume is thin, and export it rather than reading it on screen. 2. Sort by cost, highest first. Spend concentrates, so the top 50 rows usually hold most of the money. 3. Mark every term with meaningful spend and zero conversions. Define \"meaningful\" in advance, for example any term that has spent more than two to three times your target cost per conversion. 4. Sort each marked term into one of three piles: irrelevant (add as a negative), relevant but unprofitable (change the ad, page or bid), or relevant but under-sampled (leave alone and watch). 5. Group the irrelevant terms by theme, not one by one. Job seekers, free-seekers, do-it-yourself researchers and unrelated products each deserve a negative list that catches the whole family.
\nBe careful with the third pile. A term with one click and no conversion tells you nothing. Cutting on thin data is its own form of waste, because you remove queries that might have been your best buyers. Patience is part of the method.
\nOnce queries are clean, look at where, on what, when and to whom the ads are shown. Each of these is a slice of the same spend, and each slice can hide a leak.
\nFor geographies, export results by city or postal area, not just by state. Look for places where cost is high and orders are near zero, especially areas at the edge of your service territory or far from any location where a customer could pick up or receive service. Also check the location setting itself: the default option can include people who merely show interest in a place, not people who are in it.
\nFor devices, compare conversion rate and cost across phones, computers and tablets. Do not conclude from one month that a device is bad. Phone searchers often research first and buy later on a computer, so a phone that looks weak in the platform may be feeding your orders. Adjust bids modestly and verify against your order data before cutting.
\nFor hours and days, build a simple grid of clicks, cost and orders by hour. A business that answers phone leads only during standard business hours should not pay full price at midnight. A business with an online checkout may find night traffic perfectly fine.
\nFor audience layers, look at any remarketing lists or in-market segments attached to the campaign in observation mode. If a layer shows a much better or worse conversion rate, you have a bid adjustment candidate. Keep audience choices conservative where children or sensitive categories are involved, and never build layers from anyone's health, financial hardship or similar private conditions.
\nEvery slice needs the same discipline: a stated threshold, a stated action, and a written reason.
\nImpression share is the portion of eligible auctions in which your ad actually appeared. The platform also reports how much you lost because the budget ran out and how much you lost because your ad rank was too low.
\nNew buyers often chase a high number, as though 100 would mean victory. It would mean you are paying to appear in every auction, including the worst ones. A fast food chain does not try to put a restaurant on every corner. It builds where traffic justifies the rent.
\nRead the lost shares as diagnoses. Budget loss on a campaign that is profitable at the margin says you are leaving good sales unbought. Budget loss on a campaign that is unprofitable says the limit is protecting you, so leave it alone. Rank loss points to ad quality, landing page quality or bids, which you can examine one at a time.
\nOne more caution. Impression share on brand terms is usually high, and a drop there often means a competitor is bidding on your name. That is worth knowing, but it does not by itself prove you must outbid them. The next section helps you decide.
\nHere is the question that makes brand managers uncomfortable: if you turned off brand ads, how many orders would you actually lose? Some would vanish, because competitors or resellers would take the top spot. Many would simply move to the free listing right below. A matched geo test gives you a measured answer.
\nChoose pairs of regions that resemble each other in population, season, baseline order volume and marketing exposure. Within each pair, one region is the test and the other is the control. Pause brand ads in the test regions only. Keep everything else unchanged: no new promotions, no price changes, no extra media in those places.
\nRun it long enough to cover your normal buying cycle, typically four to six weeks, and begin with a stretch of pre-test data so you know how closely the paired regions moved together before anything changed. Then compare total orders from your own records, not platform conversions, in the test regions against what the pre-test relationship predicted.
\nCausalImpact, a free statistical package, does this comparison by building a forecast of what the test regions would have done without the change, then measuring the gap. Treat the result as an estimate with a range, not a verdict.
\nBe honest about limits. Few regions, noisy data or a competitor who attacks during the test will weaken the answer. If your business is small, a simpler version works: pause brand ads in half the pairs for a few weeks and compare raw order counts. Counsel and contract terms may matter if you sell through partners who co-fund brand bidding, so check agreements before pausing anything.
\nPlatform conversions and actual orders almost never match, and the gap is the audit's compass. Reasons include conversions counted more than once, conversions credited to ads for sales that closed offline, returns and cancellations, delayed attribution windows, and browsers or consent choices that stop tracking.
\nBuild a monthly reconciliation table with five columns: platform conversions, orders in your system, the ratio between them, the matched order value, and unexplained gap. Use order identifiers, not guesses, wherever you can. If your checkout passes a click identifier to your order system, you can match individual orders to individual clicks. If it does not, compare totals by week.
\nDo not force the numbers to agree. The aim is a stable ratio you understand. If platform conversions run higher than orders by about a third and stay there, you can apply that factor to forecasts. If the ratio swings from month to month, you have a tracking fault that must be fixed before any bidding decision can be trusted. Track the ratio separately for brand and nonbrand, because it is nearly always different.
\nSuppose a regional home services company spends 36,000 dollars in a month. Brand search takes 9,000 and nonbrand takes 27,000. The platform reports 720 conversions in total: 450 on brand at 20 dollars each and 270 on nonbrand at 100 dollars each. The company's own order system shows 540 orders. The ratio is three quarters, so the platform overstates by a third.
\nThe audit then runs the slices.
\n| Slice | Finding | Monthly spend flagged | |---|---|---| | Search terms | Job seeker, free repair and do-it-yourself families, zero orders | 4,200 dollars | | Geographies | Three outer areas outside the service radius, nine conversions in total | 2,100 dollars | | Hours | Midnight to five in the morning, phone leads go unanswered | 1,200 dollars | | Devices | Tablets cost twice the average per order | 1,500 dollars |
\nThe flagged total is 9,000 dollars, one quarter of spend. Because the slices were measured separately, a modest overlap likely exists, so the realistic savings are somewhat lower. The team applies negatives, tightens the location setting, schedules hours and trims tablet bids. It does not redirect the savings blindly. It holds them back until the next month's data shows orders did not fall.
\nNext comes the brand test. Paired regions run for five weeks with brand ads paused in the test half. Orders in the test regions fall by about 15 percent. The team reads the figure as roughly 15 of every 100 brand-attributed orders being truly incremental, though the range is wide. Applying that to 450 platform conversions yields about 68 incremental conversions, which puts true brand cost near 130 dollars each, well above the 20 dollars the platform showed.
\nThe team keeps a reduced brand budget as a defensive position and moves the rest toward nonbrand terms that survived the audit. The lesson is not that brand is worthless. The lesson is that its reported cost was an illusion.
\nA one-time audit decays. Queries drift, competitors change and campaigns pick up new clutter. A routine keeps the gains.
\nWeek one, export search terms, geography, device, hour and audience data into one folder. Week two, run the pass and write down every action with its reason and its expected effect. Week three, apply changes and note the date, so later swings can be traced. Week four, update the reconciliation table and compare orders against the prior month. Once a quarter, schedule one geo test, and review the account structure for duplicate keywords and mismatched landing pages.
\nAsk someone who does not manage the campaigns to review the log. People defend their own settings. A fresh reader asks why a term is still running.
\nSearch terms report: The platform's list of actual queries that triggered your ads, as opposed to the keywords you chose to bid on. Negative keyword: A word or phrase you tell the platform never to match, used to block irrelevant queries before they cost money. Incrementality: The portion of conversions that would not have happened without the ad, usually estimated with a holdout or geo test. Geo holdout: A test in which ads are paused in selected regions while matched regions continue, so total results can be compared. Reconciliation ratio: Platform-reported conversions divided by orders in your own records, tracked over time to expose tracking faults.
\n1. Judging brand and nonbrand on one blended cost per conversion, which lets cheap brand sales hide expensive nonbrand work. 2. Adding negatives for terms with one or two clicks, which removes queries that might have become buyers once they had enough data. 3. Treating impression share as a target and raising budgets until the number climbs, regardless of what those extra auctions return. 4. Running a geo test while launching promotions or changing prices in the test regions, which makes the result impossible to read. 5. Forcing platform conversions to match orders by adjusting settings until they agree, instead of explaining why they differ.
\n1. Keep brand, nonbrand and competitor campaigns separate, each with its own budget and its own reported results. 2. Write a threshold, an action and a reason for every audit finding, and keep the log where the next reviewer can read it. 3. Measure results against orders in your own system, and keep the reconciliation ratio as a standing monthly report. 4. Run one geo test per quarter on the campaign with the largest spend or the least certain contribution. 5. Hold audit savings in reserve for one month and confirm orders held steady before spending them elsewhere.
\n1. Sort the search terms report by cost, not by conversions, because waste hides at the top of the spend column. 2. Group irrelevant queries into themes and block the whole family, since new variants appear every week. 3. Check that the location setting targets people in an area, not people merely interested in it, before blaming the geography itself. 4. Pause brand ads only inside a designed test, never because someone in a meeting suspects they are unneeded. 5. Report the reconciliation ratio separately for brand and nonbrand, because the two nearly always differ.
\nDuckDB (MIT license) is an analytical database that runs on one laptop and queries exported search term files directly, so the audit pass needs no server. Metabase (AGPL license) turns those tables into a shared dashboard for the monthly review, the reconciliation ratio and the audit log. CausalImpact (Apache 2.0 license) estimates the effect of pausing brand ads in a geo test by comparing test regions with a forecast built from control regions. All three are free and can be self-hosted. No open source tool replaces the platform's own bidding, so the audit still starts with exports from it.
\nA search account can look healthy while spending money on queries that never convert, places that cannot be served, hours that never answer, and brand clicks that would have arrived free. The cure begins with separating brand from nonbrand, then running a disciplined pass through search terms, geographies, devices, hours and audience layers. Impression share guides diagnosis but is never a goal. A matched geo test measures how much brand spend is incremental, and a reconciliation table keeps platform conversions honest against real orders. A monthly routine with a written log, reviewed by someone outside the campaign, turns a single clean-up into a lasting habit.
\n1. Brand and nonbrand search behave like different businesses and must be reported separately. 2. The search terms report is evidence, and every audit action should carry a threshold, a reason and a date. 3. Impression share and lost share explain constraints, but they should never set the budget. 4. A matched geo test, read as a range, is the practical way to estimate brand incrementality. 5. The ratio of platform conversions to real orders must be stable and understood before bids are trusted.
\nThis month, split brand from nonbrand, export your search terms, and compare platform conversions to actual orders in one table before you change a single bid. Then schedule one matched geo test on brand search for next quarter.
\n1. Brodersen, K.H., Gallusser, F., Koehler, J., Remy, N., Scott, S.L. Inferring causal impact using Bayesian structural time-series models. Annals of Applied Statistics. Source methodology for the CausalImpact package. 2. CausalImpact package documentation. Apache 2.0 licensed software for estimating the effect of an intervention on a time series. 3. DuckDB documentation. MIT licensed analytical database, reference for querying exported files directly. 4. Metabase documentation. AGPL licensed business intelligence tool, reference for dashboards and saved questions. 5. Google Ads Help. Search terms report and impression share definitions, platform documentation for the metrics discussed.
"}]},{"title":"Part III: Paid Social and Creator Distribution","items":[{"id":"chapter-14-paid-social-auction-dynamics","title":"Chapter 14: Paid Social Auction Dynamics","html":"By the end of this chapter, the reader will be able to: 1. Explain why a social auction prices interruption while a search auction prices intent. 2. Calculate how bid, estimated action rate and ad quality combine into a total value that decides who wins. 3. Diagnose whether a performance dip comes from the learning phase, creative fatigue or outside competition. 4. Select a conversion event that the platform can learn from, given the volume a campaign actually produces. 5. Design a consolidated account structure that gives the delivery system enough data to work with. 6. Evaluate platform reported results against first-party event records stored in your own database.
\nSuppose two advertisers want the same thumb-stopping slot in the same feed. Brand A bids 50 dollars for each purchase it wants. Brand B bids only 30 dollars. Most buyers would say Brand A wins, because it is willing to pay more. Now add one more fact. The platform estimates that an impression of Brand A's ad leads to a purchase four times in 10,000. For Brand B the estimate is nine in 10,000. Multiply bid by estimated action rate and Brand A is worth 20 dollars per thousand impressions while Brand B is worth 27. The cheaper bidder wins. Brand A, with the bigger budget and the better intentions, loses the slot and keeps losing it.
\nThat small piece of arithmetic is the whole chapter in miniature. Many brand managers still behave as though the social feed is a marketplace where the highest offer gets the space, the way a billboard owner sells to the top bidder. It is not. The platform is running a revenue and experience calculation, and it is just as interested in whether the person in the feed will like what appears as in what you will pay.
\nRaising a bid is the lever everyone reaches for first, and it is usually the weakest one available. Editing a campaign feels like taking control, and it often quietly resets the system that was learning on your behalf. Splitting a modest budget into a dozen tidy audience segments feels like discipline, and it often starves every segment of data.
\nPaid social can be a strong part of a media plan. It can also be an efficient way to fund a platform's experiments with your money. This chapter explains the mechanics well enough that you can tell which one you are running. For platform-specific menus and settings, use each platform's current help pages.
\nA search auction starts with a person typing words. The words are a statement of need, and the advertiser is bidding to answer it. A social auction starts with a person opening an app to be entertained, to check on friends, or to kill four minutes in a waiting room. Nobody asked for your product. Your ad arrives uninvited.
\nThat one difference changes what the auction has to do. In search, the system can match your ad to a typed query and read the intent almost directly. In social, the system has to infer interest from behavior: what the person watched, paused on, saved, ignored and tapped. It runs this inference thousands of times a second, and every time it must decide which of many competing ads deserves the next slot.
\nThe practical consequence is that social demand is largely created, not captured. Someone who buys after a social ad often had no plan to buy that morning. That makes social excellent for reaching people before they search, and it makes social results harder to read, because the sale may land days later through some other door. Treat the first click or the first view as the start of a relationship, not a verdict, and judge the channel with the measurement habits covered in Chapter 44.
\nYour cost per result will also be less stable than a branded search campaign. Intent is a steady stream. Attention is weather.
\nMost social platforms rank competing ads using some form of total value. Think of it as three ingredients stirred together. The first is your bid, which is what an outcome is worth to you. The second is the estimated action rate, the platform's prediction that this particular person will take the action you are optimizing for. The third is ad quality, a measure of how the experience feels to the viewer, including signs that people hide the ad, report it, or find the promise misleading.
\nThe exact formula belongs to each platform and changes without notice, so distrust any specific weighting you read online. The shape stays stable. Total value rises with the bid, the predicted action rate and the viewer experience. An ad can win with a low bid if the other two are strong, and lose with a huge bid if people scroll past it or complain.
\nNow a second illustrative case. A regional fitness studio, which I will call Tamsin Strength, bids 24 dollars per trial signup with a stiff stock photograph, and the platform estimates three signups per 10,000 impressions. That is about seven dollars per thousand in auction terms. A rival bids only 18 dollars, but its ad is a short clip of a real member finishing a hard set and laughing. The estimate is five per 10,000, worth nine dollars per thousand. The lower bidder wins. Tamsin could raise its bid, but the better fix is changing the ad so the predicted rate climbs. That fix lasts. A bid increase only rents the slot until the next rival adjusts.
\nThe design is not hostile to advertisers. The platform wants the feed pleasant enough that people keep opening it, so the auction quietly rewards advertisers who respect the viewer.
\nWhen you launch a campaign, the delivery system does not know who will respond. It spends the early budget exploring: showing your ad to varied people, watching what happens, and adjusting its predictions. This exploration period is usually called the learning phase. Results during it are noisy and often expensive, which is not a defect. It is the system buying information.
\nPlatforms generally describe a threshold of optimization events needed in a recent window before delivery stabilizes. The number differs by platform, so check your account documentation. The system needs enough completed actions, close together, to separate real patterns from luck.
\nThis is why edits matter so much. Conceptually, a significant change tells the system that its earlier observations may no longer apply. Changing the optimization event, swapping the creative, rewriting the audience, or making a large budget jump can send delivery back toward exploration. Small cosmetic changes may not. Which edits count as significant is defined by each platform, so I will not pretend to give you a universal list. Use a simple rule instead: if the change could plausibly alter who sees the ad or who responds to it, assume it restarts some of the learning.
\nPicture the common failure. A campaign runs three days above target cost. The buyer edits the audience, exploration restarts, costs rise, and the buyer edits again on day six. The campaign never leaves the noisy period, and the buyer concludes that social does not work. The platform did not fail. The buyer kept interrupting the experiment.
\nA workable discipline: decide the evaluation window before launch and write it down, batch edits into planned changes instead of tweaking every morning, and record every edit with its date. A truly broken ad, a wrong landing page or a policy problem needs fixing at once. Everything else can wait for the window to close.
\nThe conversion event is the action you tell the platform to find more of. Pick it badly and the system will efficiently find the wrong people.
\nOptions run from shallow to deep: a click, a landing page view, an add to cart, a lead form, a purchase. Shallow events are plentiful and cheap, so the system learns fast. But a person who clicks easily is not necessarily a person who buys. Optimize for clicks and the system will find habitual clickers, who may be lovely people with no intention of purchasing. Deep events are the real goal, but they are rare, and a rare event starves the learning system.
\nThe decision rule I use is to optimize for the deepest event that your campaign can produce in enough volume to learn from. Suppose a campaign selling high priced online courses produces only three purchases a week. Optimizing for purchase will leave the system guessing for months. Moving up to a qualified lead, such as a completed application, might produce 40 events a week, which is enough for the system to find patterns. Then you track how many applications turn into enrollments on your own side, so you know the shallow event is a fair stand-in.
\nCheck that the event is real. A tag that fires on a button click instead of a completed form teaches the system to find button clickers, and a purchase event that fires twice per order distorts predictions. Audit with a test transaction before spending and after any site change.
\nBe careful with sensitive categories. Where the audience includes minors, or the product touches health, finance or other protected situations, take the conservative default: avoid optimizing around inferred sensitive traits, follow the platform's category rules, and have counsel review anything that touches regulated claims. This is operational guidance only.
\nBuyers once controlled reach with long lists of interests and demographics. Those controls have weakened on many platforms, and delivery systems increasingly infer who should see an ad from the ad itself. The picture, the opening seconds, the words and the people shown all tell the system who might respond. Your creative is a targeting instruction.
\nA video that opens on a mother in a kitchen and one that opens on a runner at dawn can promote the same product with identical audience settings and often reach noticeably different people. The system reads who engages and sends more to similar viewers.
\nThat changes how to build tests. Instead of testing five audiences with one ad, test five creative angles inside one broad audience. An angle is a reason to care: saving time, saving money, social proof, a surprising fact, a problem named plainly. Each attracts a different slice of people. The principle that creative is your largest controllable variable gets its full treatment in Chapter 49. For this chapter, remember only that the auction math rewards it directly, because stronger creative lifts the estimated action rate.
\nLead with the hook in the first moments, since scrolling is fast. Show the product in use. Make only claims you can substantiate, because misleading promises damage ad quality. Retire an ad when results fall away and frequency climbs, since fatigue shows up as rising cost with no change in your settings.
\nWhere children, creators or families appear on screen, default to clear permissions, age appropriate content and honest endorsements.
\nCompetition is the first. When a large retailer launches a sale and floods the same feed, your costs rise because the slots are suddenly in demand. Nothing you did caused it.
\nSeasonality is the second. Demand for attention swells around major shopping periods and often softens in quieter weeks, so identical creative and targeting can look expensive in a busy month. Compare like weeks with like weeks.
\nAudience size is the third. A small audience gets saturated quickly. Frequency climbs, the same people see the same ad, engagement drops, and the estimated action rate falls with it. A wider audience gives the system more fresh people to try. Tight targeting feels precise, but it can be the most expensive choice you make.
\nCreative fatigue and your own edits are the other two causes, and the only two you control.
\nWhen costs move, check the calendar for a sale or season, then frequency and creative age, then your edit log, then audience size against budget. Only after those checks should you conclude the campaign has failed. Pausing a healthy campaign during a temporary spike often throws away the learning you paid for.
\nLearning needs data, and data needs volume. This leads to the structural principle that surprises tidy-minded managers: fewer campaigns and fewer ad sets, each with more budget, usually learn faster than many small ones.
\nSuppose you have three thousand dollars a week and a target cost per lead of 30 dollars. At target you can expect about 100 leads. Split that money across 10 audience segments and each segment gets roughly 10 leads a week, far too few for the system to learn anything. Consolidate into two groups and each gets about 50 leads, which is a far better foundation. The segments you gave up were mostly illusions of control anyway, because the creative was already steering delivery toward the right people.
\nA reasonable starting structure has one campaign per business objective, a few broad audience groups, and several creative angles inside each. Keep prospecting, meaning people who have not met you, separate from retargeting, meaning people who already visited. Their costs differ so much that blending them hides both.
\nSet daily spend so the system can plausibly gather its events within your evaluation window. If the arithmetic says your budget cannot support the event you chose, move to a shallower event, run fewer campaigns, or accept that the test is underfunded. When you scale, raise budgets in steps. Doubling overnight can trigger fresh exploration and a cost spike that looks like proof the channel stopped working.
\nEvery platform reports its own conversions, and every platform has an interest in showing its own contribution favorably. View-through credit, modeled conversions and overlapping attribution windows can all inflate the dashboard relative to what your business recorded.
\nThe remedy is a first-party record you control. Log visits and conversions on your own site with campaign tags in the link addresses, and store the raw events in a database you own. Then compare weekly: platform reported conversions against your recorded conversions for the same campaigns. The two will never match exactly, but the ratio should be stable. A sudden change means something shifted in tracking, attribution settings or traffic quality. Investigate before you move budget. Chapter 41 treats data ownership in depth.
\nTotal value: The combined score an ad platform uses to rank competing ads, built from the bid, the estimated action rate and an ad quality measure. Estimated action rate: The platform's prediction of how likely a specific viewer is to complete the optimization event after seeing a specific ad. Learning phase: The early exploration period when a delivery system gathers data and results are volatile, usually restarted by significant edits. Optimization event: The single action, such as a lead or a purchase, that you instruct the platform to find more of. Frequency: The average number of times a person sees your ad in a given period, which climbs quickly in small audiences and signals fatigue.
\n1. Raising bids to fix poor delivery when the real problem is a weak estimated action rate that better creative would repair. 2. Editing audiences, budgets and ads every day, which keeps the campaign permanently in a noisy exploration state. 3. Optimizing for a cheap shallow event such as a click, then wondering why the clickers never buy. 4. Splitting a modest budget across many small ad sets so that none of them collects enough events to learn from. 5. Blaming the campaign for a cost spike that actually came from a competitor sale, a busy season or a saturated audience.
\n1. Write the evaluation window and the success threshold before launch, and judge the campaign only against them. 2. Choose the deepest conversion event you can produce in enough volume, and verify it fires correctly with a test transaction. 3. Test distinct creative angles inside a broad audience rather than many audiences with a single ad. 4. Keep prospecting and retargeting in separate campaigns so their different economics stay visible. 5. Keep an edit log with dates, so you can separate your own changes from market movement when results shift.
\n1. Compare like periods against like periods, since a busy shopping week and a quiet week are not honest neighbors. 2. Watch frequency alongside cost: a rising frequency with a rising cost usually means the audience is tired of the ad. 3. Scale budgets in moderate steps and let each step settle before taking the next one. 4. Pass real order value to the platform only when your tracking is accurate, because the system will trust bad numbers. 5. Ask whether an ad would still look honest to a skeptical parent, a creator's lawyer and a regulator, and revise it if the answer is no.
\nTo check platform reports against your own records, use Matomo (GPL license) as a self-hosted web analytics tool that records visits and conversions with campaign tags on your own server. Store the exported events in PostgreSQL (PostgreSQL license), where a simple query can line up your recorded conversions beside the platform's numbers for each campaign and week. Together they give you a first-party comparison without a software subscription. Run the comparison on a fixed schedule and keep the table.
\nThe social auction prices interruption, so it ranks ads by combining your bid with the platform's prediction that a viewer will act and a measure of how the ad feels to watch. That structure means better creative often beats a higher bid, and it means the creative itself works as a targeting instruction. Delivery systems need volume to learn, so significant edits can restart exploration, conversion events must be deep enough to matter and frequent enough to learn from, and accounts work best when consolidated. Costs also move for reasons outside your control, including competition, season and audience size, so a short diagnostic sequence beats a panicked edit. Finally, platform dashboards deserve a first-party check, which a self-hosted analytics tool and a database make possible.
\n1. A lower bid can win a social auction when the estimated action rate and ad quality are stronger, so total value, not bid alone, decides placement. 2. Significant edits can send a campaign back toward exploration, so batch changes and judge results inside a planned window. 3. Optimize for the deepest conversion event that produces enough volume to learn from, and test that the event fires correctly. 4. Creative tells the delivery system who to find, so test distinct angles inside broad audiences. 5. Consolidate budgets into fewer ad sets and compare platform reported results with your own recorded events every week.
\nBefore you touch a bid this week, divide your weekly budget by your target cost per result and count how many ad sets that volume can feed, then merge until each one clears the number of events you set as your learning threshold.
\n1. Varian, H. \"Position Auctions.\" International Journal of Industrial Organization. 2. Vickrey, W. \"Counterspeculation, Auctions, and Competitive Sealed Tenders.\" Journal of Finance. 3. Matomo documentation, self-hosted analytics and campaign tracking guides, matomo.org. 4. PostgreSQL Global Development Group. PostgreSQL Documentation, postgresql.org. 5. Platform advertiser help documentation on auction, delivery and learning behavior (consult each platform's current pages).
"},{"id":"chapter-15-audience-architecture-without-third-party-cookies","title":"Chapter 15: Audience Architecture Without Third-Party Cookies","html":"By the end of this chapter, the reader will be able to: 1. Identify which parts of audience targeting changed when third-party cookies lost reach and which parts did not. 2. Classify any audience in a media plan into one of four sources: first-party, platform-modeled, contextual, or partner. 3. Evaluate a seed list for a modeled audience and predict whether it will teach the platform anything useful. 4. Explain why hashed identifiers are still personal data and design a matching process around that fact. 5. Construct an audience map that assigns each source a job, an owner, a consent basis, and a test. 6. Decide when broad targeting paired with strong creative is a rational choice rather than a lazy one.
\nThe most common assumption I hear from brand managers is that the loss of third-party cookies broke targeting. Somebody on the team says it in a planning meeting, heads nod, and the budget conversation moves on as if a wall had come down in the middle of the stadium.
\nIt did not. What broke was one particular method of following a browser from site to site and attaching a profile to it. Targeting itself, meaning the work of deciding who should see an ad and why, was never a cookie. It was a set of decisions about people, and those decisions can be made with other raw material.
\nThe assumption fails in two directions. Some teams freeze spending until a vendor sells them a replacement identifier. Others keep buying the same rented segments and discover that reach quietly shrank under them. Both are reacting to a headline instead of reading their own account.
\nHere is the harder truth. Many audiences that brand managers believed they owned were never theirs. They were leased from a data broker through a pipe that the browser maker, the regulator, or the consumer could close at will. If your plan listed 40 segments and you cannot say where a single one came from, you did not have an audience architecture. You had a shopping list.
\nThis chapter is about building the real thing. I will walk through what changed and what held steady, the four sources you can draw on, how to give a platform a seed it can learn from, why hashing does not make a person anonymous, how consent has to be wired into the plumbing, and how to draw one map that ties it all together. I will also make the case, with some reluctance, that broad targeting is frequently the right answer.
\nStart with the mechanics. A third-party cookie is a small file placed by a domain other than the one the visitor is looking at. Ad tech used it to recognize the same browser on many unrelated sites. That recognition made cross-site retargeting and rented segments easy to buy and easy to scale.
\nWhen browsers limited that behavior and consumers gained more control over tracking, the cross-site recognition got patchy. Reach on certain browsers thinned out, and segments built on browsing history became less complete.
\nNow the list of what did not change. People still have interests, still live in places, still buy things, still respond to a message that fits their situation. Publishers still know what their pages are about. Platforms still see what logged-in users do inside their own walls. Your customers still hand you their email addresses when you give them a reason to. Every one of those facts is available today.
\nSo the working question for a brand manager shifts. It stops being \"what replaces the cookie?\" and becomes \"which sources of audience knowledge do I control, which do I rent, and which do I borrow from a partner?\" That is an architecture question, and architecture questions have answers you can write on one page.
\nEvery audience you buy or build comes from one of four places. Learning to name the source is the single most useful habit in this chapter.
\nFirst-party lists are people who have dealt with you directly: customers, subscribers, app users, event registrants, and anyone who filled in a form. You collected the data, you know the context, and you hold the permission, or you should. These lists are small compared with total reach, but they are accurate and they are yours to use.
\nPlatform-modeled audiences are groups a platform builds by prediction. You hand over a seed, such as a customer list or a pixel-fed set of converters, and the platform finds people who look similar in the signals it can see. You do not see the model or its inputs. You see the output and the results.
\nContextual audiences are reached by what the person is looking at right now, not by who they are. An ad beside an article about trail running reaches whoever is reading about trail running. No profile is required. Context is the oldest targeting method in advertising and it has become more useful, not less.
\nPartner data comes from another organization that has a real relationship with its own customers: a retailer, a publisher, a league, a venue, a co-marketing brand. Access is negotiated and usually bounded by a contract, a clean room, or a matching service.
\nA plan that uses all four, with a clear job for each, can lose any one and keep operating.
\nModeled audiences are only as good as what you teach them. The seed is the lesson plan. Hand a platform a messy seed and it will happily find you more of the mess.
\nSuppose a fictional outdoor-gear retailer, Larkspur Outfitters, uploads 12,000 email addresses of everyone who ever opened a message. The list mixes loyal repeat buyers, one-time coupon hunters, people who bought a gift, and subscribers who never purchased. The platform has no way to know which of those groups matters. It finds lookalikes of the average, and the average is mostly coupon hunters. Costs look fine, revenue disappoints, and nobody can say why.
\nNow suppose Larkspur builds the seed from the quarter of the list that purchased at full price at least twice. Same platform, same budget, different lesson.
\nA short test for seed quality:
\nIf a seed fails two or more of those, fix the seed before you touch the targeting settings. It is the cheapest improvement available.
\nThis is the misconception I most want to correct, because it causes real legal and reputational trouble. Many teams believe that once an email address is run through a hashing function, it becomes anonymous and can be shared freely. That is wrong.
\nHashing turns an input into a fixed string of characters. The same email always produces the same string. That sameness is exactly why hashing is useful for matching: two parties hash the same address and get the same result, so they can find overlap without swapping the raw address.
\nBut the same sameness is why hashing is not anonymization. The hashed value is a stable identifier for one person. Anyone who holds a list of likely email addresses can hash that list and compare. Email addresses are a small, guessable space compared with random secrets, so reversing the match is often practical. And when the hash is used to link a person's activity across companies, it is functioning as the very identifier privacy rules care about.
\nRegulators and privacy counsel generally treat hashed identifiers as personal data when they can single out an individual. Counsel decides how that applies to your situation, and this is operational guidance only. What I can tell you is the working rule my teams follow: treat a hashed email exactly as you would treat the email.
\nThat rule has practical consequences:
\nPrivacy-safe matching exists, and it is worth using. Matching through a clean room, where neither side sees the other's raw rows, reduces exposure. It does not remove obligations.
\nConsent is not a banner. A banner is the visible part. Consent plumbing is the set of pipes that carries the visitor's choice to every system that touches their data, and a leak anywhere makes the banner decoration.
\nPicture the path. A visitor lands on your page and sees a notice. They accept analytics, decline advertising. That choice has to travel to your tag manager, your analytics tool, your server-side event collector, your customer database, and your email platform. Then, later, when you build an audience upload, the system has to remember who said no.
\nMost failures happen at the handoffs. A tag fires before the banner loads, or the server-side collector never receives the choice, or an old spreadsheet export still feeds a custom audience.
\nBuild the plumbing in this order: 1. Record the choice: store the decision, the date, the notice version the person saw, and the purposes they accepted, in a table you own. 2. Gate collection: no tag, pixel or event fires until the relevant purpose is accepted. 3. Carry the flag: attach the consent status to every customer record so downstream systems can filter on it. 4. Filter uploads: audiences exported to platforms include only people whose record permits that use. 5. Honor withdrawal: when someone changes their mind, the change reaches the database, the platforms, and the partners within a defined time. 6. Audit quarterly: pick 20 records at random and trace each one end to end.
\nRespect browser-level privacy signals such as Global Privacy Control where applicable. Counsel decides what the law requires; the plumbing must carry that decision.
\nAn audience map is a single page, and its purpose is to stop the argument about which audience is \"best.\" Each row is one audience. Each column answers a question you should be able to answer in under a minute.
\nThe columns I use:
\nWalk Larkspur through it. Row one is past full-price buyers, first-party, used for repeat-purchase messages and as the seed for row two. Row two is a platform lookalike built from row one, used for prospecting. Row three is contextual placements beside trail and camping content, used for awareness with no profile at all. Row four is a co-marketing audience from a regional outfitter, partner data, matched inside a clean room under a written agreement.
\nFour rows, four jobs, four owners. When results come in, the team looks at which row delivered, and each row has a test attached so the answer is not just a platform's own report.
\nReview the map on a calendar, not when something breaks, because audiences go stale quietly. It also keeps budget talks honest: if most of the spend sits on one modeled row with a thin seed, everyone can see the concentration.
\nHere is the conclusion that surprises people after a chapter of audience building: sometimes the smartest move is to stop narrowing.
\nPlatform delivery systems have grown better at finding likely responders on their own, using signals the advertiser never sees. When you stack tight interest filters and layered exclusions on top, you may be restricting that system to a smaller pool, raising your cost per impression and starving it of the data it needs to learn. A broad audience with a clear conversion goal gives the model room to work.
\nThat does not mean targeting is dead. It means the burden shifts. When the audience is wide, the creative does the filtering. An ad that speaks directly to a trail runner, in words and images a trail runner recognizes, sorts the crowd by itself. Viewers who do not fit scroll past, and the ones who fit lean in. The message becomes the targeting.
\nBroad targeting is a rational choice under a few conditions:
\nIt is a poor choice when the product serves a small niche, when the budget is too thin to learn, when conversion signals are noisy, or when sensitive categories and young audiences are involved. In those cases I keep targeting conservative and narrow by default, and I do not let a platform widen it automatically.
\nThe test is simple. Run a broad ad set against your best narrow ad set, same creative, same budget, same period, and compare cost per outcome you trust. Decide from your own result, not from a rule of thumb.
\nFirst-party data: Information a business collects directly from its own customers and visitors through its own properties and relationships.
\nSeed audience: The source list or set of converters handed to a platform so its model can find similar people.
\nHashing: A one-way transformation that turns an identifier into a fixed string, useful for matching but not a guarantee of anonymity.
\nClean room: A controlled environment where two parties compare data and receive aggregate results without either side seeing the other's raw records.
\nContextual targeting: Placing an ad based on the content of the page or program the person is viewing, not on a profile of the person.
\n1. Treating a hashed email list as anonymous and sharing it loosely with agencies, vendors and partners without a lawful basis or a record of the transfer. 2. Uploading an entire email list as a seed, including coupon hunters and non-buyers, then wondering why the lookalike audience attracts low-value customers. 3. Installing a consent banner and assuming the job is finished, while tags and server-side collectors keep firing regardless of the visitor's choice. 4. Keeping dozens of rented segments in the plan because nobody remembers why they were added, and nobody can name their source. 5. Switching to broad targeting with weak, generic creative, which leaves the ad with nothing to sort the audience and the platform with nothing to learn.
\n1. Label every audience in the plan with one of four sources, and pause any audience that cannot be traced to its origin. 2. Build seeds from your best customers, ranked by value and recency, and refresh them on a schedule. 3. Treat hashed identifiers as personal data in every contract, workflow and conversation with partners. 4. Store consent decisions in a database you own and filter every audience export against that record. 5. Test any narrow audience against a broad one with matched creative and budget before committing to it for the quarter.
\n1. Ask every vendor selling a segment to state the original collection method in writing. A vague answer is itself an answer. 2. Keep one contextual line in nearly every plan. It requires no identity, survives consent changes, and gives you a baseline to compare other audiences against. 3. When a platform shows a low match rate on your upload, check the formatting first: stray spaces, mixed capitalization and typos in email fields are common causes. 4. Give partners a short data-sharing checklist before the first call: purpose, fields, retention period, deletion process and who signs off. 5. Review a handful of real ads in the feed as a viewer would see them. If you cannot tell who the ad is for within two seconds, the creative is not ready for broad delivery.
\nThree free, open source tools cover the foundation of this chapter. Klaro (BSD license) is a consent manager that shows notices, records the visitor's choices, and can block scripts until a purpose is accepted. Matomo (GPL license) is a self-hosted analytics platform that lets you collect first-party site data on your own server and respect consent settings. PostgreSQL (PostgreSQL license) is the database where you can keep consent records, customer attributes and audience definitions under your own control. Together they give you the record, the collection and the storage layers. Where a full-scale demand-side platform is concerned, no open source option reaches parity, so say so plainly in planning.
\nThe loss of third-party cookies ended one way of following browsers across sites, but it did not end the work of deciding who should see an ad. Audiences come from four sources: first-party lists, platform-modeled groups, contextual placement, and partner data. Modeled audiences depend on the quality of the seed, so seeds should be small, pure, recent and ranked by value. Hashed identifiers remain personal data and should be handled as if they were the original email. Consent has to travel through every system that touches a record, not stop at a banner. A one-page audience map assigns each audience a source, an owner, a consent basis, a refresh rule and a test. When the product and budget fit, broad targeting with specific creative is often the rational response, and a direct test will tell you whether it is right for your account.
\n1. The change removed a tracking method, not the need for audience decisions, so rebuild around sources you can name. 2. Seed quality sets the ceiling for any modeled audience, and a smaller pure seed usually beats a larger mixed one. 3. A hashed email is still personal data, and contracts and workflows should treat it that way. 4. Consent is a system of connected pipes, and each handoff between tools is where the failures hide. 5. Broad targeting shifts the sorting job to the creative, so it works only when the message is specific and the conversion signal is clean.
\nThis week, build the one-page audience map: list every audience in your current plan, write down its source, owner and consent record, and pause any row nobody can trace.
\n1. Global Privacy Control specification, published through the W3C Privacy Community Group. Technical standard for communicating a user's privacy preference from browser to site. 2. Federal Trade Commission. Dot Com Disclosures: Information about Online Advertising. Guidance on clear and conspicuous disclosure in digital advertising. 3. Interactive Advertising Bureau Tech Lab. OpenRTB specification. Reference for how audience and identity signals travel in programmatic auctions. 4. Klaro consent manager project documentation. Reference for configuring consent notices and script blocking. 5. Matomo project documentation. Reference for self-hosted analytics and consent configuration. 6. PostgreSQL Global Development Group. PostgreSQL documentation. Reference for relational database design and access control.
"},{"id":"chapter-16-creative-systems-for-feed-story-and-short-video","title":"Chapter 16: Creative Systems for Feed, Story, and Short Video","html":"By the end of this chapter, the reader will be able to: 1. Describe the four modules (hook, proof, offer, call) that make up a short paid social ad. 2. Apply format rules that keep one idea working in feed, story, and short video placements. 3. Design a weekly production workflow that a team of two or three people can sustain. 4. Construct a naming convention and creative library that lets anyone find what ran and why. 5. Write a creative brief that a designer or editor can execute without a follow-up meeting. 6. Evaluate a finished ad for sound-off viewing, caption quality, and contrast before it ships.
\nA buyer I know lost a quarter to a single video. It was a good video. The client loved it, the first week of results looked great, and so the whole budget went behind it. By week five the numbers had slid, and by week seven the account was paying more for every result than it had before the launch. Nobody had a replacement. Nobody had even a rough draft of one. The team spent the next month scrambling, and the scramble cost more than the original shoot.
\nThat story taught me something the platforms rarely say out loud. A winning ad is not an asset. It is a loan. It buys you attention now, and the interest comes due when the audience has seen it enough times to stop noticing it. The brands that stay steady do not have better luck finding winners. They have a machine that keeps producing candidates.
\nA creative system is closer to a kitchen line than a gallery. Ingredients are prepped in advance, stations have owners, plates go out in a known order, and a bad plate gets caught before it reaches the table. This chapter shows how to build that line with a small team and free software: modular ads, a workflow, a naming scheme, a brief template, accessibility checks, and a weekly rhythm. Testing methods and budget rules live in other chapters. Here the job is to make sure there is always something good ready to test.
\nThink of every short ad as four blocks that can be swapped independently. The hook is the first beat, the thing that stops a thumb mid-scroll. The proof is the reason to believe: a demonstration, a customer line, a before and after, a number you can legitimately support. The offer is what the viewer gets and on what terms. The call is the single action you want next.
\nWhy modules? Because a finished ad is one point in a large space of combinations. Suppose you produce four hooks, two proofs, two offers, and two calls. That is 64 possible ads from only ten pieces of work. Most of those combinations you will never run. The point is that when one hook tires out, you replace one block, not the whole ad. Production cost drops, and your tests stay clean because only one block changes at a time.
\nA useful hook test: could a stranger tell, from the first frame with no sound, what this ad is about and who it is for? If not, rewrite it. Hooks tend to come in a handful of families. There is the problem statement (\"Your gutters are overflowing again\"), the surprising result, the direct question, the demonstration already in motion, and the plain statement of who it is for. Rotate families, not just wording, because a viewer tires of a pattern faster than a phrase.
\nProof deserves the most care. Do not invent testimonials or imply results you cannot document. If a creator is paid or given product, disclosure belongs in the ad itself, in plain words. This is operational guidance only; your counsel decides what your claims require.
\nThe same four modules travel across placements, but each placement has its own physical rules.
\nFeed is the most forgiving format. The viewer is scrolling at their own pace, so a static image or a square or four-by-five video can carry more text and a slower build. Feed also shows your copy above or below the creative, which means the headline and the visual can split the work. Do not repeat the same sentence in both places.
\nStory is full-screen and vertical, and the platform lays its own interface over your ad. Profile names, reply bars, and buttons cover the top and the bottom of the frame. The practical rule is a safe zone: keep faces, key words, and logos inside the middle of the frame and leave generous margins top and bottom. Check every story export against the platform's own preview, because the covered areas differ by app and change without notice. Story viewers also tap through fast, so the hook has to land in the first second or two.
\nShort video is the most demanding format. It is vertical, it usually autoplays, and it competes with entertainment made by people who do this all day. Ads that look like ads get skipped. Ads that look like something the viewer might have chosen to watch tend to hold attention longer. That does not mean copying trends. It means a human voice, a real setting, quick cuts that serve the message, and no slow logo animation at the front.
\nOne workable approach is to design the vertical short video first, then derive the others. Crop a square for feed, pull a still frame for the static image, and trim the story version. Going the other direction, from a horizontal master down to vertical, nearly always produces tiny text and awkward framing.
\nA small table keeps the team aligned:
\n| Placement | Frame | Hook window | Text position | |---|---|---|---| | Feed | Square or four by five | First two to three seconds | Anywhere, keep it short | | Story | Full vertical | First one to two seconds | Center safe zone only | | Short video | Full vertical | First one to two seconds | Center, above captions |
\nMost small teams do not have a creative department. They have a marketer, a freelancer, maybe an owner who shoots on a phone. The workflow has to fit that reality.
\nStart with batching. Shooting one ad at a time wastes setup effort. Instead, spend one half day filming raw material for a whole month: six to eight hooks delivered to camera, product demonstrations from two angles, customer or staff lines, and a pile of clean b-roll, which is supporting footage with no speech. Record each hook three times with slightly different energy. Shoot vertical. Capture room tone for ten seconds, because you will need it to patch edits.
\nThen run an assembly line with four stations. The first station logs footage: rename clips, mark the best takes, and note timestamps. The second assembles rough cuts, hook plus proof plus offer plus call, with no polish. The third adds captions, text overlays, and any graphic elements. The fourth is review, where someone other than the editor checks the ad against a written checklist before it goes anywhere near the ad account.
\nThat fourth station is the one small teams skip, and it is the one that catches a missing disclosure, a misspelled offer, or a caption that dropped a word. Even a two-person shop can do it by swapping roles: the editor of one ad reviews the other's.
\nSix months from now, someone will ask which ad produced the best results in the spring and what was in it. If your files are called \"final,\" \"final2,\" and \"final_REAL,\" the answer is lost. A naming convention is boring, and it is also the cheapest thing in this chapter to fix.
\nBuild the file name from fields separated by underscores, in a fixed order. A workable pattern: brand, campaign, concept, hook family, hook version, proof, offer, call, format, length, version number. A real example might read: HarborCo_SpringTune_PriceFirst_Question_H2_Demo_Free-Shipping_Shop_9x16_20s_v3. Use no spaces, no special characters, and keep every field from a short controlled list so two people never spell the same offer two ways.
\nThe same string should appear as the ad name inside the ad platform. That single habit makes exports from the platform line up with your library without anyone retyping anything. When results land in a spreadsheet, you can split the name into columns and instantly see performance by hook family, by offer, or by format.
\nThe library itself can be nothing fancier than a shared folder tree and a spreadsheet. Folders hold raw footage, approved modules, finished exports, and retired ads. The spreadsheet has one row per ad with its name fields, launch date, retirement date, a link to the file, and a notes column for what you learned. Retired ads stay in the library. An ad that fatigued last spring may work again with a new audience, and the notes tell you why it was retired.
\nBad briefs create more rework than bad editors. A brief is not a mood board. It is a set of decisions made once, in writing, so nobody has to guess.
\nA good one fits on a page and answers eight questions. Who is the viewer, in one sentence? What single thing should they do after watching? What is the one message? Which block is being replaced or tested, and which blocks stay fixed? What is the placement and format? What must appear, such as an offer line, a disclosure, or a legal phrase? What must not appear? When is it due, and who approves it?
\nWhen a brief asks for a variation, say exactly what changes. \"Same ad, new hook\" is a real instruction. \"Make it pop\" is not. Name the hook family, supply the line, and state the length limit.
\nIf you work with outside creators, tell them up front what claims they can and cannot make, and that paid relationships must be disclosed. Send the checklist along with the brief rather than after the first draft comes back wrong.
\nMany viewers watch with the volume off, whether on a commute, in a waiting room, or beside someone asleep. Others cannot hear at all. Designing for sound off is therefore both a performance habit and an accessibility duty, and the two point in the same direction.
\nStart with captions. Burn-in captions, text rendered into the video itself, always display. Separate caption files can be switched on or off but may not be supported everywhere, so for paid social most teams burn them in. Keep lines short, about four to six words, break them at natural phrase boundaries, and keep them on screen long enough to read comfortably. Check the transcript for errors, especially product names and numbers, because automatic transcription gets proper nouns wrong with regularity. Position captions inside the safe zone and above the platform's interface.
\nThen contrast. Text on a busy video frame is hard to read for anyone, and harder for viewers with low vision. The Web Content Accessibility Guidelines set a minimum contrast ratio of four and a half to one for normal text, a sensible floor for overlays too. If the footage behind the words is mixed, add a solid or semi-opaque band behind the text instead of hoping the background cooperates. Avoid thin type, avoid tiny type, and avoid relying on difference of hue alone to carry meaning.
\nFinally, make sure the ad still works with no audio at all. Run the mute test: watch the whole ad silently on a phone, at arm's length. Can you tell what is offered, to whom, and what to do next? If the answer depends on the voiceover, the ad is half finished. Sound should add warmth and emphasis, not carry the only copy of your message.
\nA system needs a heartbeat. For most small teams that means a fixed weekly rhythm with a short list of tasks on fixed days.
\nMonday is review. Pull last week's results by hook family, offer, and format, read them against the notes column, and decide what to retire, what to keep, and what to test next. Keep this to 45 minutes. Do not open a debate about taste. Let the numbers and the log lead.
\nWednesday and Thursday are production. Pull from the stockpile of footage, assemble, caption, and review. If you shot a monthly batch, most of this is editing, not filming.
\nFriday is launch prep. Run the three standing checks, apply the naming scheme, upload, and schedule new ads to begin at a sensible time. Then update the library spreadsheet, so the record is current before the weekend.
\nOne more thing on cadence: set a refresh trigger rather than a calendar date. An ad is a candidate for retirement when its frequency (the average number of times each person saw it) climbs and its results soften together. Smaller audiences hit that point sooner, because the same people see the ad more often. A team with a small local audience may need new material every week or two, while a team with a wide audience may stretch to a month. Your own data decides.
\nConsider a composite, fictional example. Birchwater Outfitters is a regional outdoor gear retailer with a three-person marketing team: a manager, a part-time editor, and a freelance photographer on call. They run paid social with a modest monthly budget and had been producing about one new ad a month, each treated as a standalone project.
\nThey changed their approach in three moves. First, they moved to one filming day a month. On that day they recorded eight hooks, four proof clips, and plenty of product b-roll, all vertical. Second, they adopted the naming scheme and a shared spreadsheet, which took one afternoon to set up. Third, they locked the weekly cadence above, with Monday review and Friday launch prep.
\nAfter two months, the team was shipping about six new variations per week, nearly all of them reuses of existing blocks with one change. Their review station caught a wrong price in one caption before launch, which would have run for days otherwise. The Monday logs showed that question hooks outran statement hooks on short video but not in feed, which changed what they shot the next month. That is illustrative, and your results will differ, but the pattern holds. A modest team with a clear process can outproduce a larger team without one.
\nHook: The opening beat of an ad, usually the first one to three seconds, designed to stop the scroll and signal relevance. Safe zone: The part of a vertical frame that stays clear of platform buttons, names, and captions, where key content should sit. B-roll: Supporting footage, such as product shots or settings, laid over narration or used to build variations without new filming. Burn-in captions: Text rendered permanently into the video image so it displays on every device regardless of settings. Creative fatigue: The decline in results that occurs when an audience has seen an ad often enough to stop responding to it.
\n1. Treating each ad as a one-off project, which makes every variation expensive and leaves no record of what worked. 2. Building the horizontal master first and cropping to vertical later, which shrinks text and pushes faces into the platform's covered areas. 3. Writing briefs full of adjectives like \"bold\" and \"fresh\" instead of naming the block to change, the format, and a reference ad. 4. Skipping a second-person review, which lets misspelled offers, missing disclosures, and dropped caption words reach the ad account. 5. Retiring ads by feel or by calendar instead of by a refresh trigger tied to frequency and softening results, so good ads die early and weak ones linger.
\n1. Break every ad into hook, proof, offer, and call, and change one block at a time so each result teaches you something. 2. Shoot in monthly batches, vertical first, with several takes of each hook and a stock of clean b-roll. 3. Use one naming convention for files, ad names, and the library spreadsheet so exports match without retyping. 4. Run a mute check, a safe-zone check, and a contrast check on every ad before launch. 5. Keep a bench of two or three reviewed ads ready so a fatigued ad never leaves a gap.
\n1. Record each hook three times with different energy; the flat take and the lively take rarely perform alike. 2. Put the product or the problem in the first frame, because a viewer deciding in a second cannot wait for a build. 3. Proofread captions against the script twice, since automatic transcription tends to stumble on brand names and numbers. 4. Rotate hook families, not just wording, because audiences tire of a pattern faster than a phrase. 5. Write one sentence of feedback to each creator after every round, so the next brief starts from what you learned.
\nEvery task in this chapter can be done with free, open source software. For video editing, Kdenlive (GPL), Shotcut (GPL), and Blender (GPL) cover cutting, titling, and motion work. Krita (GPL) and GIMP (GPL) handle image editing and retouching, Inkscape (GPL) handles vector artwork, and Penpot (MPL 2.0) handles layout and design collaboration. Whisper (MIT) produces transcripts for captions, which you should always proofread. FFmpeg (LGPL and GPL) converts, trims, and resizes exports in batches from the command line. Where one tool lacks a feature, another in this list usually covers it.
\nA creative system turns ad production from a series of one-off projects into a repeatable weekly routine. Splitting each ad into hook, proof, offer, and call lets a small team produce many variations from a modest amount of footage. Format rules for feed, story, and short video keep one idea working across placements, and a sound-off, caption, and contrast checklist protects both performance and accessibility. A naming convention, a library, and one-page briefs keep the work findable and usable, while a fixed weekly cadence and a bench of ready ads prevent the gap that fatigue creates.
\n1. A winning ad is borrowed attention, so build the machine that keeps producing candidates. 2. Modular ads let you replace one block instead of rebuilding the whole ad. 3. Design vertical short video first, then derive feed and story versions from it. 4. A consistent naming convention and a simple library turn past ads into usable knowledge. 5. Sound-off design, burned-in captions, and adequate contrast are part of the ad, not an afterthought.
\nThis week, rename your last 10 ads with the field-based naming scheme, then book one half day to film eight hooks so next month's variations are already in the can.
\n1. World Wide Web Consortium. Web Content Accessibility Guidelines 2.1. Success criteria for contrast and non-text content. 2. Federal Trade Commission. Guides Concerning the Use of Endorsements and Testimonials in Advertising, 16 CFR Part 255. 3. Kdenlive, Shotcut, Blender, Krita, Inkscape, GIMP, and Penpot project documentation. Official user manuals for each tool. 4. OpenAI. Whisper repository and model card. Open source speech recognition under the MIT license. 5. FFmpeg Project. FFmpeg documentation. Reference for command line conversion and filtering.
"},{"id":"chapter-17-creator-and-influencer-paid-partnerships","title":"Chapter 17: Creator and Influencer Paid Partnerships","html":"By the end of this chapter, the reader will be able to: 1. Distinguish the main paid partnership formats and explain how whitelisting changes who controls the media. 2. Draft the core terms of a creator agreement, including deliverables, usage term, exclusivity and approvals. 3. Apply the FTC Endorsement Guides to assign disclosure duties to the brand and to the creator. 4. Evaluate a creator's audience for signs of purchased or automated engagement before money moves. 5. Design a measurement plan that combines unique codes, tagged links and a holdout. 6. Judge whether a creator fits a family-first brand standard and set a conservative default when it does not.
\nHere is a hard question. If a creator posts a video about your product and 40 people buy, who gets the credit in your reports? Most brand managers answer \"the creator,\" and most of them are wrong about at least half of those 40.
\nSome already knew your brand. Some would have found you through search that week anyway. Some grabbed the creator's code because it was the first discount they saw. A few were never real people at all.
\nHere is the second uncomfortable assumption. Many managers believe the creator carries the legal and reputational risk of a sponsored post because the creator is the one talking. Federal guidance does not read that way. The brand is on the hook too, and \"the creator forgot\" is not a defense that makes anyone feel better when a regulator or a reporter calls.
\nSo this chapter treats creator partnerships as what they are: a contract, a media buy, a compliance exercise and a measurement problem, wrapped around a human being who has an audience and a reputation of their own. The trust a creator has built is real, and it can be borrowed, but it cannot be forced. What you can control is the paperwork, the vetting, the tracking and the media plan. Those are the parts that separate a program that compounds from a string of one-off posts. This chapter stays on the paid media, contract and measurement mechanics rather than the relationship and outreach side of working with creators.
\nStart with the format, because the format decides who controls distribution. There are three that matter in practice.
\nThe first is the organic sponsored post. The creator publishes on their own account, to their own followers, and you pay a fee. You get whatever reach the creator's audience delivers that day. You do not control timing, frequency or targeting, and you cannot add budget to make a good post travel further.
\nThe second is a partnership label or branded content tag, which most major platforms now offer. The creator posts, tags the brand, and the brand can often see basic performance data. The platform may also let the brand promote that post. This is a step toward paid media, but the post still lives on the creator's account.
\nThe third is whitelisting, sometimes called partnership ads or creator licensing. Here the creator grants your ad account permission to run paid campaigns using their handle and their content. The ad shows up looking like it came from the creator, but you control the budget, the audience, the schedule and the stop button. Whitelisting is where creator content joins the rest of your paid social plan. Think of it like a player agreeing to appear in your team's broadcast spots: the person is the draw, but your network decides when and where the spot airs.
\nWhitelisting lets one strong piece of creator content be tested against several audiences and retired the minute its cost per result drifts. It also changes the risk. You are now responsible for the ad, and every claim in that content is yours to defend.
\nPlatform mechanics change often, so check current help documentation before promising a creator how access works. The principle stays fixed: ask for the narrowest access that does the job, put an end date on it, and log who granted what.
\nMost creator problems I have seen trace back to a one-page agreement or a chain of direct messages. A real agreement covers six items. Counsel should review it, since this chapter is operational guidance and not legal advice.
\nDeliverables. Spell out the count, format, length and platform. \"Two videos\" is vague. \"One 45 to 60 second vertical video and three story frames, posted on named dates\" is a deliverable. Include the number of revision rounds and what happens if a post is late.
\nUsage term and scope. Usage rights answer one question: what may the brand do with this content beyond the creator's own post? Organic reposting on brand channels, paid whitelisting, website use and email use are separate permissions. Each should carry a start date, an end date and a territory. A usage term of 30 to 90 days is common to start with, because it keeps the fee reasonable and keeps the option to renew.
\nExclusivity. Brands love to ask for it and underprice it. Exclusivity means the creator cannot promote competitors for a defined period. It costs the creator real income, so a request for broad exclusivity should come with a visible premium. Define the category narrowly and the window briefly. \"No competing running shoe brands for 30 days\" is workable. \"No athletic products\" is not.
\nApprovals. Decide in advance what you review and when. A fair structure is that the brand reviews for factual accuracy, legal claims and required disclosures, while the creator keeps control of tone, delivery and style. A brand that rewrites the script until it sounds like a brand ad has paid for a creator and received an ad, usually a worse one.
\nDisclosure obligations. State in the agreement that the creator will disclose the relationship clearly, in the manner you specify, and that the brand may require edits or removal if it does not. Put the practical method in writing so nobody guesses.
\nPayment and kill terms. Tie payment to delivery and to compliance. Add a clause covering what happens if the creator's conduct becomes a brand safety problem during the term. Creators deserve a fair clause on their side too, covering timely payment and a limit on how long the brand can hold the content in review.
\nThe FTC Endorsement Guides apply a plain idea: if a viewer would not expect a connection between the speaker and the brand, and that connection could affect how much weight the viewer gives the message, the connection should be disclosed clearly. Payment is the obvious connection. Free product, affiliate commissions and family or employment ties can count too.
\nThe word that matters is clear. A disclosure that most viewers would miss is a weak disclosure. Several habits tend to create trouble:
\nFor video, the safer pattern is a spoken disclosure early in the content plus a visible on-screen text overlay that stays up long enough to read. For stories and short clips, repeat it on every frame that carries a pitch. For captions, put it at the start, in plain words such as \"Paid partnership with\" followed by the brand name.
\nNow the part many brand managers miss. Responsibility is shared. The creator has a duty to disclose, and the brand has a duty to give clear instructions, monitor what gets posted and act when something falls short. A brand that hands over a product and a payment and then never looks at the post is relying on luck.
\nBuild a simple routine. Give every creator a one-page disclosure instruction sheet. Check each post within a day of going live. Keep a dated record, such as a saved screenshot and link. If a post misses the standard, request a fix in writing and log the response. Claims matter as well. If a creator says a product cured, prevented or guaranteed something you cannot support, the brand needs substantiation for it. Whitelisted ads compound this issue, because you are paying to spread the claim. For anything touching health, finance or children, route the copy through counsel first. Counsel decides what the rules require in your case.
\nFollower counts are cheap to inflate. Purchased followers, engagement pods and automated comment accounts are old problems, and they have not gone away. The result is a creator who looks big and sells nothing.
\nYou do not need a paid vetting subscription to catch most of it. A manual review of 15 to 20 minutes per creator catches a surprising amount. Look at these signals together, since any one alone proves little:
\nThen ask for evidence that is hard to fake. Request screenshots of native analytics showing audience location, age range and reach for the last several posts. Ask for the typical share of views from non-followers. Compare what the creator tells you to what the platform shows you. A creator who is proud of a real audience will usually share this without fuss. A refusal is information.
\nThe cleanest vetting step is a small paid test. Before a large commitment, buy one deliverable with a tagged link and a code, then see whether real traffic and real purchases show up. Suppose a creator claims a 200,000 follower audience and your tagged link brings 90 clicks, most of which bounce in under two seconds. That tells you more than any audit report.
\nBrand safety for creators is about more than avoiding scandal. It is about whether the creator's whole body of work matches the standard your brand wants to hold in front of families.
\nMy default is conservative. Review at least the last three months of a creator's content, not just the best clips. Look at language, humor, the people who appear, and the products they promote beside yours. Check comment sections too, since a creator is partly responsible for the community they allow.
\nIf your product is aimed at families or could reach minors, set stricter rules. Avoid creators whose main audience is young children unless counsel has reviewed the rules for advertising to children. Do not ask a minor creator to carry a pitch without guardian involvement and legal review, and do not use content that shows a child in a way a parent would find uncomfortable. For sensitive categories such as alcohol, supplements, financial products or anything with age limits, apply age targeting on the paid side and decline creator audiences you cannot verify.
\nA fit test I use is simple. Would you be comfortable if the creator's next three posts appeared in the same feed as your ad, and a parent was reading over a shoulder? If the answer is a hesitant yes, the answer is no. That standard costs you some reach. It protects something harder to rebuild.
\nWrite the standard into the agreement as a conduct clause and review it at each renewal.
\nCreator measurement works best as layers, because every single method leaks.
\nUnique links. Give every creator and every placement its own tagged URL with campaign, source and medium parameters. Matomo, an open source analytics platform (GPL), supports campaign tracking from tagged links, so you can see visits, goals and conversions by creator on your own server. Link clicks undercount, because many viewers see a recommendation and later type the brand name into a search box.
\nUnique codes. A discount code catches the viewers who never click. Issue one code per creator, log every redemption in a table you control, and store it in Postgres or DuckDB, both open source databases that run on a modest server. Codes overcount, because coupon sites and shared screenshots spread them to people the creator never reached. Treat a code as a floor on attention and a rough guide to ranking, not a final count.
\nHoldouts. This is the layer that answers the real question: what would have happened without the partnership? The cleanest approach for a whitelisted campaign is to split the target audience at random, show the creator ads to one group and withhold them from the other, then compare purchases between groups. When a random audience split is not possible, use geography: run the partnership in some regions and hold similar regions back, then compare the change over time. Keep the groups large enough to see a difference and run the test long enough to cover your normal buying cycle.
\nAn illustrative worked example. Suppose a fictional outdoor cookware brand, Quarry Fox Camp Kitchen, pays three creators and spends 6,000 dollars in total, including fees and whitelisting spend. Codes report 150 redemptions and tagged links report 110 purchases, with 40 of those overlapping. That is 220 unique attributed buyers, or about 27 dollars each. Now the holdout. The exposed regions produced 1,300 orders over four weeks. The held-back regions, scaled to the same size, would have produced about 1150. The lift is roughly 150 orders. At 6,000 dollars, that is about 40 dollars per incremental order, which is meaningfully higher than the 27 dollar attributed number.
\nBoth numbers are useful. The 27 dollar figure ranks creators. The 40 dollar figure shows what the program truly bought, and budget decisions should lean on it. If the two diverge wildly, check for coupon leakage or a holdout that was not truly comparable.
\nWhitelisting: A permission setup in which a creator lets a brand run paid ads through the creator's handle, with the brand controlling budget and targeting. Also called partnership ads or creator licensing.
\nUsage rights: The brand's licensed permission to reuse creator content, defined by channel, duration and territory, usually priced separately from the creative fee.
\nMaterial connection: Any relationship between an endorser and a brand that could affect how viewers weigh the endorsement, such as payment, free product, commissions or personal ties.
\nHoldout: A group of people or regions deliberately kept from seeing a campaign so results can be compared against the exposed group.
\nEngagement pod: A group of accounts that agree to like and comment on one another's posts to inflate engagement figures.
\n1. Treating a code redemption count as proof of impact, when codes spread to people the creator never reached and miss viewers who bought without one. 2. Signing a message-thread agreement with no usage term, which leaves the brand unable to say whether a paid ad is licensed or in breach. 3. Asking for broad exclusivity without paying for it, then losing the creator to a better offer or a resentful relationship. 4. Assuming the creator alone is responsible for disclosure and never checking a live post. 5. Selecting creators by follower count and skipping analytics screenshots, which leads to paying for audiences that exist only on the screen.
\n1. Separate the creative fee from usage rights and price each permission with its own term, channel list and end date. 2. Give every creator a written disclosure instruction sheet and verify each post within a day of publication, saving a dated record. 3. Run a small paid test with a tagged link and a unique code before committing to a long or expensive deal. 4. Pair attributed results with a holdout at least once per program, and base budget decisions on the incremental figure. 5. Review a creator's recent content and comment section against your family-first standard before every renewal, not only before the first deal.
\n1. Ask for native analytics screenshots showing audience location and age range, since honest creators usually share them readily. 2. Whitelist only the content that already performed well organically, and ask for the narrowest account access that works. 3. Calendar every usage expiry date and set the ad schedule to end a day early. 4. Give creators the key facts and required claims, then let them choose the words, because a stiff script reads as an ad and viewers scroll past it. 5. Record the code and link naming rules in a shared sheet before launch so reports from different creators can be compared cleanly.
\nMatomo campaign tracking (GPL) records visits and conversions from tagged creator links on your own server, so each creator's traffic appears as its own source. Postgres or DuckDB, both free and open source, can store the code redemption table, the creator roster and the holdout results. Postgres suits a shared team database. DuckDB suits a single analyst working from files on a laptop. Together they cover link tracking, code logging and basic holdout comparison without a paid subscription. For large programs, a full creator management platform has no open source equal, and a spreadsheet plus a database is the honest alternative.
\nCreator partnerships are best run as a contract, a media buy, a compliance routine and a measurement test around a person with a real audience. Paid formats range from organic posts to whitelisted ads, and whitelisting moves control of budget and targeting to the brand while adding responsibility for every claim. A sound agreement defines deliverables, usage term, exclusivity, approvals, disclosure duties and payment. The FTC Endorsement Guides place disclosure duties on both creator and brand, so the brand needs instructions, monitoring and records. Vetting for fake engagement relies on manual review, native analytics and a small paid test. A family-first standard sets a conservative default for audience and conduct. Measurement layers tagged links, unique codes and holdouts, with the holdout giving the incremental figure that should guide budget.
\n1. Whitelisting turns creator content into media you control, which also makes you answerable for the ad. 2. A usage term, channel list and end date belong in every agreement, priced apart from the creative fee. 3. Disclosure is a shared duty, and the brand's part is clear instructions, prompt checks and dated records. 4. Fake engagement is caught by comparing claimed audience data against native analytics and a small paid test. 5. Links undercount, codes overcount, and only a holdout shows what the partnership added.
\nWrite the usage end date, the disclosure wording and the holdout plan into every creator agreement, then check the first live post within a day.
\n1. Federal Trade Commission. Guides Concerning the Use of Endorsements and Testimonials in Advertising, 16 CFR Part 255. 2. Federal Trade Commission. Disclosures 101 for Social Media Influencers. Plain-language guidance for creators and brands. 3. Federal Trade Commission. Dot Com Disclosures: How to Make Effective Disclosures in Digital Advertising. 4. Matomo documentation. Tracking campaigns with campaign URL parameters. Open source analytics, GPL licensed. 5. Brodersen, K.H., Gallusser, F., Koehler, J., Remy, N., Scott, S.L. Inferring causal impact using Bayesian structural time-series models. Annals of Applied Statistics. Published methodology for estimating campaign lift.
"},{"id":"chapter-18-community-seeding-and-organic-to-paid-loops","title":"Chapter 18: Community Seeding and Organic to Paid Loops","html":"By the end of this chapter, the reader will be able to: 1. Evaluate whether a brand has earned the standing to seed a community, and name the cases where it has not. 2. Design a weekly organic posting routine that works as a low-cost creative test bed for paid campaigns. 3. Select organic winners for paid promotion using a written decision rule rather than instinct. 4. Apply a permission and rights process to user-generated content before any of it appears in an ad. 5. Construct moderation and response standards that keep a community from becoming an ad inbox. 6. Assess the loop with a simple measurement plan that separates community lift from paid spend.
\nNobody gets a Friday report from a community thread. Community seeding is slow. It is hard to attribute. It depends on people who owe the brand nothing. A paid campaign can be switched on this afternoon, it spends the budget on schedule, and a report appears by Friday. Why would a serious brand manager spend scarce hours posting into a group of strangers when the same hours could tune a bid strategy?
\nHere is the honest answer: for some brands, they should not. A brand with no product people talk about, no staff who can answer a question in public, and no appetite for criticism will do better buying reach and leaving the comment section alone. I have watched teams open a community channel, post twice a week for a month, hear silence, and declare the whole idea dead. The idea was fine. The brand simply had nothing to say that a stranger would repeat.
\nSo why bother? Because the same facts that make community slow also make it cheap to learn from. Organic posts cost staff time rather than media dollars. They show you, before you spend, which message earns a reply, which image earns a save, which question earns an argument. Paid media then gets a shortlist of tested ideas instead of a guess. This chapter covers how to build that test bed, how to promote what works, how to handle other people's content with permission, and how to keep the whole thing from collapsing into one more sales channel. Running social accounts day to day is a separate job; here we focus on the handoff to paid.
\nBefore any seeding plan, ask a plain question. If this brand disappeared from the community tomorrow, would anyone notice or care? If the answer is no, you do not yet have standing, and the first job is to earn it with useful contributions, not to schedule promotion.
\nStanding comes from subject knowledge members value, access they cannot easily get elsewhere, and consistent presence over time. Audit the places where your audience gathers: platform groups, private messaging groups, niche forums, and brand-owned spaces. Read the written rules, the culture, and the top posts first, then start with no more than two. If the community includes minors, or centers on a sensitive topic such as health, money stress, or grief, do not seed with promotional content at all. Answer questions, point to reliable resources, and keep paid activity separate from that space.
\nPaid creative testing is expensive because every bad idea costs media dollars. Organic posting lets you run many cheap tests first, provided you treat the posts as experiments rather than as announcements.
\nThink of a baseball farm system: prospects play lower-stakes games, and only proven ones get called up.
\nA usable test bed has structure. Pick four to six message angles for a month: a customer problem, a how-to, a product detail, a behind-the-scenes story, a question to the audience, a customer story. Post each angle in more than one format, such as a single image, a short video, a text-only post, and a poll, so you are not confusing the angle with the format. Which numbers? Reach and likes flatter you. Replies, saves, shares, and profile or link clicks tell you more, because they require effort. Read the comments themselves, too. A post with modest numbers but 20 replies asking where to buy is a stronger signal than one with high reach and silence.
\nA caution: organic reach depends on who the platform chose to show the post to, so an organic winner earns a paid test, not a paid budget.
\nPromotion decisions made by feel go wrong in predictable ways. The loudest person in the meeting favors the post with the best photo. The founder favors the post she wrote herself. A written rule removes both problems.
\nHere is a simple version. A post qualifies for a paid test when it meets three conditions after a fixed window, say seven days. It sits in the top quarter of your own posts for effortful actions (replies, saves, shares) relative to reach. Its comments are mostly on topic and positive or curious in tone. And it makes a claim you can stand behind in an ad without added legal review, meaning no health promises, no price promises that may expire, and no testimonial you cannot document.
\nConsider an illustrative case. Trailhead Provisions, a fictional trail snack brand, posts 12 organic pieces over four weeks. Three clear the bar. One is a short video of a staff member repacking a gear bag after a rainy weekend, one is a poll about the worst place to find a crushed granola bar, and one is a customer photo from a summit. The team puts a small, equal test budget behind each of the three, targeting a cold audience and a retargeting audience separately, and runs them against its current best ad. Suppose the repacking video produces a cost per purchase around 31 dollars with cold traffic while the control sits near 38 dollars, the poll draws cheap clicks but few purchases, and the summit photo cannot be used because the rights check fails. That is a healthy result. One winner, one traffic builder, one dropped. Nobody pretends all three succeeded.
\nAnd retire winners. A promoted post that ran well for three weeks will fatigue. Treat the end of its run as information and send the next candidate up from the test bed.
\nCommunity activity does not only supply creative. It also changes the economics of the paid campaigns around it, in several ways that you can observe even when you cannot prove them precisely.
\nFirst, social proof. An ad whose comments show real customers answering questions looks different from an ad with no visible conversation. Shoppers read comments before they trust copy. When your community members reply with specifics, the paid post carries that proof along.
\nSecond, audience quality. People who joined a community, saved a post, or asked a question are warmer than a stranger. Build audiences from them where platform rules and your privacy policy allow, and use them as seed groups for similar-audience targeting or as retargeting pools. Warm pools usually convert at lower cost, though they are small and wear out quickly, so do not plan the whole budget around them.
\nTo see whether signals are real in your account, compare the same ad running with and without the organic engagement attached for a defined period, or compare cost per result for audiences that touched the community against those that did not.
\nCustomer content is attractive precisely because it is not yours. That is also the problem. A photo posted by a customer belongs to the customer, and posting it publicly does not give you the right to run it in a paid ad. Operational guidance only here; your counsel decides what your contracts and local law require.
\nBuild a standard process in five steps.
\n1. Ask in writing, in the same place the content appeared, and say plainly that you want to use the content in paid advertising, not just reshare it. 2. Record the answer with the member's handle, the date, the exact asset, the channels, and the time period. 3. Get a short signed or clearly affirmative permission form when the content shows a recognizable person, a child, a license plate, another brand's logo, or private property. 4. Confirm any music, artwork, or third-party material in the content is also cleared, since the customer may not have rights to it either. 5. Store the record where the media buyer can find it before launch.
\nFor anything involving minors, the default is no, unless a parent or guardian gives documented consent and your counsel agrees. Do not run content from a community in which members are young unless that review is done.
\nFinally, plan for withdrawal. People change their minds, delete accounts, or ask you to take their image down. Decide before launch how fast you will pause an ad when that happens. Same business day is a sensible target.
\nA community with no standards drifts. Spam creeps in, arguments get personal, and the brand's silence gets read as approval. A community with too many rules goes quiet. The middle path is a short written code and a response routine that the team can follow on a bad day.
\nStart with a code of conduct of no more than eight lines: be respectful, stay on topic, no unsolicited promotion, no personal data, no harassment, report problems to a moderator. Post it where members see it. Then write response standards by category rather than by individual post.
\nQuestions get an answer from a named person, ideally within one business day, with a link only when the link truly helps. Praise gets a thanks that mentions something specific. Complaints get a public acknowledgment, a move to a private channel for details, and a follow-up in public when resolved, so others see the outcome. Hostile or abusive posts get removed under the posted code, with a quiet note to the author. Misinformation about the product gets a calm correction with a source. Anything touching safety, legal claims, or a minor goes to a named escalation owner at once, and the team does not improvise.
\nHere is the failure I see most. A brand builds a community, promotes winners from it, and gradually starts using the space as a place to announce sales. Members notice. Replies drop. The test bed no longer produces tests because the audience has stopped speaking freely.
\nA reasonable guardrail is a ratio. For every promotional post in the community, publish at least four that carry value with no ask: a tip, an answer, a member spotlight with permission, a useful link from someone else, a question. Adjust the ratio to the culture of the space, but keep one.
\nSeparate the channels by purpose. The community is for conversation. Paid ads are for reaching new people. Email and site pages are for conversion. If a member comes to the group and finds discount codes at every turn, the group has become a catalog. Offer promotions in a pinned, clearly labeled thread, and let members choose to read it.
\nPicture a loop with six stops. It starts at the test bed, where the team publishes a planned mix of organic posts and reads the response. From there, winners pass through the selection rule and move to the rights check, where any member content is cleared. Cleared assets go to paid testing, where small budgets run against the current best ad. Proven assets scale in paid and bring new people. Those new people arrive at the community, where the welcome routine greets them and the moderation standards keep the space healthy. Their questions, stories, and photos become the raw material for the next round of organic posts, and the loop returns to the start.
\nMeasurement follows the loop. Tag every paid link and every community link with campaign, source, and content labels so the traffic can be traced. Count the loop by stage: number of posts tested, number of winners, number cleared, number promoted, cost per result against control, new members per week, and replies per post. Compare cohorts of promoted content against non-promoted content in the same weeks. Once spend is meaningful, run an audience or geographic holdout: withhold the community-sourced creative from a matched set of audiences or regions and compare results. Last-click reports will undervalue the loop because much of the community effect happens before a final click, and they will overvalue it where members were already buyers. Treat any single number with suspicion.
\nLoops die when they depend on inspiration. A fixed routine keeps them alive with about six to eight staff hours a week for a small program.
\nMonday: Review last week's log. Mark candidates that clear the selection rule. Send rights requests for any member content in the shortlist.
\nTuesday and Wednesday: Publish the first test-bed post, answer unanswered comments, brief the media buyer, and launch small paid tests with a capped budget and end date.
\nThursday: Publish the second test-bed post. Moderate the queue, update the cleared-assets sheet, and check that any withdrawal requests have been handled.
\nFriday: Pull the numbers for the week: organic results, paid test results, new members, response times. Write five lines about what you learned and what changes next week. Retire any fatigued ad.
\nSeeding: Contributing useful posts, answers, or assets to a community so that members value the brand's presence before any promotion. Organic to paid loop: A repeating cycle in which unpaid posts are tested, the winners are promoted with paid media, and the people they attract produce the next round of content. User-generated content (UGC): Photos, videos, reviews, or posts created by customers or members rather than by the brand or its agencies. Whitelisting rights: Permission for a brand to run an ad using a creator's or customer's account handle and content, usually limited by time and channel. Holdout test: An experiment that withholds a campaign from a matched group so the difference in results estimates what the campaign added.
\n1. Boosting a post because the team likes it, without a written selection rule, so the loudest opinion decides where media dollars go. 2. Running a customer's photo in a paid ad because it was posted publicly, with no written permission, no usage period, and no record the buyer can check. 3. Posting as an anonymous fan or asking staff and friends to praise the brand in a community, then losing trust the first time someone checks. 4. Treating likes and reach as the success measure for organic tests, which rewards easy content and hides the posts that actually produce replies and clicks. 5. Turning the community into a promotion board, so members stop speaking freely and the test bed stops supplying real signals.
\n1. Write the selection rule before the month begins, with a fixed observation window and a threshold on replies, saves, and shares relative to reach. 2. Keep one cleared-assets sheet that lists each asset, owner, permission date, channels, and expiry, and require a media buyer to check it before launch. 3. Identify every staff member who speaks for the brand, and give each a short voice guide and an escalation path for safety or legal issues. 4. Test at small, capped budgets first, then scale only the assets that beat the current best ad on cost per result. 5. Review the loop monthly and cut one low performing community, format, or angle so the routine stays small enough to sustain.
\n1. Read the comments on your own best organic posts aloud in a team meeting, because customer wording often produces better headlines than anything written in the room. 2. Ask moderators of a community, politely and in private, what the group is tired of seeing from brands. Their answer is worth more than any benchmark. 3. Save the first reply from a new member of the community, since early questions reveal the objections that paid ads should address. 4. Pause promoted posts that attract a pile of off-topic or hostile comments, and tell the community lead so the conversation can be handled in the open. 5. Offer members something to make, not just something to like, such as a simple prompt, a challenge, or a template, because contributions are what feed the loop.
\nPostiz (AGPL license) is a self-hosted social scheduling tool that lets a small team plan the test-bed calendar, queue posts across accounts, and keep the posting log in one place without paying for a scheduling subscription. Matomo (GPL license) is a self-hosted web analytics platform that can record campaign, source, and content labels on community and paid links, so you can see which promoted posts send visitors who stay, sign up, or buy. Both can be run on your own server, which keeps visitor data under your control. Neither one moderates a community or clears rights for you. Those remain human routines. No open source tool matches the audience reach of a large paid platform either, so media still gets bought where the audience already is.
\nCommunity seeding earns its place when a brand has real standing, something useful to say, and people willing to answer questions in public. Organic posts act as a low-cost test bed, a written rule selects which winners earn a paid test, and community signals such as comments, warm audiences, and customer language can lower what paid media costs. User content needs documented permission, clear disclosure, and a fast withdrawal process, while moderation standards and a ratio of value to promotion keep the space from turning into an ad inbox. The loop runs on named owners, tagged links, a weekly routine, and a monthly cut, and it is measured by stage with a holdout test once spend justifies one.
\n1. A brand without standing in a community should earn it with useful contributions before it attempts any promotion. 2. Organic posts are experiments, and an organic winner earns a small paid test rather than a paid budget. 3. A written selection rule beats opinion when deciding which posts receive media dollars. 4. Customer content needs written permission, a usage period, and a record before it appears in an ad. 5. A weekly routine with named owners keeps the loop running, and a monthly cut keeps it small enough to survive.
\nThis week, write a three-condition rule for promoting organic posts and a one-page cleared-assets sheet, then test your best two posts against your current top ad at a capped budget.
\n1. Federal Trade Commission. Guides Concerning the Use of Endorsements and Testimonials in Advertising, 16 CFR Part 255. 2. Federal Trade Commission. Disclosures 101 for Social Media Influencers. Plain-language guidance on disclosing material connections. 3. U.S. Copyright Office. Copyright Basics circular. Overview of ownership and permission for creative works. 4. Matomo. Official documentation on campaign tracking and self-hosting. GPL licensed software. 5. Postiz. Official project documentation on self-hosted scheduling. AGPL licensed software.
"},{"id":"chapter-19-lead-generation-formats-and-form-quality","title":"Chapter 19: Lead Generation Formats and Form Quality","html":"By the end of this chapter, the reader will be able to: 1. Compare in-platform lead forms with landing page forms and choose between them for a stated offer and budget. 2. Design friction into a form deliberately so that it screens for intent without choking volume. 3. Apply layered defenses that catch bot submissions, fake contact details and duplicate junk before they reach a sales team. 4. Construct a simple lead scoring sheet and a follow-up standard with clear response times. 5. Evaluate consent language and call-back practices against plain-language TCPA exposure, then escalate gaps to counsel. 6. Calculate cost per lead and cost per qualified lead, and design a feedback loop that teaches ad platforms to chase the second number.
\nPicture a regional home-services brand I will call Slatewick Roofing (illustrative). The marketing manager, a careful person, reads that shorter forms convert better. She cuts her lead form to a name, a phone number and an email. Cost per lead drops by more than half within a week. She reports the win. Her owner is delighted.
\nThree weeks later the estimators are miserable. They are calling numbers that ring through to strangers, a laundromat, a dentist's front desk. Many of the people who do answer say they only wanted the free checklist. A few say they rent. One asks, politely, how the company got her number at all. The sales team stops returning to the lead queue after lunch. By month's end the campaign has produced a spreadsheet that looks like a triumph and a calendar with almost no booked inspections.
\nNothing she did was foolish. The advice is true in some settings and ruinous in others. That is the assumption this chapter challenges: that a lower cost per lead means a better lead program. A lead is a claim that somebody might buy. A form fill is only the receipt for that claim, and anyone can print a receipt.
\nThink of a stadium turnstile. Counting the people who walked through says nothing about who bought a ticket and who slipped in behind a group. The work here is learning who is actually in the seats.
\nAn in-platform form opens inside the social app, prefilled with the profile details the person already gave that platform. A landing page form sends the person to your website, where they type everything themselves. The first is a short tap. The second is a small commitment.
\nThe in-platform form wins on volume and cost per lead, almost without exception, because the effort is close to zero. The trade is that prefilled data is often stale, mistyped years ago, or a throwaway address the person created for exactly this purpose. They also submit from a scrolling, half-attentive mood. Scrolling is not shopping.
\nThe landing page wins on control. You choose the words, the order of questions, the disclosure language, the page speed, the tracking, and what happens after the submit. You can show pricing ranges. You can explain what happens next. You can screen people out politely before they cost your team a phone call.
\nA decision rule that holds up in most accounts: use in-platform forms when the offer is low commitment and your follow-up is fast and automated (a free guide, a giveaway entry, an event reminder). Use landing pages when a human has to spend real time on each lead (a quote, a demo, a consultation, a financing application). Many mature programs run both and compare them on cost per qualified lead, the number we will build near the end of this chapter. Messaging and call-based formats work when the buyer wants an answer more than a download, and they need staff who can respond in minutes.
\nFriction is every small effort between a person and the submit button. Most advice treats it as an enemy. I treat it as a dial.
\nTurn it down for a free, low-stakes offer, because a person who wants a checklist should not need to prove anything. Turn it up when your cost to serve each lead is high. A roofing estimator driving 40 minutes to an inspection is a very expensive follow-up. A question or two before that drive is cheap insurance.
\nUseful friction comes in a few forms:
\nNotice what these have in common. Each one asks for something only a real, interested person can supply easily. Useless friction, by contrast, is a long form, a forced account creation, or an unexplained request for a birth date. That kind drives off buyers and keeps bots, which do not get tired.
\nTest friction the way a coach tests a lineup: change one thing, hold the rest steady, and judge by wins. Add a single qualifying question. Cost per lead will rise. Then watch qualified leads and booked work. If those improve, the higher cost per lead bought you a better pipeline.
\nJunk comes in several kinds, and each needs its own defense. Bots submit automatically. Fakers type nonsense by hand to grab an incentive. Duplicates resubmit to claim the offer repeatedly. Competitors and curious researchers fill forms to see what you send back.
\nBuild defenses in layers, because no single check catches everything.
\nAt the form: add a hidden field that humans never see and bots fill in (often called a honeypot), a short time-on-page threshold since a real person cannot complete five fields in two seconds, and a simple challenge on landing pages when abuse appears. Prefer a self-hosted or privacy-respecting challenge to a third-party tracker.
\nAt the data: validate email syntax, check that the address domain can actually receive mail, and flag disposable-mail domains. Normalize phone numbers into one format and flag obvious fakes such as repeated digits or sequences. Compare against recent submissions to catch duplicates, matching on phone and email, not name.
\nAt the confirmation: send a confirmation message with a link or code before the lead enters the sales queue. This single step, sometimes called double opt-in or verification, removes a large share of bad contacts in most programs, at the cost of some real people who never click. For high-value leads, I accept that trade. For giveaways, I often do not.
\nAt the source: look at where junk concentrates. A single placement, a single audience, or a single creative may produce most of the fakes. Cut that source and the average quality rises without touching your form at all.
\nKeep a rejected-lead log, so you can see whether your filters are throwing out real buyers.
\nLead scoring is a way to rank leads by how likely they are to buy, so that people spend their hours on the best ones first. It does not need software or statistics to start. It needs a short sheet and honest agreement between marketing and sales.
\nStart with two kinds of points. Fit points describe who the person is: lives in your service area, right size of business, right role, project type you actually serve. Intent points describe what they did: chose a near-term timeline, answered the free-text question, asked for a call rather than a download, opened your confirmation message.
\nHere is a small illustrative scheme. Service area match, 20 points. Timeline within a month, 25 points. Project size above your minimum, 20 points. Typed a real answer to the open question, 15 points. Verified phone number, 20 points. Subtract points for a disposable email address or a mismatch between stated city and phone area code. A lead scoring 70 or more goes to a human immediately. A lead between 40 and 69 enters an email sequence. Anything lower is held, not deleted.
\nThen do the part most teams skip: review the scores against outcomes every month. If leads scoring in the 50s close as often as leads in the 80s, your scheme is guessing. Reweight it. Sales will tell you which points were fantasy, and the closed-won records will back them up.
\nScoring also gives you the word \"qualified.\" Define it in writing, such as \"score of 70 or above and reached by phone,\" or marketing and sales will argue forever.
\nInterest decays quickly. The person who tapped your ad on a phone is, within an hour, back in a conversation, a meeting or a game. Your competitor may have asked the same person for details a minute earlier.
\nSet a standard and publish it internally. A reasonable starting point for high-intent leads: an automatic text or email confirmation immediately, a human call attempt within five to 10 minutes during business hours, and further attempts spread across several days and at different times. Messages should say who you are, what the person asked for, and how to stop contact. Lower-intent leads can move more slowly, but they should never wait for someone to remember them.
\nMeasure these things, because the standard is a promise only if someone checks it: 1. Median minutes from submission to first human attempt. 2. Share of leads reached at all. 3. Share of leads reached on the first attempt. 4. Time of day when leads go cold in your queue.
\nA common finding is that the campaign performed fine and the follow-up did not. Evening submissions sat untouched until the next afternoon, and the cost per qualified lead looked terrible. The fix was staffing, a rotating call shift, or an automated message that sets expectations honestly (\"We will call you tomorrow morning at nine\"). Fix the operation before blaming the ad.
\nThis section is operational guidance, not legal advice. Your counsel decides what applies to your business, your states and your calling methods.
\nIn the United States, the Telephone Consumer Protection Act (TCPA) governs many calls and text messages sent to consumers, especially those made with automated dialing technology or prerecorded or artificial voices. The core idea in plain terms: before you send marketing texts or automated calls, the person generally needs to have clearly agreed, and you need to be able to prove they did. Penalties and lawsuits connected to this area have been serious for businesses that were careless, and the risk tends to land on the brand whose name is on the message, even when an agency or vendor sent it.
\nPractical habits that reduce exposure:
\nAd platforms learn from the signals you give them. If the only signal is \"form submitted,\" the system finds more people like those who submit forms, including the tappers, the fakers and the bored. It is doing exactly what you asked.
\nClosing the loop means sending back what happened after the form. When a lead becomes qualified, books an appointment or buys, you report that event to the platform, usually through its server-side conversion interface or an offline upload, and the algorithm begins to prefer the kind of person who did those things.
\nThree habits make this work. First, tie every lead to a click identifier or similar tag at submission, and store it with the record. Second, define one or two meaningful events (for example \"qualified\" and \"booked\") and report those, not twenty minor ones. Third, be patient about delay: if your sales cycle takes weeks, report the earliest reliable event, such as a qualified phone conversation, so the platform gets a usable signal while the campaign is still running.
\nThe new signal is thinner than raw form fills, so if qualified events are too few, optimize toward a middle event first and graduate later. Send only data you have consent to share, and let counsel review what identifiers leave your systems.
\nSuppose a brand spends 4,800 dollars on each of three setups over the same period (all figures illustrative). The first is a short in-platform form. The second is the same form with one added timing question and a confirmation message. The third sends traffic to a landing page with a fuller form.
\n| Setup | Leads | Cost per lead | Qualified leads | Cost per qualified lead | |---|---|---|---|---| | Short in-platform form | 320 | 15 dollars | 32 | 150 dollars | | In-platform with question and confirmation | 200 | 24 dollars | 50 | 96 dollars | | Landing page form | 120 | 40 dollars | 40 | 120 dollars |
\nRead it from left to right and the first setup looks like the champion. Read it from the right and it is the most expensive way to find a real buyer. The second setup costs 60 percent more per lead and is the cheapest per qualified lead by a clear margin. The landing page lands in between, and may still win if its qualified leads close at a higher rate.
\nCarry the arithmetic further when you can, to cost per booked appointment and cost per sale. With small samples a few odd weeks can flip the ranking, so let tests run long enough to trust.
\nLead form ad: A paid social format that opens a form inside the app, often prefilled from the person's profile, so they can submit without visiting a website. Qualified lead: A lead that meets a written definition of fit and intent, agreed by marketing and sales, such as a verified contact in your service area who wants service soon. Speed to lead: The time between a person submitting a form and a human or automated first response. Double opt-in: A two-step confirmation in which the person must verify their contact detail, usually by clicking a link or entering a code, before entering your follow-up system. Offline conversion upload: A method of sending outcomes that happen after the click, such as booked appointments or sales, back to an ad platform so it can optimize toward them.
\n1. Judging a campaign by cost per lead alone, which rewards the cheapest and least interested audience and hides the real cost of finding a buyer. 2. Shortening every form to the minimum fields without asking what the sales team spends to chase each submission. 3. Running consent language as a footnote, with a pre-checked box and no saved record of what the person saw. 4. Letting leads sit unattended overnight or over weekends, then blaming the ad creative for poor results. 5. Reporting only form submissions back to the ad platform, which trains the algorithm to find more people who tap forms rather than people who buy.
\n1. Write the definition of a qualified lead with sales before building the form, then design every question to predict it. 2. Test one form change at a time and judge each by cost per qualified lead, not cost per lead. 3. Layer defenses at the form, the data, the confirmation and the traffic source, and keep a log of rejected leads. 4. Store consent wording, form version, timestamp and source with every lead, and check suppression lists before every message. 5. Report qualified and booked events back to the ad platform, tied to click identifiers captured at submission.
\n1. Use in-platform forms for light offers with automated follow-up, and landing pages when a person must invest real time per lead. 2. Add one typed question that only an interested person will bother to answer, and read the answers weekly. 3. Check the hour of submission against the hour of first contact, because the gap often explains weak results. 4. Compare placements and audiences by junk rate before touching creative, since a single placement can poison an average. 5. Review lead scores against closed sales every month and reweight the points that failed to predict anything.
\nThree free, self-hostable tools cover the follow-up side of this chapter. Mautic (GPL license) is a marketing automation platform that handles lead scoring, forms, segments and nurture sequences, and records contact activity. Listmonk (AGPL license) is a fast newsletter and mailing list manager suited to confirmation emails, double opt-in flows and simple sequences. Postal (MIT license) is a mail delivery server you run yourself, which gives you control over sending and bounce handling. Together they can verify, score and nurture leads without paid subscriptions. None replaces the ad platform's own form or conversion interface, and you remain responsible for hosting, security and sender reputation.
\nA low cost per lead is a receipt, not proof of demand. In-platform forms deliver volume and speed, landing pages deliver control, and the right choice depends on what each lead costs your team to pursue. Deliberate friction, layered bot defenses and a written scoring sheet turn raw submissions into a queue worth calling. Fast follow-up, honest consent language with saved records, and conversion data returned to the ad platform keep the whole system pointed at buyers. The worked illustration shows how the cheapest lead can be the most expensive qualified lead.
\n1. Cost per lead is an input; cost per qualified lead is the number that tells you whether the program works. 2. Friction is a dial you set by your cost to serve each lead, not a flaw to remove. 3. Bot and junk defenses work in layers: the form, the data, the confirmation and the traffic source. 4. Speed to lead, written consent language and saved consent records are operating standards, with counsel deciding the legal specifics. 5. Platforms optimize toward whatever signal you send, so send qualified and booked events rather than form fills.
\nA lead campaign should launch with a written definition of a qualified lead, one added question that predicts it, and that event reported back to the platform. Then judge the campaign by cost per qualified lead, never by cost per lead.
\n1. Federal Communications Commission. Telephone Consumer Protection Act rules and consumer guidance, including consent requirements for calls and texts. 2. Federal Trade Commission. Dot Com Disclosures: How to Make Effective Disclosures in Digital Advertising. 3. Federal Trade Commission. CAN-SPAM Act: A Compliance Guide for Business. 4. Mautic project documentation. Contact scoring, forms and campaign automation. 5. Listmonk project documentation. Subscriber management and opt-in confirmation.
"},{"id":"chapter-20-frequency-fatigue-and-refresh-cadence","title":"Chapter 20: Frequency, Fatigue, and Refresh Cadence","html":"By the end of this chapter, the reader will be able to: 1. Explain why average frequency misleads and describe frequency as a distribution of people. 2. Estimate where diminishing returns begin for a campaign using reach and conversion data by exposure count. 3. Identify creative fatigue from leading indicators before cost per result moves. 4. Apply a fatigue detection checklist to a live paid social account. 5. Design a refresh cadence and a set of suppression rules scaled to budget size and objective. 6. Evaluate whether a refresh worked using a holdout comparison rather than a before and after glance.
\nOpen the dashboard and it will tell you the campaign has a frequency of 2.4. Green light. Healthy. Move on to the next meeting.
\nNow put the same campaign in front of a finance partner who has never heard the word \"impressions.\" She asks one question: how many of those 2.4 were the same person, and what did the last few cost us? You do not have an answer. The dashboard divided total impressions by total people reached and handed you a number that describes nobody. Somewhere inside that average, a large group saw the ad once and drifted off. A smaller group saw it 16 times and quietly decided they dislike your brand. Both groups are folded into 2.4, and the budget was spent on both.
\nThat is the contradiction this chapter lives inside. The platform reports a single comfortable figure. A person who has to defend the spend would reject it in about 30 seconds, because averages hide exactly the cases where money leaks.
\nThe usual assumption is that fatigue is a creative problem you notice when results sour, and you fix it by commissioning something new. I think that gets the order backwards. Fatigue is a distribution problem first, a measurement problem second, and a creative problem third. By the time cost per result climbs, you have already paid for weeks of decline. The teams that stay ahead of it do not have better artists. They look at different numbers, they decide in advance what counts as too much, and they build the replacement ads before the old ones are tired.
\nNone of this requires a large budget. It requires a habit of asking who saw what, and how many times.
\nFrequency is impressions divided by reach, where reach is the count of unique people who saw the ad at least once. It is a ratio, and ratios flatten shapes. Two campaigns can share an average of three and behave nothing alike.
\nPicture a campaign with 100 thousand people reached. In one version, nearly everyone saw the ad two to four times. In the other, 70 thousand people saw it once, and five thousand people saw it 20 times or more while the remaining 25 thousand sat in between. The averages match. The experiences do not. The second campaign is wasting spend on a small, over served pocket and under serving everyone else.
\nSo stop asking for the average and ask for the histogram: how many people at one exposure, two, three, and so on up to a final bucket for 10 or more. Most ad platforms report frequency breakdowns in some form, either as exposure buckets or as a downloadable reach and frequency report. If yours only gives the average, export the log level data where you can and build the buckets yourself.
\nThen read the tail. The right edge of that distribution is where fatigue lives. A useful working number is the share of reached people in your top bucket. If that share is growing week over week while the average sits still, the platform is concentrating delivery on a narrowing group, often because your targeting is small or your bids favor people who respond. The average will not warn you. The tail will.
\nThink of a stadium. Average attendance across the season tells you little about the fan who sat in the same seat for the same promotional giveaway 11 games in a row.
\nEach additional exposure does something less than the one before it. The first impression introduces the brand. The second confirms it. The third may nudge a person who was thinking about acting. Somewhere after that, the additional lift per impression shrinks toward zero, and for some people it turns negative.
\nThere is no universal number for where this happens, and anyone quoting one is guessing. Your own data can show it. Take conversions per thousand impressions and group them by the exposure count the person had reached when the conversion occurred. Suppose the first exposure produces a conversion rate you set as a baseline of 100 index points, the second produces 85, the third 70, the fourth 52, and by the seventh the figure sits at 20. Those numbers are illustrative, but the shape is common: a steep fall, then a long flat tail where you are paying full price for nearly no response.
\nThe practical point is where the marginal cost crosses what the customer is worth. If the fifth impression costs the same as the first and delivers well under half the response, the fifth is not a good buy. It might still be defensible for a high value product with a long consideration period. It is not defensible for a low priced impulse purchase.
\nTwo things bend the curve. Creative quality moves it right: an ad that entertains or teaches holds attention across more exposures than one that merely repeats an offer. Audience size also moves it: a broad audience lets the platform find fresh people, while a narrow one forces repeats. Treat the curve as a property of a specific ad in front of a specific audience, not a law of nature.
\nCost per acquisition is a lagging indicator. It reflects spending that already happened and conversions that may take days to report. If you wait for it to move, you are steering by the rear view mirror.
\nLeading indicators move first, because they measure attention and not purchase. Watch these in roughly this order.
\nClick through rate on the ad usually slips before anything else. People stop reacting to something they have seen. A steady downward drift over several days at a stable audience is a stronger signal than one bad afternoon.
\nCost per thousand impressions often rises as engagement falls, because many auction systems reward ads that people respond to and charge more for ads they ignore. A rising delivery cost with unchanged targeting hints that the platform has read your ad as less relevant.
\nNegative feedback matters. Hides, \"not interested\" taps, and comment sentiment turn sour when people feel pestered. In family oriented categories, set a stricter tolerance here and treat a small rise as a reason to act.
\nVideo completion rate and early view thresholds drop when the opening seconds are familiar. A viewer who recognizes the first frame swipes away. For video, the hook fatigues before the story does.
\nFinally, watch the gap between frequency in the top bucket and engagement among those people. When the group seeing the ad most often also shows the weakest response, you have found the fatigued segment directly.
\nNone of these alone proves anything. Together, moving the same direction for several days, they justify acting while cost per result still looks fine.
\nRun this weekly for every ad carrying meaningful spend. Any three items flagged means the ad goes on the refresh list.
\nThe thresholds above are starting points, not standards. Calibrate them against your own account history in the first two months, then tighten them.
\nRefresh speed should follow how fast you consume the audience, and that depends on spend relative to audience size more than on any calendar. Still, budget tier gives a workable starting rhythm.
\nA small account, spending a few hundred dollars a week against a narrow local or niche audience, burns through its people quickly. The tradeoff is that it cannot afford a production calendar of new films. Rotate in one fresh variant roughly every two to three weeks, built by changing a single element: the opening line, the first image, or the offer framing. Keep the proven body of the ad.
\nA mid sized account, spending a few thousand dollars a week, should run a standing rotation of perhaps three to five active ads per audience and retire the weakest as the checklist flags it. Plan a new batch on a monthly rhythm, with enough variety that the platform has real options.
\nA large account, spending tens of thousands a week or more, usually has the opposite problem: volume of creative. Fatigue arrives by ad, not by audience, so the work is a continuous pipeline with new concepts entering testing every week and winners promoted. Here the discipline is concept variety, not slight edits, because 20 variations of the same idea tire together.
\nA rule beneath all three: the cadence is a ceiling on how long you tolerate, not a promise to refresh on schedule. If an ad is still strong at week four, leave it. Replacing a working ad because the calendar said so is how teams throw away their best performer.
\nA frequency cap limits how many times one person can see an ad within a time window. Suppression excludes specific people from seeing it at all. They solve different problems, and you will usually need both.
\nCaps protect against the tail. If your distribution shows a pocket of people at 16 exposures, a cap of, say, four per week is a direct response. Not every placement or buying type supports a hard cap, particularly on some auction based social buying where you instead steer with reach and frequency reservation or bid and budget settings. Know which control your platform actually gives you, and do not assume the setting exists because the concept does.
\nSuppression is about who should not see the ad. Existing customers should usually be excluded from acquisition messages, and recent purchasers from promotions for the thing they just bought. People who converted, those who asked to stop hearing from you, and anyone on an opt out list belong in an exclusion audience. Build those exclusions from your own first party customer lists where permitted, and keep them current, because a stale list pays to annoy your best customers.
\nAdd a sensible cooling off period for people who engaged negatively. Someone who hid your ad should not see its sibling the next morning.
\nFor audiences that may include minors, or sensitive categories, set a conservative default: lower caps, narrower targeting, and a rule that nothing escalates frequency simply because it performs. Operational guidance only here; counsel and platform policy decide what is permitted.
\nA single cap across a whole account is a sign nobody thought about the goal behind each campaign.
\nAwareness campaigns want broad, even reach. Their job is to introduce the brand to as many distinct people as the budget allows, so a tight cap is usually right, because every repeat exposure displaces a first one. For cross-screen planning, where the same person may meet your ad on a phone, a TV, and a laptop, frequency has to be managed across devices, a problem Chapter 27 takes up.
\nConsideration and retargeting campaigns work differently. The audience already knows you. A higher cap is reasonable because the message is closer to a decision, but it needs a time limit: perhaps a week or two of intensity, then rest.
\nConversion campaigns aimed at a hot, small audience, such as cart abandoners, tolerate more repeats in a short window and then decay quickly. A person who has not bought after a week of reminders is rarely moved by a 10th. Stop, or change the message entirely.
\nLaunch moments and events are different again. A time boxed promotion may justify a deliberate spike, accepted in advance and ended on a date.
\nThe principle: the cap expresses how much repetition the objective can earn. State it per campaign, write down why, and review it when results shift.
\nComparing performance before and after a refresh is tempting and unreliable. Seasons turn, competitors change bids, and a news event can move everything at once. You may credit the new creative for a lift that came from the calendar.
\nA cleaner method is a holdout. Split a comparable audience into two cells and keep the old ad running in one while the new ad runs in the other, with the same budget, timing, and placements. Compare the leading indicators first, then conversions once enough have accrued. Suppose the new ad shows click through rate roughly a fifth higher and the same cost per conversion after two weeks. That is a modest result worth recording, not a triumph.
\nDecide your success measure and the length of the test before launching, so you cannot choose a flattering window afterward. Record what element changed, what the hypothesis was, and what happened. After a dozen entries, that log tells you which kinds of changes extend an ad's life and which are noise. A refresh that nobody measured teaches nothing, and you will repeat the same guess next quarter.
\nReach: The count of unique people who saw an ad at least once in a given period. Frequency cap: A setting that limits how many times one person can see an ad within a time window. Suppression audience: A list of people deliberately excluded from seeing a given campaign, such as existing customers. Creative fatigue: The decline in response to an ad as the same audience sees it repeatedly. Holdout: A comparison group kept on the old treatment so a change can be judged against a fair baseline.
\n1. Trusting the average frequency on the dashboard and never looking at the share of people in the highest exposure buckets. 2. Refreshing on a fixed calendar date, which retires strong ads early and lets weak ones linger. 3. Waiting for cost per acquisition to rise before acting, when click through rate and negative feedback signaled the problem days earlier. 4. Replacing an entire ad when only the opening line or first image had tired. 5. Applying one frequency cap to every campaign regardless of whether the goal is reach, consideration, or conversion.
\n1. Export or build the exposure histogram for each major campaign monthly, and track the top bucket share as a standing metric. 2. Keep a tested replacement ready for every ad carrying significant spend before it shows fatigue. 3. Change one element at a time on small and mid sized accounts so you learn which element mattered. 4. Maintain current suppression audiences for customers, opt outs, and negative responders, and refresh them on a schedule. 5. Judge every refresh with a holdout and a written hypothesis fixed before launch.
\n1. Narrow audiences fatigue faster, so expand targeting slightly before commissioning new creative when reach has flattened. 2. Ads that teach or entertain tend to last longer than ads that repeat an offer, so lead with value where the category allows. 3. Check negative feedback weekly even when conversions look strong, because irritation is cheap to fix early and expensive to repair later. 4. Retire retargeting ads on a time limit, then rest the audience before returning with a different message. 5. Keep a simple refresh log with date, element changed, and result, since patterns appear only after many entries.
\nDuckDB (MIT license) is an embedded analytical database that runs on a laptop and reads exported platform files directly. Load your reach and frequency exports, or log level delivery data where you have it, and write a short query that groups people by exposure count to build the histogram described earlier. Metabase (AGPL license) is a self-hosted dashboard tool that can sit on top of those queries and chart the top bucket share, click through rate drift, and negative feedback week by week. Together they replace the single average figure with the distribution your finance partner would actually accept. Neither tool buys media, and neither includes the platform data itself, so you still depend on what each ad platform will export.
\nAverage frequency describes no actual person, and the real risk sits in the tail of the distribution, where a small group sees an ad again and again while many others see it once. Returns diminish with each exposure, and your own conversion data can show where the marginal impression stops paying. Fatigue shows up first in leading indicators such as click through rate, delivery cost, negative feedback, and early video completion, well before cost per result moves. A weekly checklist turns those signals into a decision. Refresh cadence should follow budget and audience size, treated as a ceiling and not a schedule, while caps and suppression should reflect what each objective can earn. A holdout comparison tells you whether any refresh actually helped.
\n1. Frequency is a distribution of people, and the top exposure buckets matter more than the reported average. 2. Diminishing returns can be measured in your own data by grouping conversions by exposure count. 3. Leading indicators such as click through rate and negative feedback reveal fatigue before cost per result does. 4. Refresh cadence scales with budget and audience size, and should respond to signals rather than a calendar. 5. Caps should differ by objective, and a holdout is the fair way to judge a refresh.
\nThis week, export an exposure histogram for your largest campaign and write down the share of people at 10 or more exposures. Put that number on the dashboard next to the average, and make three checklist flags the trigger for your next refresh.
\n1. Media Rating Council. Viewable Impression Measurement Guidelines. Standards for impression validity and exposure counting. 2. Media Rating Council. Cross-Media Audience Measurement Standards. Reference on de-duplicated reach across platforms. 3. DuckDB Foundation. DuckDB documentation. Reference for analytical queries on local files. 4. Metabase. Metabase documentation. Reference for self-hosted dashboards.
"}]},{"title":"Part IV: Video, Connected Tv, and Audio","items":[{"id":"chapter-21-online-video-buying-skippable-non-skippable-and-bumper","title":"Chapter 21: Online Video Buying: Skippable, Non-Skippable, and Bumper","html":"By the end of this chapter, the reader will be able to: 1. Distinguish skippable in-stream, non-skippable in-stream, and bumper formats by their billing event, length, and viewer control. 2. Explain why a \"view\" means different things on different platforms and why that changes how reports are compared. 3. Match each format to a reach, consideration, or action objective and defend the choice. 4. Evaluate a placement report for quality problems and build an exclusion list from it. 5. Interpret the core elements of a VAST tag well enough to catch a broken or misleading one. 6. Design a pre-launch quality check for video files using free, open source tools.
\nSkippable video is the best value in online video. Plenty of buyers repeat that line because the billing model sounds so fair: you pay only when someone chooses to keep watching. Why would anyone pay for a person who walked away?
\nNow the strongest counterargument. A skip is not a free event. The viewer who skipped still saw the first few seconds of your brand, and you paid nothing for that exposure, which is good. But the viewer who stayed may have stayed because the video autoplayed on a second screen, because a phone was face down, or because the thumb was busy elsewhere. A paid view is a rough proxy for interest, not proof of it. And a format that charges only on a view tends to shape creative into a trap, where every second is written to prevent the skip rather than to say something worth remembering.
\nSo the claim survives only with conditions attached. Skippable video is a strong value when the creative earns its first moments, when the view definition is understood, and when the objective is consideration rather than guaranteed reach. Change any of those three and a different format may win.
\nI have watched brand managers spend a quarter running one 30 second spot across every video line item, then wonder why the reports looked so uneven. The formats were never interchangeable. They behave like different positions on a field: a long snapper, a short-yardage back, a specialist who comes in for one play. You would not ask all three to do the same job. This chapter treats each format as a tool with a specific purpose, shows where the reporting hides differences, and gives you a way to read the technical plumbing so a vendor cannot hand you a tag you cannot check.
\nSkippable in-stream ads play before, during, or after other video content, and the viewer can skip after a short countdown. The buyer typically pays on a view event rather than on every impression, which makes the format forgiving to test. Length is flexible, but the real unit of design is the opening few seconds. If the brand and the reason to care are not on screen before the skip option appears, the ad is asking the viewer to do you a favor.
\nNon-skippable in-stream ads remove the choice. They are usually bought on impressions, and the length is capped by the platform, so the creative has to hold someone who cannot leave. Exposure is guaranteed. Attention is not. A person can look at a kitchen wall while your spot plays, and you will still be billed.
\nBumpers are the shortest unit, a few seconds at most, unskippable, and bought on impressions. They are not a trimmed version of a longer ad. A bumper carries exactly one idea: a name, a product shot, a date, a single offer. They are well suited to repeating a message the audience already half knows. They are poorly suited to explaining anything that needs a second sentence.
\nA useful way to hold all three: skippable video sells the chance to earn attention, non-skippable sells the right to demand it, and bumpers sell a cheap reminder. Each costs something different, and each can be wasted in a different way.
\nThis is the part of video buying that costs people money without them noticing. Platforms define a billable or reportable view differently. One may count a view after the viewer passes a time threshold or interacts, another may count after a shorter threshold, and another may count a percentage of the ad. Some count a view when the player simply starts. The definitions change over time, so treat the platform's current documentation as the source, not memory or a vendor deck.
\nThe consequence is simple. Two reports that both say \"views\" can describe very different amounts of watching. A cost per view on one platform cannot be placed next to a cost per view on another without first translating both into the same measure, such as cost per completed view, cost per thousand viewable impressions, or cost per second watched. Even then, autoplay and sound settings alter what the number means.
\nBuild a one-page definition sheet for every platform in the plan. For each, record the billing event, the view threshold, whether sound is on by default, whether the placement autoplays, and which third-party verification (if any) can be attached. Put that sheet at the front of the media plan. When a stakeholder asks why platform A looks cheaper than platform B, the answer is on page one rather than buried in an email thread.
\nCompletion rate is the share of started plays that reach the end. On non-skippable inventory it will almost always look healthy, because the viewer has no exit. That does not make it a quality signal. A high completion rate on a non-skippable placement is partly a feature of the format, so comparing it to skippable completion is comparing different species.
\nViewability is the check on whether the ad had a fair chance to be seen. Industry standards for video generally require that a majority of the player be in view for a short stretch of continuous playback. Vendors can attach verification to measure this, and the Media Rating Council guidelines in the references describe the accredited approach. A few cautions apply. A placement can be viewable and still unwatched. A player can be technically on screen while muted, paused in a background tab, or auto-advancing through a playlist. And verification coverage differs by environment, so some inventory reports \"not measurable\" instead of a clean pass or fail.
\nTreat completion, viewability, and audible playback as three separate questions. Did it play to the end? Was it on screen? Could anyone hear or read it? A placement that answers yes to all three is worth a premium. A placement that answers yes to only the first is mostly delivering a reporting number.
\nObjectives drive format, and the mapping is looser than most plans admit.
\nFor reach, meaning getting the message in front of as many distinct people as the budget allows, bumpers and non-skippable spots work because they control frequency and guarantee exposure. For consideration, meaning moving people from vaguely aware to actively interested, skippable video earns its place because the people who stay are self-selected. For action, video rarely closes the sale alone. It primes the audience, and the click happens in search, retargeting, or a later visit.
\nHere is an illustrative case. Suppose a regional boat dealer, a fictional composite called Tidewell Marine, has 12,000 dollars for a spring push. A sensible split is not equal thirds. Most of the money goes into a skippable spot built around a real owner story, because the buying decision is slow and the viewers who stay are the ones likely to visit the lot. A smaller share funds bumpers aimed at people who already watched the longer spot, repeating one line and the showroom date. A modest remainder buys a non-skippable placement only where the dealer can verify audible, viewable delivery. The dealer then judges the plan by showroom visits and branded search lift in the test region, not by view counts.
\nThe decision rule underneath: assign each format a job before assigning it a budget, and write the job in the plan. If a line item cannot name its job, it is spending by habit.
\nI avoid exact timings here because platforms revise limits, and your own tests will teach more than my numbers can. The qualitative rules hold up well across changes.
\nFor skippable ads, put the brand and a human reason to care into the opening moments. Do not spend them on a logo animation. After that, the creative can run as long as the story earns, but a longer ad should carry a payoff the viewer can feel, such as a demonstration, a surprise, or a useful answer. Padding a longer skippable ad with filler simply pays you less for more.
\nFor non-skippable ads, structure matters more than hook. The viewer is captive, so the danger is boredom, not escape. Build a clear beginning, middle, and end. Show the brand early and again near the close. Make sure the spot works with sound off, since a lot of watching happens that way.
\nFor bumpers, write one sentence first. If the idea cannot be said in one sentence, it does not belong in this format. Then show it. Avoid fine print, since nobody can read it in that time, and put required legal language in the landing destination or a longer companion unit where counsel has approved it.
\nOn orientation, produce the creative for the placement: widescreen for most desktop and television-style players, vertical for phone-first placements. Cropping one into the other cuts off faces and text. Plan a few alternate edits at the start of production rather than rescuing a single master at the end. Family-first judgment applies here too: if the audience may include children, keep tone, products, and offers conservative, and check the platform's rules for youth audiences before you target anything near them.
\nVideo inventory ranges from premium long-form programming to low-quality sites that exist mainly to host an autoplaying player. The format name tells you almost nothing about where the ad ran. Placement reports do.
\nPull the placement list after the first week, not the last. Sort by spend, then look at the top placements one by one. Red flags include channels or sites you cannot identify, content that does not match the brand's standards, players that are tiny or hidden, very short plays with suspiciously high completion, and traffic with no corresponding site engagement. In-stream, accompanying content, and standalone placements behave differently, as the IAB Tech Lab format definitions explain, so check that the delivery type matches what you purchased.
\nExclusions are the control that fixes problems after they appear. Maintain three lists: exclude specific channels and sites that failed review, exclude content categories that clash with the brand, and exclude app inventory that cannot be verified. Keep the lists in a shared document, with the date and the reason for each entry, so a new teammate does not undo the work. Review them on a schedule, because a placement that was fine last month can drift.
\nOne restraint: do not exclude so aggressively that you starve delivery. If a campaign stops spending, loosen one list at a time and watch what happens, rather than reopening everything at once.
\nVAST stands for Video Ad Serving Template, an XML standard from the IAB Tech Lab that tells a video player what to play and what to report. You do not need to write one. You need to read one well enough to see whether something is wrong.
\nStart with the top-level wrapper, which tells you whether the tag is an InLine ad, with the creative included, or a Wrapper, which points to another tag. Wrappers can chain through several vendors, and each hop adds latency and a chance for failure, so long chains are a reason to ask questions.
\nInside, look for the Impression elements, which are the pixels fired when the ad is counted. Count them and note who owns each. Look at the Linear section, which holds the Duration, the media files, and any skip offset that sets when the skip control appears. Compare the Duration to the file you approved. Look at the tracking events: start, first quartile, midpoint, third quartile, and complete. These are the markers behind the completion numbers in your reports, and if one is missing or fires on the wrong event, the report is wrong. Finally, check the click-through URL, which should land where you intended, with your tracking parameters intact.
\nMost vendors offer a tag tester. Use it, then compare what it shows against the booking. A mismatch between the booked length, the file length, and the declared Duration is the single most common cause of a delivery or reporting argument I see. Settle it before launch, when it costs nothing.
\nA video file that fails on upload costs a day. A video file that uploads but plays with the wrong audio level or a clipped end costs a campaign. Build a short preflight and run it every time.
\nUse a probe tool to confirm the container, codec, resolution, frame rate, duration, and audio stream for each file against the placement specification. Check loudness, because a spot that is much louder or quieter than the surrounding content draws complaints or gets ignored. Confirm the last frame is clean and the first frame is not black, since the opening moments are your hook. Play each file on a phone, a laptop, and a connected television if the plan includes one. Record what you checked, who checked it, and when.
\nThen test the tag. Fire it in a tester, confirm the events listed in the previous section, and check that your measurement parameters survive the click. Only then release the budget. For a launch with several creative versions, keep a naming scheme that carries format, length, orientation, and version, so reports can be sorted without guessing which file is which.
\nSkip offset: The delay before a skip control appears in a skippable ad, set in the tag and by the platform. Completion rate: The share of started plays that reach the end of the ad, which is heavily shaped by whether the format allows skipping. Viewable impression: An impression that met a defined standard for how much of the player was on screen and for how long. Wrapper: A VAST tag that does not contain a creative itself but redirects the player to another tag, often through a vendor chain. Companion unit: A secondary display element, such as a small banner or card, that appears alongside the video while it plays.
\n1. Running one master spot across skippable, non-skippable, and bumper line items, which guarantees that at least two of them carry the wrong structure. 2. Comparing cost per view across platforms without checking each platform's view definition, so the cheapest-looking line is simply the loosest definition. 3. Treating a high completion rate on non-skippable inventory as proof of attention, when it mostly reflects that the viewer had no way out. 4. Reviewing the placement report only at campaign end, after most of the budget has already gone to inventory that should have been excluded. 5. Skipping the tag test because the vendor said it was fine, then discovering a duration mismatch or missing tracking event after launch.
\n1. Write the job of each format into the media plan before assigning budget, and cut any line item that cannot name its job. 2. Keep a one-page view definition sheet per platform so every cost comparison starts from a shared measure. 3. Build separate edits for each format and orientation at the start of production instead of cropping a single master late. 4. Review placements within the first week, log every exclusion with a date and reason, and revisit the lists on a schedule. 5. Judge video by an outcome that sits outside the platform, such as a regional holdout comparison or branded search movement, rather than by platform-reported views alone.
\n1. A skip is information. If most viewers leave at the first opportunity, rewrite the opening before you buy more reach. 2. Make every non-skippable spot work with the sound off, because the captive viewer is often not listening. 3. Use bumpers to repeat something people have already seen in a longer unit, not to introduce a brand cold. 4. Ask for the tag, not just the screenshot. A screenshot proves a picture exists; a tag proves what will be counted. 5. Keep a standing relationship with whoever runs verification on your account, since coverage gaps and unmeasurable inventory are easier to resolve before a dispute than after one.
\nTwo free tools cover most of the production-side quality work. FFmpeg (licensed under LGPL and GPL, depending on how it is built) includes a probe utility that reports the container, codec, duration, frame rate, and audio details of any file, plus converters for making each edit and orientation from a master. Use it as the core of the preflight in this chapter. Kdenlive (GPL) is a desktop video editor suited to cutting the separate skippable, non-skippable, and bumper versions and adding captions. Neither replaces a full demand-side platform, and no open source option reaches parity with one for buying, so the buying itself still happens on the platforms where the media is sold.
\nOnline video is three different formats with three different billing events, and treating them as one product is where most waste begins. Skippable video earns attention and suits consideration, non-skippable video guarantees exposure but not attention, and bumpers deliver one cheap reminder. A view means different things on different platforms, so every comparison starts with a definition sheet. Completion, viewability, and audible playback are separate questions and deserve separate answers. Placement reports and exclusion lists protect the budget, and a working knowledge of VAST lets a buyer check what a vendor will actually count. A short preflight on files and tags, built on free tools, prevents the expensive errors.
\n1. Each video format has a different billing event, and the plan should name a specific job for each before money is assigned. 2. Platform view definitions differ, so compare costs only after translating them into a shared measure. 3. Completion on non-skippable inventory is mostly a property of the format, not a sign of attention. 4. Review placements early and keep dated exclusion lists so quality problems get corrected while budget remains. 5. Read the VAST tag and test it before launch, since mismatches in duration or tracking events are a common source of reporting disputes.
\nBefore the next video flight, write one sentence naming the job of each format in the plan and one line stating each platform's view definition, then cancel any line item that has neither.
\n1. Interactive Advertising Bureau Tech Lab. Video Ad Serving Template (VAST) specification, current version. 2. Interactive Advertising Bureau Tech Lab. OpenRTB specification, current version, video object and skip-related attributes. 3. Interactive Advertising Bureau. Digital Video Ad Format Guidelines, covering in-stream, accompanying content, interstitial, and standalone definitions. 4. Media Rating Council. Viewable Impression Measurement Guidelines, video sections. 5. FFmpeg project documentation, ffprobe and ffmpeg command references.
"},{"id":"chapter-22-connected-tv-and-streaming-inventory","title":"Chapter 22: Connected TV and Streaming Inventory","html":"By the end of this chapter, the reader will be able to: 1. Describe the three layers of connected TV inventory (device, app, and channel) and the ad-supported tiers that sit inside them. 2. Distinguish the measurements a connected TV buy can verify from the ones it can only estimate. 3. Evaluate a placement for made-for-advertising risk using observable signals rather than seller assurances. 4. Apply a structured set of supply path questions to any seller, reseller, or platform before committing budget. 5. Design a connected TV test with a hypothesis, a holdout, a frequency rule, and a stop condition. 6. Analyze exported delivery logs with Python and DuckDB to find duplicated reach and unexplained placements.
\nSuppose you have 40,000 dollars for a four-week connected TV flight. The seller quotes a 20 dollar CPM (cost per thousand impressions). Do the arithmetic. That buys two million impressions.
\nThe slide deck says \"reach two million households.\" Now suppose the flight runs through three buying paths, because your agency, your platform rep, and a reseller each had a good idea. Each path enforces a frequency cap of four per week, which sounds disciplined. But the three caps do not talk to each other. A household in the overlap can legitimately receive twelve impressions a week and every dashboard will report that nobody exceeded four.
\nOver four weeks, an average household in that overlap might see the ad 40 times or more. Say the average frequency across the whole flight turns out to be 10. Then two million impressions reached 200,000 households, not two million. The reach claim was off by a factor of ten, and nobody lied.
\nThat is the assumption I want to challenge in this chapter: that a big screen, a premium feel, and a clean-looking report mean you know what you bought. Connected TV is television delivered over an internet connection, and it inherited the prestige of television along with the opacity of programmatic buying.
\nI am not arguing against the channel. I am arguing that the buyer's job here is closer to a scout's than a shopper's: you verify what the player did, not what the program says he did. For the engineering detail of how streaming ad delivery works, the VAST and SupplyChain specifications in the references are the place to start.
\nConnected TV (CTV) inventory is any video ad shown on a television screen that reaches the internet, whether through a smart TV, a streaming stick, a set-top box, or a game console. Simple definition. Messy reality, because three kinds of seller can hold the same impression.
\nThe first is the app. A streaming service that offers a cheaper tier supported by advertising sells ads inside that tier. You buy known content in a known environment, usually with the strongest brand safety.
\nThe second is the free ad-supported streaming channel, called a FAST channel. Picture a continuous lineup of older shows or niche programming with commercial breaks, bundled with dozens of others inside one app. It resembles cable. The economics differ, because many of these channels license content cheaply and depend entirely on ad revenue.
\nThe third is the device or operating system layer. The maker of the television or streaming stick often runs its own home screen and its own ad sales, and it sees what households launch. That visibility is valuable. It also means audience claims sometimes arrive that the buyer cannot inspect.
\nThe word \"CTV\" can describe a premium series or a looping channel nobody chose on purpose. A buyer who treats them as one pool pays the average price. When a seller says \"CTV inventory,\" ask which layer, and ask for the apps and channels by name. If the seller will not provide the list before the buy, that is an answer.
\nDelivery is the clearest. A video ad served to a television produces a technical record: requested, started, reached the quarter marks, finished. Completion looks high because the viewer is on a couch, the ad is often unskippable, and the remote is across the room.
\nGeography is dependable too. Devices connect through household internet, so location at the postal code or metro level is usually reasonable. A regional retailer buying 12 counties can enforce that boundary better than a broadcast signal that spills across market lines.
\nHousehold-level exposure is the third strength. Because the unit is often the household, you can compare exposed households with unexposed ones using a holdout (a group deliberately kept from seeing the ads). With clean matching, a well-designed holdout gives a lift estimate that television buyers once could only dream about.
\nFrequency inside a single platform is also controllable, though only inside that platform's walls.
\nSo the strong zone is what happened on the screen inside one controlled environment. Build your report on it and be modest about everything else.
\nThe weak zone is wider than most sellers admit.
\nCross-platform reach is the first gap. No single seller sees a household's viewing across every app and device. Each reports its own reach, and adding the reports overcounts because the same household sits in several. Overlap estimates are modeled opinions. Treat combined reach as a ceiling.
\nIndividual attention is the second. A television may be on while the viewer is in the kitchen. A completed ad did play. Whether anyone watched is a separate question the platform cannot answer.
\nAttribution is the third, and it costs the most money. A \"view-through conversion\" says a household saw an ad and later did something you value. The household may have done it anyway. Heavy users of your brand are heavy users of everything, so they appear in exposed groups at higher rates. The number fits a story. It does not prove cause. The holdout, covered below, gets closer.
\nIdentity is the fourth. Matching a television to a household, and a household to a customer file, relies on partners and methods you often cannot inspect. A low match rate quietly shrinks the audience you thought you targeted. Ask for the match rate in writing.
\nIn short: what happened on the glass is measurable, what happened in the room is not, and what happened afterward is estimated. Decide on the first. Label the rest directional when you present it.
\nEvery channel attracts people who harvest ad revenue without serving an audience. On the open web they were made-for-advertising sites, pages built to carry ad slots and fed with cheap traffic. Connected TV has its own version.
\nThe pattern: an app or channel assembled from low-cost video, sometimes looped, with commercial breaks arriving at a punishing rate. In worse cases, requests that look like a television in a living room are produced by software pretending to be one. The log shows a completed video. Nobody was there.
\nI have no reliable public number for how much of any marketplace is affected, and I will not invent one. The risk is real enough that serious buyers keep it as a standing review item. Start with signs you can see without anyone's permission:
\nNone of these proves fraud alone. An odd placement can be a small, legitimate channel. The point is to turn a feeling into a list and ask the seller about every item. A seller with nothing to hide answers quickly and specifically. A seller with something to hide says \"proprietary.\"
\nFor a family-first brand the stakes rise. A children's product ad in an adult-oriented channel is a brand problem no completion rate can repair. Set a conservative default: approve placements by name, not by category, until early logs justify relaxing.
\nWhat should a CTV CPM cost? There is no public benchmark, and anyone who quotes one as settled fact is selling something.
\nPrice varies with the layer, the targeting, the buying method, live content, the season, and how many intermediaries sit between you and the screen. Two buyers can run the same creative against the same audience on the same day and pay different prices.
\nSo judge price against two things you control. The first is your own history. Log every CTV buy: layer, targeting, method, price paid, delivery quality. After a few flights you hold a private benchmark worth more than any published one, because it reflects your category and your negotiating position. The second is the supply path. A high price can be fair if most of it reaches the content owner. A low price can be poor if middlemen consumed most of it and the remainder bought junk.
\nA worked example, illustrative. Seller A quotes 14 dollars per thousand and Seller B quotes 24 dollars. You assume Seller A saves you ten dollars per thousand. Now suppose a log review shows one third of A's impressions landed in placements you reject. Your effective cost for acceptable impressions is 14 divided by two thirds, or 21 dollars. The gap shrinks from ten dollars to three. If B also offers named placements and a data match that supports a holdout, the pricier quote may be the cheaper buy. Do the second calculation every time.
\nIf the seller will not describe the fee structure or the included placements, you are not negotiating. You are accepting an invoice.
\nThe supply path is the chain of companies between your ad and the screen: your platform, resellers, the exchange, the publisher, the app. Each link takes a share and offers a chance for substitution, where the impression you bought is swapped for a cheaper one. You need a list and the discipline to use it every time.
\nStart with ownership. Does the seller own this inventory or resell it, and from whom? Ask for the names of the app or channel owners, not just the seller's brand.
\nThen ask about the paperwork the industry already publishes. Publishers can post a public file listing which companies may sell their ad space, and a companion record states whether a seller is the owner or an intermediary. The supply chain information can also travel with the bid so the chain is visible. Ask whether the inventory is covered. Your counsel or technical lead can confirm how this applies to your buy.
\nNext, fees. How many parties take a cut, as a percentage or a fixed amount per thousand? Then placements. Can you approve a list before the buy, block a list after, and receive impression-level logs? Finally, incidents. If a placement proves invalid, is there a credit, who decides, and how fast? Get it in writing. This is operational guidance, not legal advice, and counsel should review any remedy language.
\nThe meeting version:
\n1. Who owns the inventory, and who resells it? 2. What do the public authorization files say about this seller? 3. How many parties take a fee, and how much reaches the publisher? 4. Will you give me placement-level logs, by name? 5. What is the remedy, in writing, when a placement is invalid?
\nFive direct answers, proceed. Three or four, proceed smaller. Two or fewer, move on.
\nMost CTV tests fail before they start, because nobody wrote down what the test should prove. A good test is a small, boring document with five parts.
\nFirst, a hypothesis you can fail. \"CTV drives awareness\" cannot fail. \"Exposed households will purchase at a rate at least five percent higher than matched unexposed households within 30 days\" can. Pick one primary outcome, not five.
\nSecond, the holdout. For a regional brand, the cleanest option is geographic: choose comparable markets, advertise in some, withhold in others, and compare. Pick markets that behaved alike before the test, because that history makes the comparison fair. Decide before launch and protect it. Someone will want to add the held-out markets late. Do not let them.
\nThird, frequency across paths. If you buy through more than one path, give each path a separate slice of the audience, or set conservative caps so the total stays under your target. For a first test I like a combined ceiling of about four a week, enforced by dividing the audience rather than trusting a single cap.
\nFourth, named placements, because a first test is no time to explore the long tail.
\nFifth, a stop condition written before you spend a dollar. If impressions in rejected placements pass a threshold, pause. If the seller misses the log delivery date, pause.
\nAn illustrative case: a fictional furniture brand, Cedar Row Home, split 12 markets into six test and six hold, approved 20 named apps and channels, and set a combined weekly cap of four. Midway, the log review showed one reseller delivering heavily into unapproved placements. The stop condition existed, so they paused that path within days and finished cleanly. The lift estimate was modest and honest, and it produced a decision.
\nReports are summaries, and summaries hide things. The impression-level log is where the truth sits. If the seller will not provide one, that is a finding.
\nLoad the files into DuckDB, a fast database that runs on your own computer, and query them with plain SQL. How many impressions ran in each placement? Which show near-perfect completion at every hour? How many distinct households is that, compared with the reach the seller reported?
\nThe last question rescues you from the introduction's trap. Divide impressions by distinct identifiers and you have your real frequency. Household identifiers are personal data in many settings, so keep logs in a restricted place, delete them on a schedule, and have counsel confirm your handling.
\nFAST channel: A free ad-supported streaming television channel, usually a continuous lineup of programming with commercial breaks, delivered through an app or a device home screen. Made-for-advertising: Inventory built primarily to carry ads and collect ad revenue, with little genuine audience or content value behind it. Supply path: The chain of companies between the buyer and the screen, including platforms, resellers, exchanges, and the publisher or app owner. Holdout: A group of households or markets deliberately kept from seeing the ads, used as a comparison to estimate what the ads actually caused. View-through conversion: A conversion credited to an ad because the household saw it and later acted, without proof that the ad caused the action.
\n1. Treating \"CTV\" as a single product and accepting a price without asking whether it covers an app, a channel, or a device placement. 2. Setting a frequency cap in each buying path and assuming the caps add up to control, when they only limit each path in isolation. 3. Reading a view-through conversion report as proof of cause, which credits the ad for sales that heavy brand users would have made anyway. 4. Comparing sellers by quoted CPM alone, without adjusting for the share of impressions that land in placements you would reject. 5. Launching a test with no holdout, no stop condition, and no outcome defined in advance, then arguing about what the results mean afterward.
\n1. Approve placements by name before launch, and require impression-level logs so that you can compare what ran against what you approved. 2. Keep a private price log of every CTV buy, recording layer, targeting, method, price, and delivery quality, so your own history becomes your benchmark. 3. Ask the five supply path questions at every negotiation and record the answers in the insertion order or contract, with counsel reviewing any remedy language. 4. Divide the audience among buying paths, or set a combined weekly ceiling, so that total frequency stays visible and bounded. 5. Run a holdout for any claim about sales or leads, and write the stop condition before the first impression is served.
\n1. Ask for the list of apps and channels before pricing. 2. Review logs at the end of week one, not the end of the flight. A problem found in seven days costs a quarter of what the same problem costs in 28. 3. Build the creative for a living room: clear audio, readable text from across the room, and a message that works if the viewer glances up late. 4. For children's products and sensitive categories, use a named allow list and keep it conservative until the logs justify expanding it. 5. When a seller offers a \"guarantee,\" ask what is guaranteed: impressions delivered, impressions verified, or placements approved. They are three different promises.
\nPython and DuckDB (MIT) are enough to audit a connected TV flight on a laptop. Ask the seller for impression-level logs, load the files into DuckDB, and use SQL to count impressions by app or channel, spot placements with implausibly perfect completion, and divide impressions by distinct household identifiers to find true average frequency. Full-scale demand-side platforms have no open source option that reaches parity, so the buying itself will still happen on a seller's system. The review of what that system delivered does not have to.
\nConnected TV inventory arrives in three layers (app, channel, and device), and the same label covers premium content and placements nobody chose. Measurement is strong for what happened on the screen inside one platform, including delivery, completion, location, and single-platform frequency. It is weak for cross-platform reach, individual attention, attribution, and identity matching. Made-for-advertising risk is real but cannot be sized with a public figure, so buyers screen placements by name and review logs early. No public price benchmark exists, so your own history and your supply path knowledge decide whether a quote is fair. A short list of supply path questions, a test with a holdout, a frequency rule, and a stop condition, and a routine log review with Python and DuckDB give a brand manager a defensible way to buy.
\n1. A CTV quote can mean an app, a channel, or a device placement, so ask which one before discussing price. 2. Frequency caps apply inside each path only, so total household frequency across paths must be controlled by you. 3. Trust delivery, completion, and location more than reach, attention, and view-through credit. 4. Compare sellers on price per verified impression in approved placements, since no public benchmark exists. 5. A test needs a failable hypothesis, a holdout, a frequency rule, named placements, and a stop condition written before launch.
\nPut impression-level logs and a named placement list into the CTV contract, then divide impressions by distinct households to see the reach you actually bought.
\n1. IAB Tech Lab. ads.txt and app-ads.txt specifications. Public files that declare which companies are authorized to sell a publisher's inventory. 2. IAB Tech Lab. sellers.json and the SupplyChain object specifications. Records and bid-level fields that disclose the chain of sellers behind an impression. 3. IAB Tech Lab. Video Ad Serving Template (VAST) specification. Standard for video ad delivery and the playback events that delivery logs record. 4. Media Rating Council. Outcomes and Data Quality Standard and Video Viewability guidance. Accreditation framework for measurement claims. 5. DuckDB Documentation. Reading CSV and Parquet files with SQL. Project documentation for local log analysis.
"},{"id":"chapter-23-programmatic-guaranteed-private-deals-and-open-exchange-for-video","title":"Chapter 23: Programmatic Guaranteed, Private Deals, and Open Exchange for Video","html":"By the end of this chapter, the reader will be able to: 1. Define programmatic guaranteed, private marketplace, and open exchange buying and state what the buyer actually receives in each. 2. Explain how a deal ID works and identify the points where a deal can fail without any visible error. 3. Calculate the effective cost of a video impression after floors, fees, and curation layers are applied. 4. Audit a seller's supply path by reading ads.txt and sellers.json files. 5. Diagnose a deal that is under-delivering by working through a fixed order of checks. 6. Evaluate a proposed deal against a written scorecard and decide whether it earns budget.
\nHere is the assumption I hear most often: a private deal is a safe deal. The word \"private\" does the work. It suggests a velvet rope, a vetted guest list, a premium seat. Buyers hear it and relax.
\nThen the campaign runs. Delivery is a third of what was promised. Or it delivers fully, and the placements turn out to be the same resold inventory that was sitting in the open auction last week, now wearing a deal ID and a higher price. Nobody lied. The label described the packaging, not the contents.
\nI have spent enough time around media budgets to know that this is where money quietly leaves. A deal ID is a pointer. It tells a system which set of terms to apply to a bid. It does not tell you who owns the inventory, how many hands touched it on the way to you, who took a cut at each hand, or whether the video player at the end of the chain is even visible on the screen.
\nThis chapter does not argue against private deals. They are often the right tool for video. It argues that every deal, guaranteed, private, or open, deserves the inspection you would give a used truck before handing over cash: look under it, ask who owned it before, and walk away if the answers are vague. Operational guidance only. Your contracts and counsel decide the legal terms.
\nStart with the three structures and be precise about the promise each one makes.
\nProgrammatic guaranteed is a fixed commitment. The seller reserves a defined volume of impressions at an agreed price for a defined window, and you commit to buy it. It behaves like a traditional insertion order. You get certainty of volume and price. You give up flexibility, and you carry the risk that the placement does not perform, because the commitment does not care.
\nA private marketplace deal, often shortened to PMP, is an invitation to bid, not a promise to deliver. The seller offers access to a defined pool of inventory, sets a floor price (the minimum bid they will accept), and lets approved buyers compete. Nothing is guaranteed. If your bid is below the floor, or another buyer bids higher, or the seller prioritizes a different source, you win nothing.
\nOpen exchange is the public auction. Any approved seller can list impressions and any approved buyer can bid. You get no promised volume or placement, but you get reach, speed, and the ability to apply your own filters.
\nOne way to hold the difference: guaranteed sells volume, private sells access, and open sells a lottery ticket with good odds if you do the filtering. A video impression is only worth something if the player loads, the ad plays with sound where sound matters, and a person can actually see it. None of the three deal types promise that by themselves. Those are separate conditions you negotiate and verify.
\nA deal ID is a short code your buying platform sends in the bid request so the seller's system knows which terms apply. The seller creates it. You enter it in your platform. Both sides must agree on the price, the floor, the dates, and the targeting attached to it.
\nThe trouble is that the code lives in two systems that do not talk to each other in real time. That gap is where most failures hide.
\nThe common ones are mundane. The deal was never activated on the seller side. The ID has a transposed character. The deal is tied to one buyer seat and your platform is using another. The seller quietly raised the floor above your bid. The deal expired on a date nobody wrote down. Or your creative misses the seller's video specifications, so their ad server rejects it at the last step. Each produces the same symptom: nothing happens, with no readable error.
\nThere is a second class of problem that is worse because delivery looks fine. Some deals attach a code to inventory that is also available elsewhere at a lower price. You pay the deal premium for impressions you could have won in the open auction. Ask the seller directly: is the inventory under this deal exclusive to the deal, or is it also offered in the open auction? A seller who answers cleanly is a seller worth working with.
\nKeep a deal register, a spreadsheet with one row per deal: the ID, the seller, the agreed floor, the dates, the targeting terms, the creative specs, the named contact on both sides, and the date you last confirmed it was active. It feels clerical until a campaign stalls and you can answer every question in 10 minutes instead of three days.
\nThe price you bid is not the price the publisher receives. Between your bid and the publisher, several parties may take a share, and in video the chain is often longer than in display.
\nHere is an illustrative worked example. Suppose a brand manager for a fictional regional retailer, Harborline Outdoor Supply, agrees to a private deal at a 24 dollar CPM (cost per thousand impressions). That sounds like the price of premium video. Now trace where the money goes. Suppose the buying platform charges a share of media spend, a data provider attaches an audience fee, a curator layered on top charges a fee for packaging the inventory, and a verification vendor charges per thousand impressions. If those four layers together consume eight dollars of the 24, the publisher is receiving 16 dollars for the impression and the buyer is paying a 50 percent premium over what reached the content owner. These numbers are made up for the example. The point is the shape: the headline price hides a stack.
\nA curation layer is a third party that assembles a package of inventory, often by audience or theme, and sells it as a ready-made deal. Curation can add real value when the curator has knowledge you lack, such as which sports streaming placements hold up on a weekend. It adds cost without value when the curator only repackages inventory you could reach directly.
\nFloors deserve their own attention. A floor is the lowest price the seller will accept. Sellers sometimes raise floors when demand is high and rarely announce it.
\nBefore accepting any deal, ask which fees are inside the stated price, whether the floor is fixed for the deal period, and what the curator does that you could not do yourself. Write the answers into the deal register. If the seller will not answer in writing, treat that as an answer.
\nYou do not have to take a seller's word for who they are. Two public files, both defined by the IAB Tech Lab, let you check.
\nThe ads.txt file sits on a publisher's website. It lists the companies the publisher has authorized to sell its inventory, along with an account ID at each company and whether the relationship is direct or through a reseller. The sellers.json file sits on the domain of an ad exchange or supply platform. It lists the sellers that platform represents, and tells you whether each is the publisher itself or an intermediary. Many intermediaries are legitimate. Some are not.
\nA third piece, the SupplyChain object, travels inside the bid request itself. It records each company that handled the impression on the way to you. If the publisher's ads.txt does not list the platform that sent you the bid, or the sellers.json entry for that account is hidden or does not match, you have found a gap worth questioning.
\nHere is a practical audit sequence: 1. Ask the seller for the publisher domains or app bundle names attached to the deal. 2. Pull each publisher's ads.txt and confirm the seller's platform and account ID appear, marked as direct where the seller claims to be direct. 3. Pull the platform's sellers.json and confirm the account ID resolves to a named entity that matches the publisher or a disclosed intermediary. 4. Count the intermediaries in the chain. More hands means more fees and more places for error. 5. Record anything that fails and send it back to the seller as a written question.
\nDo not read a clean file as proof of a good placement. It proves only that the sale was authorized. Whether the ad played to a real viewer is a verification question for a different chapter.
\nA deal is under-delivering when it spends far less than the plan expected. Work through the causes in a fixed order, from cheapest to check to most expensive. Random poking wastes days.
\nFirst, confirm the deal is live on both sides. Ask the seller to confirm in writing that the deal is active, the dates are correct, and your seat is approved.
\nSecond, check the ID and the seat. Compare the code character by character. Confirm your platform is bidding from the seat the seller approved.
\nThird, check the bid against the floor. If your bid is at or barely above the floor, you are exposed to any small increase. Raise the bid in modest steps and watch whether volume appears. If it does, the floor was the constraint.
\nFourth, check the creative. Video creative can fail on length, file format, bitrate, or missing tracking tags. Ask the seller for the exact rejection reason code if one exists. Many sellers can see it. Few volunteer it.
\nFifth, check your targeting. Layering audience data on top of a deal that already has narrow inventory can shrink the pool to almost nothing. Remove one layer at a time and watch whether bid requests rise.
\nSixth, check frequency caps. A cap set too low for a small pool will choke delivery.
\nSeventh, ask the seller about available volume. Some pools are simply small. If the pool cannot hold your plan, the fix is a bigger pool or a smaller plan, not a higher bid.
\nLog each step in the deal register. A record of what you tried turns the seller conversation from complaint into diagnosis.
\nOpen exchange tends to win in four situations. First, when you need broad reach quickly and have no relationship with specific sellers. Second, when you can build strong filters yourself: allow lists of verified domains, exclusion lists, supply path preferences, and verification on every impression. If your own controls are better than the seller's packaging, you are paying for something you already own. Third, when the private deal has a thin pool. A deal that cannot spend is worth nothing regardless of how premium it sounds. Fourth, when you are testing audiences and formats before committing guaranteed budget.
\nPrivate deals tend to win when the seller offers something the open auction cannot: a specific live event, exclusive content, a confirmed placement against a particular show, audience data the seller owns, or a genuinely cleaner supply path with one hop between you and the publisher.
\nFor buyers in a family-first category, children's products or anything near sensitive topics, set the conservative default: start with private or guaranteed inventory and a short allow list, and expand to open exchange only after verification results are clean.
\nOne honest limit: no open source tool replaces a full-scale demand-side platform, the system that actually places bids across exchanges at volume. The tools in this chapter help you audit and verify. The bidding itself still happens in whatever buying platform you have contracted.
\nA scorecard turns gut feeling into a repeatable decision. Score each criterion from zero to five, then apply a minimum threshold. This is an illustrative structure. Adjust the weights to your own campaign.
\nConsider a composite case. A fictional trail-running footwear brand, Quarry Fox Trail Co., receives three proposals for a six week video flight: a guaranteed deal on a streaming sports channel, a private deal curated by an intermediary, and an open exchange plan with its own allow list.
\n| Criterion | What a high score looks like | | --- | --- | | Supply path clarity | Seller is direct, ads.txt and sellers.json both match | | Fee transparency | All fees disclosed in writing, curator role explained | | Floor stability | Floor fixed for the deal period | | Delivery confidence | Forecast covers at least 130 percent of planned volume | | Exclusivity | Inventory is not offered cheaper in the open auction | | Creative fit | Specs and rejection reasons are shared up front | | Exit terms | Pilot period or cancellation without penalty |
\nSuppose the guaranteed deal scores strongly on delivery but a zero on exit terms. The private deal scores a two on supply path clarity because the chain has four intermediaries, and a one on fee transparency. The open exchange plan scores well on exit terms and floor stability. Quarry Fox's decision in this illustration: take the guaranteed deal only after negotiating a short pilot, decline the private deal until the curator discloses fees and the chain shrinks, and run a small open exchange test with strict filters. The scorecard made the choice defensible and gave the team a written reason to ask the seller for changes.
\nDeal ID: A code passed in a bid request that tells the seller's system which agreed terms, such as price and inventory pool, apply to the bid.
\nFloor price: The lowest bid a seller will accept for an impression, set by the publisher or the platform and sometimes changed without notice.
\nCuration: The packaging of inventory by a third party into a ready-made deal, often by audience or theme, usually for an added fee.
\nsellers.json: A public file on an ad platform's domain listing the sellers it represents and whether each is a publisher or an intermediary.
\nSupplyChain object: A record inside the bid request listing each company that handled the impression between the publisher and the buyer.
\n1. Assuming a deal ID guarantees premium inventory. The code only applies terms, so confirm what sits behind it. 2. Accepting a quoted CPM without a fee breakdown. Layered charges can consume a large share before the publisher sees any of it. 3. Launching a deal and not checking spend for several days. A deal that never activated, or one with an expired date, looks the same as a quiet start. 4. Treating a clean ads.txt match as proof of quality. It shows the sale was authorized, not that the ad was seen by a real person. 5. Writing off open exchange without testing it.
\n1. Keep a deal register with one row per deal, including floors, fees, dates, specs, and named contacts on both sides. 2. Ask in writing whether the inventory in each deal is exclusive to the deal or also sold in the open auction. 3. Audit every publisher domain and app bundle in a deal against ads.txt and sellers.json before the first dollar is spent. 4. Negotiate a pilot period or exit clause into guaranteed commitments so a poor fit costs a small test, not a full flight. 5. Score every proposal on the same written criteria so decisions can be explained and compared across campaigns.
\n1. When a deal stalls, ask the seller for the exact rejection or no-bid reason code. Most sellers can see it and will share it if asked directly. 2. Request a weekly forecast of matching impressions before you set targeting. A small pool cannot absorb a large plan no matter how high you bid. 3. Count the hops in the supply chain. One or two is healthy. Four or more deserves a direct question about why. 4. Review floors weekly on any private deal. A quiet floor increase is one of the most common reasons a healthy deal slows down. 5. Treat a seller who refuses to put fee terms in writing the way you would treat a contractor who refuses to put a price in writing.
\nPython and Scrapy (BSD license) let you write a small script that fetches ads.txt files from a list of publisher domains and sellers.json files from a list of platforms, then compares the entries against the seller IDs in your deal register. Run it before a deal launches and again monthly. The Prebid documentation (Apache 2.0 license) explains how supply chain information and seller declarations are passed through bid requests, which helps you read the SupplyChain object and ask sharper questions of your sellers. Neither replaces a buying platform. They give you evidence to challenge a supply path before money moves.
\nProgrammatic guaranteed sells volume and price certainty, private marketplace sells access to a pool of inventory with no promise of delivery, and open exchange sells broad reach that depends on your own filters. A deal ID is only a pointer and can fail silently when two systems disagree. The listed price often hides a stack of fees, so ask for the breakdown in writing. The ads.txt and sellers.json files, read alongside the SupplyChain object, let you check that a seller is authorized and see how many hands touched the impression. When a deal stalls, work through the checks in order, from confirmation of activation to the size of the pool. Open exchange can beat a private deal when your controls are strong, and a written scorecard keeps every decision consistent and defensible.
\n1. A private label describes packaging, not contents, so verify the inventory, the exclusivity, and the supply path behind every deal. 2. Programmatic guaranteed commits you to volume and price, while private marketplace and open exchange commit the seller to nothing. 3. The quoted CPM is a stack of fees, and the publisher may receive only a portion of it. 4. ads.txt, sellers.json, and the SupplyChain object are public evidence you can check before spending. 5. Under-delivery has a fixed order of causes, and a deal register plus a written scorecard make both troubleshooting and buying decisions faster.
\nIf a video deal is about to go live, pull the ads.txt and sellers.json entries for every publisher in it, and get the fee stack and floor terms in writing. If the seller will not provide either, move that budget to a deal that will.
\n1. Interactive Advertising Bureau Tech Lab. ads.txt specification, current version. Authorized digital sellers file. 2. Interactive Advertising Bureau Tech Lab. app-ads.txt specification, current version. Authorized sellers for apps. 3. Interactive Advertising Bureau Tech Lab. sellers.json specification, current version. Seller identity disclosure. 4. Interactive Advertising Bureau Tech Lab. OpenRTB specification, SupplyChain object extension. Supply chain transparency in bid requests. 5. Prebid. Prebid documentation. Supply chain object support and seller declarations.
"},{"id":"chapter-24-streaming-audio-and-podcast-programmatic","title":"Chapter 24: Streaming Audio and Podcast Programmatic","html":"By the end of this chapter, the reader will be able to: 1. Identify the three main audio ad formats: streaming spots, host-read, and dynamically inserted. 2. Explain why audio earns reach credit but struggles to earn click credit. 3. Apply codes, vanity links, and lift studies to measure an audio campaign. 4. Write a 30 second script that works for the ear rather than the eye. 5. Evaluate a frequency plan across ad pods, shows, and devices. 6. Design a host-read quality check using a free speech to text tool.
\nThe first audio campaign I ever judged, I judged by the wrong scoreboard. The reports came back with a big impression count and almost no clicks, and I called it a miss. Then the phones at the client's shop got busier, branded search crept up, and a few customers walked in saying they had \"heard something on the radio,\" except there was no radio buy. I had measured a channel that has no click with a tool built for clicks. That mistake cost the client about a month of misplaced doubt, and it taught me to ask a different question before I spend a dollar on sound.
\nHere is the assumption I want to challenge: that audio is a cheaper, quieter version of video, so it should be bought and judged the same way. It is not. Nobody looks at an audio ad. Nobody scrolls past it or taps it. The listener is driving, folding laundry, or lifting weights, and your message arrives inside a voice the listener already chose to trust. That is a very different place to stand.
\nIt also means the usual comfort metrics fail you. Click through rate, the percentage of people who tap an ad, is close to meaningless when there is nothing to tap. If your reporting culture demands a click for every dollar, audio will always lose the argument, and the reason will be the ruler, not the channel. As Chapter 1 argued, a number being trackable is not the same as it being measured.
\nThis chapter gives you a working method instead. You will see the formats and what each one is good for, how programmatic buying fits into audio, how to count results without clicks, how to write copy people can hear, and how to keep frequency from turning a good listener into an annoyed one. I will also show a simple way to quality check host reads with free software, because the biggest risk in podcast buying is not the price. It is paying for a read that never said what you paid it to say.
\nMost audio buys come down to three formats, and each one carries a different kind of trust.
\nA streaming audio spot is a recorded ad, usually 15 or 30 seconds, played in a break between songs or shows on a music or radio style service. The voice belongs to your brand or an actor you hired. The listener knows it is an ad. Its strength is scale and speed, because you can reach a large audience in a defined place and time without waiting on any one creator.
\nA host-read ad is spoken by the person who makes the show, in their own words, often working from a short brief and a few talking points. The listener hears a recommendation instead of a commercial. In many accounts this format earns stronger recall and action than a produced spot, though I would not promise you a multiple, because it depends on the show and the offer. Its weakness is control. You trade precision for believability.
\nA dynamically inserted ad is not a creative style at all. It is a delivery method. The episode file is stored without ads, and the ad is stitched in at the moment of download or play, which lets you swap messages, target by location or device, and retire a campaign without touching the episode. Dynamic insertion can carry a produced spot or a host-read recording. It is the plumbing that makes programmatic podcast buying possible, and it is also why old episodes can keep serving fresh ads for years.
\nPosition matters inside each format. Ads placed before content tend to ride on attention that is still arriving. Ads in the middle benefit from a listener who is settled in. Ads at the end catch whoever stayed, which is cheaper for a reason. Suppose a local home services brand wants to promote a spring tune up. A produced 30 second spot in a commute window brings reach. A host-read mid-roll on a home improvement show brings trust. The same offer, two jobs.
\nProgrammatic buying means software places your ad through an automated auction or a direct deal instead of a phone call and an insertion order. In audio, the software sits between you and many publishers, and each break of ads is described to buyers using the same bid request language that display uses, with fields adapted for audio such as duration and the position of the ad inside the break.
\nTwo realities shape this channel. Audio inventory is finite, because a listener can only hear so many minutes in a day. And audio is linear. A person cannot skim past a spoken ad the way an eye skips a banner, so each impression carries more weight and each bad placement hurts more.
\nI would divide the plan into two lanes. The first lane is programmatic streaming audio, where you set an audience, a geography, a daypart (a time window such as the morning commute), and a frequency cap, and the system delivers. This lane is useful for steady reach and for testing. The second lane is host-read podcast sponsorship, which is still mostly arranged person to person or through a network, even when a platform offers some programmatic access to the same shows.
\nOne honest caution: a full scale demand-side platform, the software a buyer uses to bid across many exchanges, has no open source equivalent that reaches parity today. You will rent that access through a partner or a managed buy. Be clear about who holds the audience data and the reporting, because that decision decides how much you can verify later. The open tools in this chapter handle your creative and your quality checks, not the auction.
\nAudio asks you to count three kinds of evidence. None is perfect. Together they are good enough to decide.
\nThe first is direct response signals that need no click. A promo code spoken in the ad (\"say SPRING at checkout\") can be counted at the register or the cart. A vanity link is a short, easy to say web address such as yourbrand.com/spring that redirects to a tracked page. People remember a short word far better than a long address, so keep it simple, and never read out a string of random characters. Both tools undercount, because many listeners will simply search your name later instead of typing what they heard. Treat the totals as a floor.
\nThe second is lift. A brand lift study asks a group of people who heard the ad and a matched group who did not the same questions about awareness or intent, and the gap is the lift. A geographic holdout does the same thing with places. Suppose you run audio in 12 test markets and withhold it from 12 similar control markets for six weeks. If branded search and store visits rise faster in the test group, and the markets were truly alike before the campaign, the difference is a reasonable estimate of what audio added. Pick markets by matching their past trend, not their size alone.
\nThe third is the industry standard for counting podcast delivery. The IAB Tech Lab publishes podcast measurement technical guidelines, and IAB Tech Lab runs a compliance program built on them. At name level, the idea is simple: downloads and ad plays should be counted by agreed rules, filtering out robots and repeat requests, so that two publishers reporting \"one thousand\" mean the same thing. Ask any seller whether their numbers follow these guidelines, and whether an independent party has checked them. A vague answer is information.
\nCopy that reads well on a page often sounds terrible out loud. The ear has no scroll bar. It cannot reread a confusing sentence, so every line must land the first time.
\nStart with one idea. A 30 second spot has room for roughly 75 spoken words, and a good one uses fewer. Say the brand name early, and again near the end, because the listener may have been distracted for the first half. Say the action once, and make it a single action. \"Search for us\" or \"say SPRING at checkout\" will work. A list of three actions will produce none.
\nWrite the way people talk. Contractions, short sentences, and plain words. Read the draft aloud at normal speed with a stopwatch before you record anything. If you stumble, the listener will too. Avoid tongue twisters, similar sounding words side by side, and numbers that are hard to hear. A phone number is a poor ask in audio. A short word is a good one.
\nSilence is a tool, not a hole. A half second of space before the offer helps it stand out. Music under the voice should support it, never fight it. And test the spot with the screen off. If a friend cannot repeat your brand name and your offer after one listen in a car, the script is not finished.
\nFor host-read ads, the brief should be short. Give the host the offer, two or three true points, the required disclosures, and the one action. Then let them speak in their own voice. Over-scripting a host drains the very trust you are paying for.
\nFrequency means how many times one listener hears your message. In audio it behaves differently than on a screen, for three reasons.
\nThe first is the pod. A pod is the group of ads in one break. If three ads play back to back, your spot competes with the others for the same few seconds, and the third position usually gets the weakest attention. Ask for position control where you can, and avoid breaks stacked with competing brands from the same category.
\nThe second is repetition by habit. Listeners often return to the same few shows and playlists. A cap that looks gentle across the whole network can still hammer a loyal listener inside one show. Set caps by listener and by show, and watch the numbers by daypart.
\nThe third is the identity problem. One person may hear you on a phone, a smart speaker, and a car dashboard, and many systems will count three people. Your reported frequency is therefore a guess, usually a generous one. Plan a little lower than feels necessary, and treat a flat, steady schedule as better than a burst that wears people out.
\nA practical rule: aim for enough repetition that the brand name sticks, then stop. Three to five exposures over a few weeks is a sensible starting test for many local offers. Then use your lift data to move that number up or down, instead of defending a rule of thumb.
\nHere is a scenario I have seen more than once. You pay for a host-read on a show, the episode goes live, and no one on your team ever listens to it. The host dropped the vanity link, mispronounced the brand, or left out the legally required line. By the time anyone notices, the campaign is over.
\nWhisper, a free speech to text model, fixes the listening problem at almost no cost. You download the episode audio, cut out the ad segment, and run it through Whisper to produce a transcript. Then you check the transcript against a short list: Was the brand name said correctly? Was the code or link said, and said once clearly? Was the required sponsorship disclosure included? Did the read run at the agreed length? Did the host add a claim you never approved?
\nTreat the transcript as a screening tool, not a final judge. Speech to text can mishear names and unusual words, so a flagged line deserves a human ear. Used that way, it turns a vague hope into a checklist. Keep the transcripts with your campaign records, because they are your proof if a seller disputes whether an ad ran as ordered.
\nI also use the transcripts for learning. Over a few campaigns, you will see which hosts say the code clearly and which bury it, and that pattern guides the next buy better than any dashboard.
\nAudio often plays in cars and kitchens, and children are frequently in the room. Set a conservative default. Avoid placing ads for sensitive categories next to family or children's content, and do not target by traits that people would not expect you to know. Where your product is regulated, the disclosures a host must say, and the claims you may make, are a matter for your counsel and the applicable rules. This chapter gives operational guidance only. Build a short approved claims sheet, share it with every host, and keep a record of what was said.
\nDynamic ad insertion: Stitching an ad into an audio file at the moment of play or download, so messages can change after publication. Host-read: An ad spoken by the show's own host in a personal voice, usually from a short brief. Ad pod: The group of ads that plays together in one break between content. Vanity link: A short, memorable web address spoken in an ad that redirects to a tracked page. Geographic holdout: A test that withholds a campaign from matched places to estimate what the campaign added.
\n1. Judging audio by clicks. There is nothing to tap, so a click based report will undervalue the channel and push you toward worse choices. 2. Reading a web page aloud as an ad. Long addresses, lists of offers, and phone numbers are hard to hear and almost never remembered. 3. Buying host-reads and never listening to them. A skipped disclosure or missing code is a real cost, and it is easy to catch with a transcript. 4. Trusting reported frequency as exact. Devices and accounts are counted separately in many systems, so true repetition is often higher than the number you see. 5. Running one big burst instead of a steady schedule. Concentrated spend can fatigue loyal listeners while missing the occasional ones.
\n1. Write one idea and one action per spot, then read it aloud against a stopwatch before recording. 2. Pair every campaign with at least one counting method that needs no click, such as a spoken code or a vanity link. 3. Plan a holdout from the start, so lift can be estimated instead of argued about later. 4. Ask sellers whether their podcast numbers follow the IAB Tech Lab measurement guidelines and whether anyone independent has verified them. 5. Keep a one page claims and disclosure sheet for every host, and check each read against it.
\n1. Ask for the show's older episodes in your test, since dynamically inserted ads may run on back catalog content long after release. 2. Use short, spellable words for vanity links, then check the redirect on a phone before launch. 3. Record two versions of a spot with different openings and rotate them, so you learn which first three seconds hold attention. 4. Compare branded search before and after audio flights, but check for other activity that week before crediting the audio. 5. Put pod position in the contract language or insertion order whenever the seller allows it, because a premium break near the start of a pod is worth asking about.
\nThree free tools cover most of what a small team needs here. Audacity (GPL) is an audio editor for cutting, cleaning, and leveling your spots and for trimming ad segments out of episodes. FFmpeg (LGPL and GPL) is a command line tool that converts audio between formats, adjusts loudness, and clips segments in bulk, which saves hours when you handle many files. Whisper (MIT) converts speech to text so you can check each host-read against your brief. None needs a subscription. For the auction itself, no open source option reaches parity with a full scale demand-side platform, so say so plainly to your team and plan to work through a partner.
\nAudio is a channel of voices, not clicks. Streaming spots deliver scale, host-reads deliver trust, and dynamic insertion delivers flexibility, and each should be bought for the job it does well. Because the listener has nothing to tap, measurement leans on spoken codes, vanity links, lift studies, and geographic holdouts, supported by publishers who follow agreed counting rules. Copy has to land in a single hearing, with one idea and one action. Frequency needs care across pods, shows, and devices, and host-reads need a quality check that free speech to text software makes cheap. A conservative default for family settings protects both the listener and the brand.
\n1. Judge audio by lift, codes, and links, never by click through rate. 2. Match the format to the job: scale from streaming spots, trust from host-reads, flexibility from dynamic insertion. 3. Write for a single hearing with one idea and one action. 4. Cap frequency by listener and show, and assume reported frequency understates real repetition. 5. Transcribe every host-read and check it against the brief and disclosures.
\nEvery audio buy needs a spoken code, a vanity link, and a matched holdout in the plan, and every host-read should be transcribed within a day of airing so you catch errors while they can still be fixed.
\n1. IAB Tech Lab. Podcast Measurement Technical Guidelines, current version. 2. IAB Tech Lab. OpenRTB specification, audio object fields, current version. 3. IAB Tech Lab. Podcast Measurement Technical Guidelines and Compliance Program. 4. Federal Trade Commission. Guides Concerning the Use of Endorsements and Testimonials in Advertising, 16 CFR Part 255. 5. OpenAI. Whisper speech recognition project, MIT license, public repository documentation.
"},{"id":"chapter-25-sports-streaming-and-live-event-digital-inventory","title":"Chapter 25: Sports Streaming and Live Event Digital Inventory","html":"By the end of this chapter, the reader will be able to: 1. Distinguish rights-holder inventory from aggregator inventory and explain why the difference decides whether a delivery promise can be kept. 2. Explain how dynamic ad insertion and regional versus national rights shape what a live sports impression actually is. 3. Compare direct, programmatic guaranteed, private marketplace and open exchange routes for live events. 4. Apply a supply path checklist to a proposed live sports deal before any budget is committed. 5. Design a live moment buying plan with fixed rules for triggers, caps and fallbacks. 6. Evaluate a live flight by reconciling the plan, the delivery log and the outcome data in one place.
\nHere is a question. If a seller promises you 500,000 impressions on one specific Saturday football game, what are the odds that the promise can be kept? Most brand managers answer that the seller must know, because the seller wrote the number down. That answer is often wrong.
\nA live game is not a product on a shelf. It happens once. The audience size is unknown until the whistle, the ad breaks arrive when the broadcast decides, and the number of streams in each ad slot depends on who tuned in, from where, on what device. A seller who has a direct relationship with the rights holder can forecast with some honesty. A seller who has only assembled scraps of inventory from exchanges is guessing, and the guess is printed on an insertion order with your name on it.
\nSports feels like the safest buy in video. The audience is paying attention, the content is brand-safe by reputation, and nobody gets up to make a sandwich during overtime. All of that can be true and the buy can still go badly, because the risk in live sports is rarely the audience. It is the paperwork between you and the audience: who actually holds the rights, who actually inserts the ad, and who actually counts the impression.
\nThis chapter is built around that paperwork. I will walk through where live inventory comes from, how rights and geography cut it into pieces, what dynamic ad insertion does to a plan, and how to buy moments rather than just minutes. I will also show how to check a flight afterward with free tools, so you can find out what happened without waiting for a seller's recap. Treat live sports the way a good coach treats a road game: respect the opponent, scout the building, and never assume the schedule will hold.
\nMost digital media is forgiving. If a banner underdelivers on Tuesday, you buy more on Wednesday. If a prerolled video campaign runs light, the algorithm catches up the following week. Live sports removes that forgiveness. The game starts at a set time, it ends at a set time, and the audience for that single event will never be assembled again.
\nThree planning changes follow from that.
\nFirst, the buy is one-shot. A regular season game, a playoff game and a championship are different products, even when the same teams play on the same service. Demand for the last of those can be many times the first, and so can the price. When you plan a season-long program, you are really planning dozens of separate events, each with its own audience and its own risk of being a blowout that people leave early.
\nSecond, the buy is time-bound. An impression served 20 minutes after the final whistle is worth far less than one in the last two minutes of a tied game, so spend should concentrate around the event with firm cutoffs afterward.
\nThird, the attention is real but not uniform. A fan watching a close game leans forward. A fan watching a lopsided game checks a phone. Fan attention follows the scoreboard, and a media plan that ignores the scoreboard is a plan written for a different sport.
\nThe cleanest way to sort live sports inventory is by how close the seller stands to the game.
\nRights-holder inventory comes from the entity that owns or licenses the broadcast: a streaming service that bought the rights to a league package, a league that streams games on its own platform, or a team with its own channel. These sellers can tell you which games they will carry, how many ad slots each game has, and roughly how many viewers have been in similar events. They can also be held to account, because they control the stream.
\nAggregator inventory is assembled. An aggregator is a company that gathers ad slots from many publishers, apps and connected TV channels, packages them, and resells them at a margin. Good aggregators are honest about what they hold and sell you a package of audiences with sports leanings. Weak ones describe that package as if it were a named game.
\nHere is the distinction that matters. If the package is described by audience (\"adults 25 to 54 who watch college football\"), the aggregator is selling something it can deliver, because audiences travel across many apps. If the package is described by event (\"this Saturday's rivalry game, guaranteed\"), the aggregator is promising access to a specific stream it may not touch. Trade press has reported that some middlemen guarantee impressions on specific games they cannot reach through the inventory they buy. I will not repeat anyone's numbers, because the useful lesson is simpler: the more precisely a seller names an event, the more precisely you should ask how the seller gets into it.
\nSports rights are sliced geographically in ways that confuse newcomers and sometimes confuse sellers. A game can be national on one service and regional on another. A local team's games may stream inside its home market only, with blackout rules that prevent viewing outside it. A championship may carry one set of ad slots nationally and let local affiliates insert their own in specific markets.
\nFor a brand manager, this matters in three ways.
\nThe first is geography matching. A regional furniture retailer has no use for a national impression in a city it does not serve. A national package with the ability to geotarget (limit delivery to chosen metro areas) may fit better than a regional rights deal, or the reverse. The right answer depends on whether your store footprint lines up with a team's home territory.
\nThe second is the stacking problem. Buy the national and regional streams of the same game and one fan can see you twice, with no frequency cap spanning both contracts. Keep geography clean in writing.
\nThe third is the rights fine print. Whether you may show team marks or player images in your creative is a licensing question outside the media buy, so have counsel confirm it before production begins. Operational guidance only.
\nOlder broadcasts baked ads into the video itself. Streaming usually does not. In dynamic ad insertion, the video player or a server in the delivery chain swaps in an ad at the moment a break begins, chosen for that viewer. Two people watching the same play can see two different ads, one national and one local, one for your brand and one for a competitor.
\nFor live events, that creates two practical consequences.
\nThe break is signaled, not scheduled. A marker in the stream tells the system that a break is starting, and the ad decision happens in a very short window. If the ad server is slow or the creative is not ready, the viewer may see a blank, a fallback ad or a repeat of something already shown. Ask every seller what happens when the ad decision fails. A good answer includes a fallback, a timeout and a log.
\nServer-side insertion also blurs what a viewer-level impression means. When the ad is stitched into the stream, the player may not report back the way it does for a banner. Some platforms offer pings (small signals that fire when the ad plays) and some offer only server counts. Server counts can be reasonable, but they measure what the server sent, not what a household saw. For a live game, that gap is one more reason to compare logs rather than trust a single number.
\nFour routes reach live inventory. A direct buy is a contract with the rights holder: best certainty and premium placements, but high minimums and slow timelines. Programmatic guaranteed reserves a set volume at a fixed price, executed through ad technology. A private marketplace is an invitation-only auction where the seller picks the buyers and sets floor prices. The open exchange is anyone-can-bid, and premium live events run thin there because rights holders sell them elsewhere first, so anything labeled as a specific game deserves extra scrutiny.
\nLive moment buying is the practice of defining, in advance, which moments in an event you want to be present for and what you will do if they do not happen. It sounds exciting. It is really a rulebook.
\nSuppose you sell sporting goods and want to appear in the final five minutes of close games. Write the rule down: \"if the score margin is within a stated number of points with a stated time left, release a defined budget.\" Then write the fallbacks. If the margin is large, spend nothing. If the game goes to overtime, extend the cap by a fixed amount, no more. Cap total spend per game and per week. Decide who is allowed to override the rules and who is not.
\nA rulebook protects you from yourself. In the heat of a game, a spend dial feels like a lever of power. Without caps, it becomes a way to blow through a quarterly budget during a single triple overtime.
\nNot every live sports dollar goes into a 30 second spot. Rights holders also sell integrations: a branded segment (for example, a \"player of the game\" graphic), a presenting sponsorship on a replay package, or a sponsored highlight clip that lives on a team channel after the game.
\nIntegrations are harder to compare with ad breaks because they are sold as packages and measured differently. A useful way to evaluate them is to ask four questions. What is the exposure, in minutes on screen, in appearances, and on how many platforms? What control do you have over the creative and over adjacency to other sponsors? What happens if the event is postponed or the game is a blowout? And what assets do you receive afterward: clips, logs, a recap you can audit?
\nBe honest about the limits. An integration builds familiarity, but you cannot prove it with a click. Plan it as brand work and measure it with a holdout comparison.
\nStart with a plan for three records. The first is the plan itself: games, dates, impressions promised, price, geography. The second is the delivery log: what each seller says was delivered, by game and by hour. The third is the outcome record: site visits, sign-ups, store visits or whatever matters to your business. Most failed flights share one cause. The three records live in three places, owned by three different people, and nobody ever lays them side by side.
\nHere is an illustrative example. Suppose a regional lender, call it Halvard Credit Union, buys three college basketball games with a promised 200,000 impressions per game. After the third game, the seller reports 580,000 total impressions delivered. That looks close to the promise of 600,000. But the delivery log shows 140,000 of the impressions on one game and 60,000 on another, with 380,000 arriving on the third game, which was a blowout that many viewers left early. The total hides the shape.
\nNow look at the outcomes. Online applications rose on the night of the first game, barely changed after the second, and did not move after the third. The report says \"97 percent of goal delivered.\" The reconciliation says the plan failed on two of three events.
\nYou need a place to do this comparison. Python, the general-purpose programming language, and DuckDB, a small database that runs on a laptop and reads spreadsheet-style files directly, are enough. The method is simple: load the plan, the log and the outcomes as three tables; join them on game and date; and calculate delivered share, price paid per delivered impression, and outcome change by game. A page of queries does the job.
\nCross-screen viewing adds a wrinkle: a fan who moves from television to phone may be counted twice. Ask each seller how it deduplicates, and record the answer.
\nThree public records help. An ads.txt file is a list a publisher posts naming who may sell its ad space. A sellers.json file is a list an ad platform posts naming who it sells for. A supply chain object travels with a bid and lists every company that touched it. Check that the seller appears in the publisher's list, count the intermediaries and ask why each exists, require game-level logs you can load yourself, and put a make-good clause in writing. Contract terms are for your counsel; this is operational guidance only.
\nEverything above assumes you will deal with several parties: rights holders, platforms, aggregators and measurement vendors. For many brand managers, particularly those with modest budgets, that web of relationships is the real cost of live sports. Each party adds a margin, a delay and a point of failure.
\nThis is the first of three places in this book where I describe my own companies,. If your goal is sports-anchored digital placement and athlete partnerships, there is a shorter path. Sports Media Inc., through SportsMedia.net, packages sports-focused digital placement and audience-based activation so that the supply path is short enough for you to inspect. For athlete and creator partnerships, including name, image and likeness arrangements structured with a conservative approach to minors and school rules, MySportsMedia.com/NIL provides the entry point.
\nI run these programs, so weigh my view accordingly. Apply the same questions to us that you apply to everyone else.
\nDynamic ad insertion: A method in which an ad is chosen and stitched into a stream at the moment a break begins, so different viewers of the same event can see different ads. Programmatic guaranteed: A deal that reserves a set volume of inventory at a fixed price and executes it through ad technology. Supply chain object: A record carried with a bid that lists each company that handled the ad opportunity between publisher and buyer. Blackout: A rights restriction that prevents an event from streaming in certain locations, often the home market of a team. Make-good: A contract remedy in which the seller provides credit or replacement inventory when promised delivery is not met.
\n1. Accepting a guarantee on a named game from a seller who cannot show a path to the rights holder. The precision of the promise should raise your suspicion, not lower it. 2. Judging a flight by total impressions instead of by game. A strong total can hide two failed events and one blowout that carried the numbers. 3. Buying national and regional streams of the same game without checking overlap. Frequency caps do not cross contracts, so the same fan can see you repeatedly. 4. Letting live moment spending run without caps and written override rules. The excitement of a close game is a poor budget manager. 5. Producing creative before rights are cleared. Team names, player images and league marks usually need separate permission, and delays follow when no one asked.
\n1. Ask who holds the rights and who sits between you and them before negotiating price. The answer shapes everything that follows. 2. Write a one page live buying rulebook for each program, covering triggers, caps, fallbacks and who may override. 3. Reconcile every flight with the plan, the delivery log and the outcome record side by side, by game and by day. 4. Prepare two or three creative versions matched to game states, plus a quiet fallback for situations where a celebratory message would be tone deaf. 5. Keep a conservative default for anything involving minors, athletes in school programs or family audiences, and route compliance questions to counsel.
\n1. Ask what happens when an ad decision fails mid-break. Sellers with a clear fallback and a log have usually thought about it before you asked. 2. Request delivery data as a file you can load, not a slide. A slide cannot be rechecked after the call ends. 3. Schedule a post-game review within a week while the seller's team still remembers what happened. 4. Negotiate make-good terms before the season starts, when sellers are friendliest and the game list is still hypothetical. 5. For regional brands, test one home-market stream before committing to a season. A single game teaches you how the seller reports.
\nPython (Python Software Foundation License) and DuckDB (MIT License) are all you need for flight reconciliation. Python handles file loading and cleanup, such as reading a seller's delivery log and standardizing game names and dates. DuckDB then runs ordinary SQL queries directly against those files, joining plan, delivery and outcome tables on a laptop with no server and no subscription. A first reconciliation can be as small as one script and a handful of queries: delivered share by game, cost per delivered impression, and outcome change by event. No open source option reaches parity with a full-scale demand-side platform for buying itself, and I will not pretend otherwise. For checking what you bought, these two are sufficient.
\nLive sports is a one-shot, time-bound environment where the main risk is the chain of parties between advertiser and audience. Rights-holder inventory can be forecast and held to account, aggregator inventory varies in honesty, and the more precisely a seller names a game the more closely the supply path deserves inspection. Regional rights, dynamic ad insertion and cross-screen viewing change what an impression means. Live moment buying works only under written rules, and measurement works when plan, delivery and outcomes are reconciled game by game with free tools.
\n1. A guarantee on a named game is only as good as the seller's path to the rights holder. 2. Totals hide the shape of a flight, so reconcile by game and by day. 3. Regional and national streams of the same game can overlap, and caps do not cross contracts. 4. Live moment buying needs triggers, spending caps, fallbacks and named override authority written in advance. 5. Verify the supply path with ads.txt, sellers.json and the supply chain object before signing, and write make-good terms into the contract.
\nIf a seller will not name the rights holder above them and deliver game-level logs you can load yourself, skip the buy, and reconcile every flight that does run in Python and DuckDB within a week of the final whistle.
\n1. Interactive Advertising Bureau Tech Lab. ads.txt Specification, version 1.1. Authorized digital sellers standard. 2. Interactive Advertising Bureau Tech Lab. sellers.json and SupplyChain Object Specification. Supply path transparency standards. 3. Interactive Advertising Bureau Tech Lab. Open Measurement SDK documentation. Viewability and verification signals for video. 4. Society of Cable Telecommunications Engineers. SCTE-35 Digital Program Insertion Cueing Message. Ad break signaling standard. 5. DuckDB Foundation. DuckDB Documentation. Querying CSV and Parquet files with SQL.
"},{"id":"chapter-26-video-creative-specifications-and-testing","title":"Chapter 26: Video Creative Specifications and Testing","html":"By the end of this chapter, the reader will be able to: 1. Identify the spec families that cover most video placements and state what each one requires. 2. Apply safe-zone and audio-off design rules to a finished edit. 3. Design a single-variable test of hooks and lengths with a stated decision rule. 4. Evaluate test results honestly against the sample size that produced them. 5. Diagnose a broken VAST wrapper using a short, repeatable check sequence. 6. Construct a pre-flight QA list that a second person can run in under 20 minutes.
\nAsk a team what its last video test taught it and listen for the pause. Most video tests are a waste of money. A brand manager running a small budget splits it across four versions of a spot, waits a week, sees one version ahead by a few clicks, declares a winner, and calls it learning. It is not learning. It is noise with a spreadsheet around it.
\nThere is a second objection, and it is a fair one. Testing takes budget away from the thing you are trying to sell. Every dollar spent on a losing variant is a dollar that did not go to the best one. A good editor with a strong instinct will often beat a mediocre testing program, and a mediocre testing program can make a team timid, because it rewards small safe variations over big swings.
\nI hold both objections to be partly right. Sloppy testing is worse than no testing, because it gives you confidence you did not earn. And a team that only tests button colors in video form, a slightly different logo placement here, a slightly different end card there, will converge on a very comfortable local peak.
\nSo why write the chapter? Because the alternative is not wisdom. The alternative is guessing in a way nobody can audit. The real answer is narrower than \"always test.\" Test when the decision is large, the budget can bear it, and you will read the result with honest limits. Before any of that, get the basics of the file right, because a perfect hook inside a clipped, mis-sized, silent-on-mute video never gets a fair trial. Think of a quarterback with a great arm throwing to a receiver standing out of bounds. The throw is fine. The play is dead.
\nThis chapter covers the specs first, then the tests.
\nEvery platform publishes its own spec sheet, and every sheet changes. Memorizing them is a losing game. What lasts is the family each placement belongs to, because the creative decisions follow the family.
\nThink in four families. Wide (widescreen) is the format of television screens, desktop players and most connected TV inventory. Tall (vertical) is the phone held upright, used by short-form feeds and story placements. Square and its near cousin, the slightly tall portrait frame, live in social feeds where the video shares the screen with text and other posts. Short fixed formats are the bumper of about six seconds and the 15 or 30 second spot that connected TV and many in-stream placements expect.
\nLength is its own family. Bumpers cannot be skipped, so they carry one idea and no setup. Skippable in-stream gives you a short window before the viewer can leave, so the first moments must work as a standalone message. Connected TV spots are usually not skippable, which changes the craft: you earn attention by being worth watching, not by shouting early.
\nPractical rule: produce a master in the widest, highest-quality form you can, then plan the cut-downs before you shoot. Shooting wide and cropping to tall afterward usually leaves faces too small and text too tiny. Shoot with the tall frame in mind, keep the subject centered, and leave room around it.
\nAlways confirm current numbers in the platform's own documentation before you export. Treat any number in a book, including this one, as a starting point.
\nA safe zone is the part of the frame the platform will not cover with its own interface. Buttons, captions from the app, the account name, the progress bar and the call-to-action overlay all sit on top of your video. Put a headline where a button lands and the viewer never reads it.
\nPlatforms publish safe-zone guides, often as downloadable templates. The consistent pattern: keep key text and logos away from the bottom band and the upper edge in tall formats, and away from the far left and right edges in wide ones. When in doubt, hold important elements to the central block of the frame and test the export on an actual phone, not just an editing monitor.
\nNow the matter of sound. Many viewers watch with the sound off, especially in feeds and in public places. Other placements, such as connected TV and some short-form apps, are mostly heard. You cannot assume either, so design for both.
\nAudio-off design means the video makes sense with no sound at all. Burn in captions that match the speech, keep them short, and keep them readable at phone size. Show the product doing the thing within the first moments. Use on-screen text for the one claim that matters. Then check the reverse: with the sound on, does the voice or music add something the visuals cannot? A video that works in both modes has two chances to hold the viewer instead of one.
\nCaptions also help accessibility, which is both a courtesy and, in some contexts, an obligation. Counsel decides what applies to you. The operational habit is to caption everything.
\nFile weight is the quiet reason otherwise good creative underdelivers. A heavy file loads slowly, and a viewer on a weak connection leaves before the first frame. Many ad servers also reject files above a size limit, or quietly serve a lower-quality transcode that looks soft.
\nThe goal is the smallest file that still looks clean at the size viewers will see. That means choosing a sensible frame rate, a bitrate matched to the content, and a widely supported container and codec, typically H.264 video with AAC audio in an MP4 wrapper. Motion-heavy footage needs more bitrate than a talking head against a plain wall.
\nHere is a short habit that saves arguments: export two versions, one at the platform's ceiling and one noticeably lighter, and compare them side by side on a phone. If you cannot see the difference, ship the lighter one. FFmpeg, the free command-line converter covered in the callout below, makes this a two-line job and keeps the settings in a script anyone on the team can rerun.
\nKeep a naming convention that carries the concept, the length, the aspect ratio and a version number in the file name. When a result comes back six weeks later, you will want to know exactly which file it was.
\nA hook is the opening moment that earns the next few seconds of attention. A length test asks whether the same idea works at 15 seconds and at 30. Both are cheap to build and both can mislead, so the design matters more than the edit.
\nChange one thing. If version A has a different opening, a different voiceover and a different end card than version B, you learned that A and B differ. You did not learn why. Keep everything identical except the opening three seconds, and you have a clean hook test. For a length test, cut the longer version so the shorter one is a strict subset, same opening, same close.
\nWrite the decision rule before launch. For example: \"We will compare completion rate to the 15 second mark and cost per desired action, and we will declare a winner only if one version leads on both after each has been served at least a stated number of impressions.\" If you decide what counts after you see the numbers, you will find a way to like the numbers.
\nSplit the audience evenly and run the versions at the same time. Running A this week and B next week tests the calendar, not the creative. If the platform's own split tool rebalances spend toward an early leader, understand that it may starve a late bloomer before it gets a fair look. When a clean split matters, run the variants as separate, equally funded ad sets, and use an open source experimentation tool such as GrowthBook to log which version each exposed visitor saw when the landing page is yours.
\nOrder matters, because the biggest swings come first. A rough ranking from largest likely effect to smallest:
\n1. The core concept and promise, meaning what the video is about. 2. The opening three seconds. 3. The length. 4. The call to action and end card. 5. Captions style, music and small visual treatments.
\nTeams usually do the reverse. They polish the end card and never test the concept, because the end card is easy to change. A different concept costs real production effort, so it feels risky. It is also where most of the winning and losing happens.
\nStart with two clearly different concepts, not two shades of the same one. Suppose a regional furniture retailer, Juniper Lane (an illustrative brand), tests a \"problem and fix\" story against a \"room makeover\" montage. Whichever wins becomes the control. Next round, hold the winning concept and test two openings. Then test length. Each round narrows the field, and each result is small enough to explain.
\nIf the budget is thin, one test per month, done well, beats six sloppy ones. And leave room for a wild card now and then, a concept that breaks your pattern. Testing should protect you from bad guesses, not from good bold ones.
\nHere is the hard part. Small numbers swing wildly. Flip a coin 10 times and eight heads is unremarkable. Flip it a thousand times and 800 heads means something is wrong with the coin.
\nTake an illustrative case. Suppose two hooks each get about two thousand views, and version A produces 30 clicks while version B produces 40. B looks 33 percent better. But the gap is 10 clicks, and a handful of lucky clicks could produce that gap. Run the same test again and the order might flip. The right move is to keep the test running, or to call it a tie and choose on a secondary factor such as production cost.
\nPractical guardrails, none of which need a statistics degree:
\nBe equally honest about what the platform reports. A view-through conversion may include people who would have bought anyway. When the stakes are large, a geographic holdout, where some regions see the ads and comparable regions do not, gives a cleaner answer than any dashboard. How attribution models bend results is a subject for its own chapter, and I will not repeat it here.
\nFinally, creative wears out. A winner that was strong in week one can sag by week six as the same people see it again. Watch frequency and the trend line, and rotate in a fresh cut before the decline becomes expensive.
\nVAST, the Video Ad Serving Template from the IAB Tech Lab, is the standard format ad servers use to tell a video player which file to play, where to send tracking pings, and what happens on a click. A wrapper is a VAST response that points to another VAST response, so a tag can pass through several systems before reaching the actual video.
\nWrappers break quietly. The ad slot stays empty, the player shows a spinner, or impressions are counted that never played. Run these checks before launch:
\n1. Open the tag URL in a browser or fetch it from the command line and confirm it returns valid XML, not an error page. 2. Follow the wrapper chain by hand. Count the hops. Players often cap how many wrappers they will follow, and every hop adds delay. 3. Confirm the final response contains a media file in a format the target player supports, and that the file URL loads. 4. Check that the impression, quartile and click tracking URLs are present and fire once each. 5. Confirm everything is served over a secure connection, since many players refuse mixed content. 6. Verify the skip offset and duration in the XML match what you booked.
\nDo this in a test player or the ad server's own preview before spending money. The payoff is catching, for 10 minutes of effort, the error that would otherwise burn a week of budget showing nothing.
\nPre-flight is a short checklist, run by someone other than the editor, because the editor can no longer see the video fresh. Copy this list, trim it, and make it yours:
\n1. Aspect ratio, length and file name match the placement and the naming convention. 2. File plays cleanly from start to end on a phone and a desktop, with sound on and off. 3. Key text and logos clear the safe zones for the placement. 4. Captions are present, accurate, spelled correctly and readable at phone size. 5. First three seconds make a clear point without the sound. 6. File weight sits under the platform limit and loads quickly on a throttled connection. 7. VAST tag, where used, returns valid XML with working media and tracking. 8. Landing page matches the promise made in the video. 9. Claims, music and talent permissions are cleared for the placements and dates booked. Operational guidance only, counsel decides the rest. 10. Test records are filled in: variable, hypothesis, decision rule, minimum exposure.
\nAnything involving children or sensitive categories gets the conservative default: no targeting around minors, plain claims, and a second reviewer.
\nHook: The opening moment of a video designed to earn the viewer's next few seconds of attention.
\nSafe zone: The region of the frame that stays clear of platform interface elements such as buttons, captions and progress bars.
\nVAST: The IAB Tech Lab standard for describing a video ad to a player, including the media file, tracking and click behavior.
\nCompletion rate: The share of started views that reach the end of the video.
\nCreative fatigue: The decline in response that happens as the same audience sees the same video repeatedly.
\n1. Testing several changes at once, so a result cannot be traced to a cause. Change the opening and nothing else. 2. Declaring a winner after a few hundred views and a small gap. Set minimum exposure before launch and stick to it. 3. Cropping a wide master into a tall frame and leaving key text under platform buttons. Plan the tall version before the shoot. 4. Relying on sound to carry the message, then running the ad in placements where most viewers are muted. 5. Skipping the VAST chain check and discovering an empty ad slot after the budget is gone.
\n1. Shoot with every planned aspect ratio in mind, and keep the subject centered with room to crop. 2. Caption every video and review the captions as carefully as the script. 3. Write the hypothesis and decision rule before launching any test. 4. Test concept first, opening second, length third, and finishing details last. 5. Keep a creative log recording what was tested, what happened, how much data supported it, and what you decided.
\n1. Export a heavy and a light version and compare them on a phone. If you cannot tell them apart, ship the light one. 2. Name files with concept, length, ratio and version so results stay traceable for months. 3. Judge the first three seconds by watching with the sound off and asking whether a stranger would know what the ad is for. 4. Build modular edits, with swappable openings and endings, so one shoot yields many testable versions. 5. Run a rotation calendar for long campaigns, planning fresh cuts before frequency climbs rather than after results drop.
\nFFmpeg (LGPL and GPL, depending on build options) converts, compresses and inspects video files from a script, which makes spec checks repeatable. Kdenlive (GPL) and Shotcut (GPL) are full video editors for building the master and its cut-downs. Whisper (MIT) transcribes speech to text, giving you a first-draft caption file to correct by hand. GrowthBook (MIT core) records which version of a page or experience each visitor saw and helps you read the outcome. Free tools do not fully replace a paid ad server's tag preview, so use the browser and the command line for the VAST checks described above.
\nVideo tests are only as good as the files and the discipline behind them. Place each video in the right spec family, protect it with safe zones and audio-off design, and keep file weight under control so it gets a fair showing. Test big things first, change one variable at a time, write the decision rule in advance, and treat small gaps as leans, not verdicts. Check the VAST chain before launch and run a short pre-flight list through a second pair of eyes. Rotate fresh creative before fatigue costs real money.
\n1. Spec families (wide, tall, square, short fixed) guide decisions better than memorized spec sheets. 2. Safe zones and audio-off design protect the message from the platform's interface and from muted viewers. 3. A clean test changes one variable, uses an even split, and states its decision rule before launch. 4. Small samples produce convincing but unreliable winners, so set minimum exposure in advance. 5. VAST checks and a second-person pre-flight catch failures that no dashboard will warn you about.
\nWrite one sentence naming the single variable you are testing and the minimum number of views each version must receive ahead of the next video launch, then have someone other than the editor run the pre-flight list.
\n1. IAB Tech Lab. Video Ad Serving Template (VAST) specification, current version. Standard for video ad response format, wrappers and tracking. 2. IAB Tech Lab. Video Multiple Ad Playlist (VMAP) specification, current version. Standard for ad break scheduling. 3. Media Rating Council. Viewable Impression Measurement Guidelines. Standards for video impression validity. 4. FFmpeg project documentation. Command-line reference for video conversion and inspection. 5. Platform creative specification pages and safe-zone templates from each video platform you buy on, consulted at time of export.
"},{"id":"chapter-27-cross-screen-reach-and-frequency","title":"Chapter 27: Cross-Screen Reach and Frequency","html":"By the end of this chapter, the reader will be able to: 1. Distinguish summed reach from deduplicated reach and explain why the gap changes a media plan. 2. Calculate combined reach across channels under an independence assumption and state where that assumption fails. 3. Interpret a frequency distribution and identify where spend turns into waste. 4. Evaluate the identity, panel and census limits that make cross-screen counts uncertain. 5. Construct a simple reach curve from campaign data and use it to set a frequency target. 6. Design a channel comparison a small team can repeat each quarter, including a plain-language clean-room request.
\nPicture a regional furniture brand, entirely illustrative, that I will call Harlan and Pine. The marketing manager buys three things: connected TV (streaming television on a living-room screen), online video, and streaming audio. Each platform sends a tidy report. Connected TV says it reached 300,000 households. Online video says 250,000 people. Audio says 150,000 listeners. The manager adds them up, tells the owner the campaign reached 700,000 people, and feels good about it.
\nThe market area holds about 500,000 households. Somebody just reached 140 percent of the town.
\nOf course nobody did. The same families were counted three times, once by each platform, and the report that looked like a victory measured how often the budget touched the same living rooms. A large block of shoppers saw nothing at all, and a smaller block saw the pine sofa ad 20 times in a week and started muting it.
\nThat is the assumption this chapter challenges: that bigger platform numbers mean bigger reach. They do not. Every platform counts only what it can see, using its own identity system, and none of them knows what the others did. Adding reports together is like adding the attendance figures from three games played in the same stadium on one weekend and announcing how many different people love your team.
\nWhat follows is how to estimate both honestly, across screens, with a team of two and a spreadsheet.
\nEvery ad platform reports its own audience, and each report is true inside its own walls. The trouble starts when a buyer carries those numbers into one slide and treats them as additive.
\nSummed reach counts exposures per platform. Deduplicated reach counts unique people across all platforms. The first number is always the larger one, and the gap between them is the overlap. Overlap is not a defect in the plan. People watch television on the couch and scroll video on a phone beside it, and in many households the same evening produces both. Some overlap is simply how modern viewing works.
\nWhat matters is that you know its size, because overlap changes three decisions. It changes how you judge a new channel, since a channel that mostly reaches people you already reached is worth much less than its standalone number. It changes how you split the budget, since money moved toward a channel that reaches fresh people often buys more new reach than money added to the strongest one. And it changes what you promise the owner.
\nWhen you have no cross-platform data at all, you can still build a rough estimate with one formula. It rests on an assumption, independence, which means that being reached by one channel tells you nothing about being reached by another. Under that assumption:
\nCombined reach = one minus the product of each channel's unreached share.
\nHere is a worked example with round, illustrative figures. Suppose a market of 100,000 households. Connected TV reaches 30 percent, online video reaches 20 percent, and streaming audio reaches 10 percent.
\nThe unreached shares are 70 percent, 80 percent and 90 percent. Multiply them: 70 percent of households, then 80 percent of those, then 90 percent of those, leaves about 50 percent unreached. Combined reach is therefore about 50 percent, or roughly 50,000 households. The summed figure, 30 plus 20 plus 10, would have claimed 60 percent.
\nSo even under the friendliest assumption the sum overstates reach by about 10 points, and the overstatement grows as individual reaches grow. Two channels that each reach 60 percent of a market cannot combine to 120 percent; under independence they reach 84 percent.
\nNow the caution, and it is a serious one. Real audiences are not independent. Heavy media users tend to be heavy on every screen, so the same people sit in several audiences at once. That positive correlation means true combined reach is usually lower than the independence formula says. Treat the formula as a ceiling for planning, not a forecast. If your buy depends on 50 percent reach and the formula gives you exactly 50, you have no cushion.
\nAverage frequency is the number that hides the most. Suppose a campaign reports an average frequency of five. That sounds like a healthy, deliberate level. It could also describe a campaign in which half the audience saw the ad once and a small group saw it 30 times.
\nA frequency distribution shows what share of the exposed audience received one impression, two, three, and so on. Ask for it every time. It is the picture that average frequency flattens.
\nHere is a simple illustration. Suppose 100,000 people were reached and the average frequency was six. The distribution might look like this:
\n| Exposures received | Share of reached people | | --- | --- | | One to two | 40 percent | | Three to five | 25 percent | | Six to nine | 15 percent | | 10 or more | 20 percent |
\nAverage frequency says six. But look at the shape. Two in five people got a thin, probably forgettable dose, while one in five absorbed a heavy dose, and that heavy group probably consumed a disproportionate share of the impressions, and therefore of the money.
\nHow much is too much? I will not give you a universal number, because none exists, and anyone who gives you one is guessing. The right ceiling depends on creative quality, category, how simple the message is, and how long the purchase cycle runs. A short, memorable spot for a familiar product needs fewer exposures than a first introduction of an unfamiliar financial service. What I will say is this: the second and third exposure usually do more work than the 15th, and the high end of the distribution is where money quietly leaks. For brands that reach households with children, or any audience in a sensitive category, I set the ceiling conservatively. Repeated exposure to a product message is a fairness question as much as an efficiency one, and a low cap is the right default.
\nTo deduplicate, someone has to recognize that the person on the television and the person on the phone are the same. That recognition is called identity resolution, and it is the weak joint in every cross-screen plan.
\nThere are three common ways platforms try. The first is logged-in identity: the viewer signs into an app on several devices, so the platform knows it is one account. This is the most reliable, but it only works inside that platform. The second is a household graph, a model that links devices sharing a home network and infers they belong to one household. It is useful for television but cannot tell which person in the house saw what. The third is probabilistic matching, which uses signals like timing and location to guess that two devices belong together. Guesses are fine for averages and unreliable for any single person.
\nThe honest takeaway is simple. Cross-screen reach is always an estimate with a margin of error, and the margin is wider for individuals than for groups. Plan for households and audiences, not for following one person from sofa to phone.
\nTwo kinds of measurement sit underneath most reach numbers, and a buyer should know which one is speaking.
\nPanel data comes from a sample of people or households who agree to be measured, usually with a meter or an app. The sample is then scaled up to represent the population. Its strength is that it observes real people across many channels, including those that cannot be matched any other way, and it can report demographics. Its weakness is size. A panel can be statistically sound for a broad audience and noisy for a small one, such as a narrow zip-code cluster or a niche hobby group. A local business with a small trade area should be wary of panel reach claims that slice very thin.
\nCensus data is a full count of events, such as every ad served or every stream started, taken from the platform's own logs. Its strength is completeness, since nothing is sampled. Its weakness is that it counts devices and accounts, not people. A census log knows an ad was delivered to a device. It does not know that three people watched it or that nobody did.
\nNeither is better in every case. Census data is the right tool for delivery and frequency inside one platform. Panel data is the right tool for estimating how many different people were reached across platforms.
\nA reach curve plots how reach grows as spend or impressions grow. It always bends. The first dollars find the easiest, most available people, and each additional dollar has to work harder to find someone new, so the curve flattens.
\nA simple way to describe the curve is a saturating shape. Reach starts near zero, rises quickly, and approaches a ceiling, which is the share of the market that the channel can ever reach. Two numbers matter most: where the ceiling sits and how fast you approach it. A channel with a low ceiling, such as a small local audio network, stops being useful quickly. A channel with a high ceiling, such as broad streaming television, keeps paying longer but still bends.
\nSuppose, for illustration, an online video buy produces these results across four weekly budgets.
\n| Weekly spend | Reach of the market | | --- | --- | | 10,000 dollars | eight percent | | 20,000 dollars | 14 percent | | 40,000 dollars | 21 percent | | 80,000 dollars | 27 percent |
\nRead the steps. Going from 10,000 to 20,000 dollars bought six more points of reach for 10,000 extra dollars, under two thousand dollars per point. Going from 40,000 to 80,000 dollars also bought six more points, but for 40,000 extra dollars, over six thousand dollars per point. Same reach, four times the price. That is diminishing returns in plain arithmetic.
\nThis is where the curve earns its keep. It tells you the spend at which a channel stops being your cheapest source of new people, and that is the point at which you should move the next dollar to another screen.
\nYou do not need a media-mix modeling department. You need three steps, a spreadsheet, and the discipline to repeat them.
\nStep one is to build a reach curve for each channel. Pull reach and spend from past flights, even small ones. If you have run a channel at two or three different budget levels, you have points on a curve. If you have run it once, ask the platform for a reach forecast at several budget levels, and treat the answer as a hypothesis to test. Fit a simple saturating curve through the points. A spreadsheet trendline can do it, and so can a short script in Python or R. When data is thin, use conservative curves and widen your margins.
\nStep two is to set a frequency target. Work backward from the message. A familiar product with a short, clear ad might aim for a modest range. A new product with a more complex pitch needs more. Whatever you choose, write down a target range and a ceiling, for example three to six per person per month with a hard stop at eight, and make the ceiling a planning input, not a hope. Then apply it across screens. If three platforms each cap at four per week, the household total can reach 12, so lower each cap until the combined ceiling matches your intent.
\nStep three is to compare channels at your target frequency. For each channel, ask what incremental reach it gives per dollar once the prior channels are already in the plan. Start with the channel that has the best curve. Add the second channel at the point where the first has flattened. Re-estimate combined reach using the independence formula, then shade the result downward to reflect real overlap. Repeat until the next dollar buys more in a new channel than in an existing one.
\nEventually a good plan runs into a wall. You want to know how many of your customers saw the campaign, but the platform will not hand over its logs, and you will not hand over your customer list. A data clean room is a controlled environment where two parties can match their data and ask questions without either side seeing the other's raw records. Think of two sports teams comparing scouting reports through a neutral third party who only reports the overlap, never the notes.
\nIn practice, you upload a hashed (scrambled, one-way) version of your customer identifiers. The platform does the same with its viewers. The room matches them and returns only aggregate answers, such as how many matched people saw the ad and how many times. Usually it refuses to answer when the group is too small, which protects individuals.
\nFor a small team, the practical move is modest. Before you spend anything on a clean-room project, ask each major platform a plain question in writing: what overlap, frequency and match-rate reporting do you provide for my first-party list, and what are the minimum group sizes? The answer tells you whether the tool is within reach. Any clean-room work involving personal data raises contract and privacy questions that your counsel should review.
\nDeduplicated reach: The count of unique people or households exposed to a campaign across all channels, with repeat counts removed. Frequency distribution: A breakdown of how many reached people received one exposure, two, three, and so on, which shows the spread that an average hides. Identity resolution: The process of linking devices, accounts or households so that exposures can be attributed to the same person or home. Panel: A sample of consenting people or households whose viewing is measured and then scaled to represent a larger population. Data clean room: A controlled environment where two parties match data and receive only aggregate answers, without exposing raw records to each other.
\n1. Adding platform-reported audiences together and presenting the total as unique reach, which counts the same households several times. 2. Reporting average frequency alone and never asking for the distribution, so a heavy tail of overexposed people goes unnoticed. 3. Setting a per-platform frequency cap on each channel without adding the caps together, which lets one household receive far more exposures than intended. 4. Treating the independence formula as a forecast when real audiences overlap, which leaves the plan with no cushion. 5. Comparing a television household impression and a phone impression as if both were one person, which inflates reach in mixed plans.
\n1. Ask every platform for reach and frequency by flight and record them in one shared sheet before building any combined number. 2. Write a frequency target range and a hard ceiling for the whole campaign, then divide the ceiling across screens. 3. Fit a simple reach curve for each channel and update it after every flight with the actual results. 4. Shade independence-based combined reach downward and label it as an estimate with a stated assumption. 5. Report reach as a range with the method named, so the owner knows what was measured and what was modeled.
\n1. Request a frequency distribution, not an average, in every vendor reporting template you sign off on. 2. When a platform quotes reach for a narrow local audience, ask how large the underlying sample was before you rely on it. 3. Test a new channel at a small budget and read its reach curve before you scale, since the first dollars always look cheaper than the later ones. 4. Prefer logged-in identity within a platform for frequency control, and treat household graphs as audience-level tools only. 5. Keep your own first-party customer list clean and current, because any future overlap study depends on it.
\nPython, R and PyMC-Marketing (Apache 2.0) cover the curve-fitting work in this chapter. A short Python or R script can fit a saturating reach curve to your flight results, and PyMC-Marketing adds Bayesian methods that express uncertainty when data is thin, which is the usual state of a small team. DuckDB (MIT) lets you query large exports of delivery logs and campaign files on a laptop with plain SQL, so you can count unique households and build a frequency distribution without a database server. No open source tool reproduces a full panel or an identity graph, because those depend on data you cannot download. Plan to rely on platform reports for that input and on these tools for the analysis.
\nPlatform reports each count their own audience, so adding them overstates how many different people you reached, and average frequency hides the overexposed tail. A planning formula for combined reach under independence gives a useful ceiling, but real audiences overlap, so the true figure is usually lower. Identity resolution across screens is imperfect, panel and census data answer different questions, and reach curves show where each channel stops being a cheap source of new people. A small team can build curves, set a frequency ceiling, compare channels, and learn from each flight, and clean rooms offer a closer look when the data and the contract allow it.
\n1. Summed platform reach is never unique reach, and the gap between them is the overlap you are paying to repeat. 2. Combined reach under independence equals one minus the product of the unreached shares, and it is a ceiling because real audiences overlap. 3. Frequency distribution, not average frequency, shows where spend turns into waste. 4. Identity, panel and census limits make cross-screen reach an estimate, so report it as a range with the method named. 5. A reach curve built from your own flights tells you when to move the next dollar to another screen.
\nYour cross-channel plan starts with one sheet that lists each channel's reach, an independence-based combined reach shaded down for overlap, and a household frequency ceiling that sums the caps across screens. Report the range to your owner, then compare it with what happened.
\n1. Ehrenberg, A.S.C. Repetitive Advertising and the Consumer. Journal of Advertising Research, on frequency and response. 2. Naples, M. J. Effective Frequency: The Relationship Between Frequency and Advertising Effectiveness, Association of National Advertisers. 3. Media Rating Council. Outcomes and Data Quality Standard. Also see the Council's cross-media audience measurement standards. 4. Interactive Advertising Bureau and Media Rating Council, Cross-Media Audience Measurement Standards, Phase One Video. 5. Interactive Advertising Bureau Tech Lab, Data Clean Room Guidance and Recommended Practices. 6. PyMC Labs, PyMC-Marketing documentation, media mix modeling and saturation curves.
"}]},{"title":"Part V: Programmatic Display and Native","items":[{"id":"chapter-28-programmatic-display-fundamentals","title":"Chapter 28: Programmatic Display Fundamentals","html":"By the end of this chapter, the reader will be able to: 1. Describe the sequence of events that turns a page load into a served display impression. 2. Compare the four common ways programmatic inventory is bought and match each to a campaign goal. 3. Evaluate a display placement for viewability, creative weight and supply quality before spending against it. 4. Diagnose why a display campaign reports strong activity but weak business results. 5. Design a display plan that assigns reach and retargeting jobs instead of click targets. 6. Select the open source tools that serve owned inventory and strengthen a publisher's header bidding setup.
\nHere is a contradiction I see in almost every account review. The display dashboard says the campaign delivered millions of impressions, a healthy click rate and a few hundred conversions. The finance partner across the table asks one question: if you had turned it off, would we have sold less? Nobody in the room can answer.
\nThat gap is the whole subject of this chapter. A dashboard is built by the system that sold you the media, and it counts what that system can see. Finance counts what the business actually earned. Those are different ledgers, and display is the channel where they drift furthest apart, because display sits in the margins of a page while the person is busy doing something else.
\nMany brand managers inherit a quiet assumption: display is a cheap click channel, so judge it by cost per click. That assumption does damage in two directions. It pushes budget toward the placements that produce accidental clicks, such as thumbs landing on a banner in a mobile game. And it starves the placements that do the real work of reminding a person who already heard of you.
\nI am not going to tell you display is worthless. It is not. A banner can be the third touch that finishes a sale begun on television or search, and it can keep a small brand in front of a buyer for two weeks at a price no other format matches. But it earns that credit only when you understand how the impression is bought, what you actually receive, and what job you gave it.
\nEvery programmatic impression begins with a person opening a page or an app. The page contains an empty slot. Before the content finishes loading, the publisher's software sends a request describing that slot to an ad exchange: the size, the site, the device type and whatever user signals are permitted.
\nThe exchange forwards that request to buyers. Each buyer's software, called a demand-side platform (the system that bids on behalf of advertisers), decides in a few milliseconds whether this impression fits a campaign, and if so, how much to offer. The exchange picks a winner, commonly through an auction, and returns the winning ad to the page, where the browser renders it.
\nThe standard that describes these messages is OpenRTB, maintained by the IAB Tech Lab. What you need to know is that a bid is a prediction about a person the buyer cannot fully see.
\nThree failure points matter to you. The request can be thin, carrying little useful information, so the bid becomes a guess about nothing. The bid can win, but the ad can load too slowly for the person to see it. And the render can happen below the fold or in a background tab. In each case the log reads \"impression delivered\" and your money has gone.
\nSuppose a campaign wins one million auctions. Maybe 100,000 of those ads never finish loading before the visitor scrolls on. Another 300,000 load in a corner nobody looks at. Your invoice counts all one million. This is why the later sections of this chapter focus on what happens after the win.
\nThe same slot can reach you through different doors, and the door determines price, control and transparency.
\nThe open auction is the widest door. Any approved buyer may bid on available inventory, and the highest offer wins. Prices run low and scale is enormous, but you inherit whatever supply the exchange carries.
\nA private marketplace is an invitation-only auction. A publisher or a curator decides which buyers may bid and sets a price floor. You get a narrower, cleaner pool of sites, usually at a higher cost per thousand impressions.
\nA preferred deal fixes the price and gives you first look at inventory, but you are not obliged to take it. It suits a buyer who likes a publisher and wants predictable access without a commitment.
\nA guaranteed deal reserves a set volume at a set price. It looks like a traditional insertion order with automated trafficking, and it carries the most certainty about where your ad runs.
\nThe decision rule is simple. When you are testing and do not yet know which sites matter, start in private marketplaces, because the open auction teaches you mostly about fraud. When a placement has proven it moves your numbers, convert it to a preferred or guaranteed deal. Open auction becomes useful later, once you hold a tight list of exclusions and known-good domains, and when reach matters more than precision.
\nOne honest note on tooling. A full demand-side platform needs bidder infrastructure, exchange connections and fraud filtering at a scale no open source project reaches at parity, so buying runs through the in-house planning path rather than a self-built bidder.
\nDisplay creative comes in a small set of standard shapes, and knowing the family helps you plan assets without memorizing dimensions.
\nThe medium rectangle, roughly 300 by 250 pixels, is the workhorse because it fits inside article text and sidebars on both desktop and mobile. The leaderboard, about 728 by 90, sits across the top of a page. The mobile banner, about 320 by 50, lives at the bottom of a phone screen. The half page, about 300 by 600, is the tall unit that gives a headline and a picture room to breathe.
\nThe practical point is not the dimensions. It is that each unit gives you a different amount of space for a message, and shorter units demand sharper copy. A mobile banner holds one idea, one logo and one action. If your creative team tries to fit a three-step offer into it, nobody reads any of it.
\nPlan at least four sizes for any campaign. Fewer, and you are locked out of inventory the buyer's system cannot fill without matching creative.
\nKeep animation restrained. A short loop that ends on a clear frame works. Fast flashing annoys readers and sits badly against a conservative family standard. The final frame must stand alone, because many people see only that frame.
\nViewability is a measure of whether an ad had a fair chance to be seen. The common industry standard counts an impression as viewable when at least half of the ad's pixels were on screen for at least one second. It is a floor, not a promise. An ad can meet it and still be ignored.
\nWhy does this matter to your budget? Because a nonviewable impression is a bill with no delivery. If a placement is only viewable about half the time, your effective price per viewable impression is double the quoted price. That arithmetic is rarely printed on the report.
\nUse viewability as a filter, not a goal. Set the target too aggressively and the buyer's software chases safe placements at inflated prices. Set a minimum threshold, review domain level results monthly and drop the bottom performers.
\nThird party measurement tags add weight and cost, so ask who pays for the verification.
\nCreative weight is the file size your ad asks a browser to download. Heavy creative loads slowly, and a slow ad is a missed ad. Visitors on weaker mobile connections are the first to be lost, and they are often the people you most want to reach.
\nMost exchanges and publishers cap initial load at a modest file size, and many reject or throttle anything larger. A banner built from a large photograph, three web fonts and a chain of third party scripts competes against the page's own content for bandwidth, and the page usually wins.
\nHere is a simple discipline. Compress images until the difference is barely visible. Use a single font, or convert the text to a graphic. Avoid pulling in scripts from other domains unless they are required. Test every creative on a throttled mobile connection before launch, and watch what appears first.
\nLike a broadcast director cutting to the replay, you get one moment. Put your message and logo in the first frame and treat everything after as a bonus.
\nAlso track load failures. Some platforms report a gap between impressions won and impressions rendered, and that gap is a direct measurement of wasted money. A large gap on one domain tells you either the creative is too heavy for that environment or the supply is poor. Both are reasons to stop buying there.
\nSupply quality is the question of whether the page where your ad appeared was a real place for real people. The open internet includes excellent publishers and thoroughly bad ones, and the auction does not tell them apart for you.
\nThe industry built a few public files to help. A publisher posts an ads.txt file naming the sellers authorized to sell its inventory. Companion files and a supply chain object describe who is involved in the chain between the page and your bid. These are verification tools, not guarantees, but a placement that fails them is a placement to refuse.
\nMade for advertising sites exist mainly to host ads, load pages stuffed with units and recycle traffic. Domain spoofing disguises cheap supply as a premium name. Invalid traffic, produced by software instead of people, can pass a quick look and still fail a closer one. Apps aimed at children raise a separate concern, and a conservative default says to exclude them unless the campaign is explicitly designed for that audience and counsel has reviewed the rules.
\nThe defense is dull but effective. Build an approved list of domains you have checked by hand. Keep an exclusion list that grows with every audit. Review the placement report weekly during a launch and monthly afterward, sorted by spend. The top 20 domains by spend deserve a human look. If you cannot describe what a site is in one sentence, you should not be paying for it.
\nA click is a poor proxy for the work display does. Consider who actually clicks a banner. Some are interested. Many are accidental. A sizable share are already your customers on their way to a login page, and a small number are software.
\nNow suppose a retargeting campaign shows 80 clicks and 12 purchases on a spend of a thousand dollars. The platform claims 12 sales at about 83 dollars each. But suppose those 12 buyers were people who had put items in a cart yesterday and would have returned anyway. The banner did not create those sales. It stood near them and took a photograph.
\nThat is the heart of the click trap. Optimizing display toward clicks teaches the buying system to find people who click, and those are not the people most likely to buy. Cost per click falls, the click rate looks fine, and the business result stays flat.
\nBetter questions: did sales in exposed regions rise relative to matched regions where you held back? Did new customers, as opposed to returning ones, increase? These answers come slower, and finance will respect them more than a click report.
\nTreat view-through credit with suspicion. A platform that claims a sale because someone saw an ad earlier in the week will find many sales to claim. A holdout test is the cleanest way to discover how many were real.
\nDisplay does two jobs well, and both are modest.
\nThe first job is reach: repeated, low cost presence in front of a defined audience over time. This works best when you pair it with a stronger medium. A person who saw a television spot or a video ad and then meets your banner on a news site recognizes the brand faster. Display here is the echo, not the voice. Set a frequency cap, because the same person seeing the same ad 30 times in a week is not being reminded, they are being irritated.
\nThe second job is retargeting: showing ads to people who already visited your site or engaged with your product. It converts well on paper because the audience is already warm. That is the trap, and the reason to test it with holdouts. Use it for people who showed real intent and then left, such as visitors to a pricing page, and keep it away from anyone who has already purchased. Build a short window, a clean exclusion of buyers and a creative that answers the likely objection.
\nA worked example. Suppose a regional furniture brand spends six thousand dollars on a month of reach display and two thousand dollars on retargeting. It holds back one matched metro area for the whole month. If the exposed metros show higher sales than the held back one, the difference is the incremental lift, and finance can work with it. If they show none, the retargeting budget moves to something else.
\nOwned inventory sits on the other side of the table. If you run a website with traffic, you can sell or serve your own display slots. Revive Adserver does this on your own server, handling campaigns, targeting and reporting without a per-impression fee. If you are a publisher, Prebid.js lets multiple buyers compete for your slots in a single header auction. These are the open tools for the inventory you control.
\nImpression: One instance of an ad being counted as served to a page or app, whether or not anyone saw it. Viewable impression: An impression where at least half of the ad was on screen for at least one second under the common industry standard. Private marketplace: An invitation-only auction in which a publisher or curator controls which buyers may bid and sets a price floor. Header bidding: A publisher technique that lets several buyers bid on a slot at the same time before the page's ad server makes its decision. ads.txt: A public file on a publisher's domain listing the companies authorized to sell that publisher's ad space.
\n1. Judging display by click rate and cost per click, which rewards accidental clicks and teaches the system to hunt for them instead of buyers. 2. Buying open auction supply from day one with no approved list, then discovering months later that a large share of spend landed on low quality sites. 3. Shipping a single heavy creative in two sizes, which locks the campaign out of inventory and loads too slowly on mobile connections. 4. Retargeting everyone who ever visited the site, including recent buyers, and then crediting the platform with sales that were already going to happen. 5. Accepting platform reported conversions as incremental without a holdout, then presenting those numbers to finance as proof of effect.
\n1. Start tests in private marketplaces and move proven placements into preferred or guaranteed deals once they show lift. 2. Build at least four creative sizes, keep each file light, and put the message and logo in the first frame. 3. Review placements by domain every month, sorted by spend, and add every questionable site to the exclusion list. 4. Set frequency caps on every campaign, and set separate caps for reach and retargeting audiences. 5. Run a matched geography holdout on any display line large enough to matter, and report incremental results next to platform results.
\n1. Ask for the gap between impressions won and impressions rendered on each domain. A wide gap is a quiet sign of a weak or heavy placement. 2. Test creative on a slow phone connection before launch. The page your design team sees is not the page your visitors see. 3. Verify ads.txt for any publisher you buy direct. A missing file or a missing seller is a reason to pause the conversation. 4. Keep a short exclusion window on retargeting so buyers drop out of the audience quickly after purchase. 5. Pair display flights with a stronger medium and schedule them to overlap, so the banner echoes a message already heard.
\nTwo open tools matter here. Revive Adserver (GPL) is a self-hosted ad server you can run on your own infrastructure to deliver, schedule and report on ads for inventory you own, such as your website or a newsletter. Prebid.js (Apache 2.0) is a header bidding library that publishers install on their pages so several buyers can compete for each slot at once, which usually improves price discovery. Neither one replaces a demand-side platform for buying across the open web, and I will not pretend otherwise. They protect your position as a publisher and your control over owned placements. Operational guidance only: confirm contract and privacy terms with counsel before sharing visitor data.
\nA display impression is the end of a fast chain: a request from the page, a bid from a buyer, an auction and a render. At every link, something can go wrong while the log still reads as delivered. You choose among open auction, private marketplace, preferred and guaranteed deals based on how much you know about the placement. You plan a small set of ad sizes with light files, and you filter placements by viewability and supply quality. Display fails when it is asked to produce clicks, and it works when it is given reach and retargeting jobs and tested with holdouts. Owned inventory can run on open tools, while buying across the open web runs through the in-house planning path.
\n1. A dashboard counts what the seller can see, and finance counts what the business earned. Learn to reconcile the two before you report. 2. The deal type you choose, open auction, private marketplace, preferred or guaranteed, sets your level of control over where the ad appears. 3. Viewability, file weight and supply quality decide how much of your spend becomes something a person could actually see. 4. Click optimization on display finds clickers, not buyers, so give the channel reach and retargeting jobs instead. 5. A matched geography holdout is the cleanest way to find out whether display added anything beyond what would have happened anyway.
\nHold back one matched region on every display flight, cap frequency, and judge the result on sales finance can verify rather than on clicks the platform reports.
\n1. IAB Tech Lab. OpenRTB specification, current version. Message formats for requests and bids in real time auctions. 2. IAB Tech Lab. ads.txt specification, current version. Authorized digital sellers file format and usage. 3. IAB Tech Lab. sellers.json and SupplyChain object specifications. Supply path transparency standards. 4. Interactive Advertising Bureau and Media Rating Council. Viewable impression guidelines. Standard viewability definition for display. 5. Prebid.org. Prebid.js documentation. Header bidding concepts and publisher setup.
"},{"id":"chapter-29-header-bidding-and-supply-path-optimization","title":"Chapter 29: Header Bidding and Supply Path Optimization","html":"By the end of this chapter, the reader will be able to: 1. Describe how a header bidding auction runs and where the buyer's bid request enters it. 2. Explain why one impression can reach a buyer through many routes and why those routes are not equal. 3. Interpret a publisher's ads.txt file, an exchange's sellers.json file, and the SupplyChain object in a bid request. 4. Select a small set of direct and near-direct paths using stated criteria instead of habit. 5. Conduct a path audit on a live campaign and calculate the share of spend reaching the publisher. 6. Draft path requirements that can be written into an insertion order and enforced.
\nMost buyers will never shop for header bidding, and they should stop worrying about it. Header bidding is a publisher-side tool. The publisher installs it, the publisher tunes it, and the publisher decides which exchanges get a seat. You do not configure it, and no setting in your buying platform turns it on or off.
\nNow the strongest counterargument, and it is a good one: you cannot ignore it either. Header bidding is the reason the same impression now arrives at your door from four, six or a dozen directions. Before it, a publisher passed inventory down a ranked list, one partner at a time. After it, many partners bid at once, and every partner that wins the right to sell can forward that same opportunity to every buyer it knows. Your platform hears about one page view repeatedly. It bids on each copy. You may even win against yourself.
\nSo the assumption worth challenging is the comfortable one, that the buying platform handles routing and the brand manager handles creative and audience. Routing is where margin leaks. Every extra hop between you and the publisher is a company taking a cut, and none of those cuts show up as a line item you approved.
\nThis chapter takes the buyer's seat. You will not learn to write auction code. You will learn what the auction does to your bids, how to read the three public documents that describe who is allowed to sell what, how to choose a few direct paths, how to audit the paths you already use, and how to write the results into the contract your agency or trading desk signs. Where the law or a contract is involved, this is operational guidance only, and counsel decides what the paper must say. Where children's content or sensitive categories are involved, the conservative default applies: a path you cannot verify does not carry that spend.
\nPicture a stadium that used to sell each seat by calling one ticket broker, then the next, then the next, accepting the first yes. That was the waterfall. A publisher's ad server offered an impression to partner one at a set price floor. If partner one passed, partner two got a look, and so on. The highest bidder did not always win. The first partner willing to meet the floor did.
\nHeader bidding changed the sequence. A small piece of code runs when the page loads, before the ad server decides anything. It asks many partners for a bid at the same moment, collects the answers, and passes the best of them into the ad server's own decision alongside any direct deals. The most common open implementation is Prebid.js for web pages, with Prebid Server handling the same job from a server so the browser does less work.
\nFor you, three consequences follow.
\nFirst, more of your bids are competitive. A bid that once lost silently in a waterfall can now win on price. Second, the publisher chooses a handful of exchanges to call, so the exchange you happen to use may or may not be one of them. Third, and this is the part buyers miss, the exchanges that win seats in the wrapper then resell what they hold to many buying platforms. One impression becomes many bid requests, each carrying a different route.
\nThat is why the question is no longer whether you can reach an impression. You can reach almost any impression. The question is which route you take and who gets paid along it.
\nSuppose a mid-sized regional news site, which I will call the Harbor Gazette (illustrative), sells through five exchanges, and two of those exchanges also resell through a third. A single page view can reach your buying platform through as many as seven distinct routes. Your platform receives seven bid requests for the same person on the same page at the same moment.
\nA well built buying platform tries to detect these copies and bid on one. Many do this imperfectly. Detection depends on matching identifiers that the sellers do not always pass cleanly, and a platform may prefer the route that gives it the highest margin rather than the one that sends you the most inventory value. If you are unlucky, you send a bid down two routes, win both, or win one at a higher price than the other would have cost.
\nIllustrative arithmetic helps. Suppose you plan 60,000 dollars of monthly spend on the Gazette and similar publishers. If four of every 10 impressions arrive on routes that carry a heavier fee stack than the cheapest available route, and the difference is five cents per dollar of spend on those routes, you give up about twelve hundred dollars a month without buying a single extra impression. The figures are made up. The mechanism is not. Duplicate routes are a tax on the buyer who does not pick a lane.
\nThere are two basic relationships between a seller and a publisher. A direct relationship means the publisher itself has authorized the seller, usually an exchange it integrates with, to sell its inventory. A reseller relationship means the publisher authorized one party, and that party passes the opportunity along to another.
\nResellers exist for sound reasons. A small publisher may not have the staff to manage a dozen partner integrations. A network may package thousands of small sites into something a buyer can actually purchase. A reseller can also bring a specific audience or format. None of that is a crime. The cost is that each layer takes a fee, adds a few milliseconds, and moves you one step further from the party you can hold accountable.
\nHere is a working rule I use. A direct path is the default. A one-hop reseller is acceptable when the publisher or network offers something you cannot get directly, such as scale across many small sites, and when the reseller declares itself properly in the public documents. Anything past one reseller needs a written reason. The rest of the chapter shows how to find out which category a given route falls into.
\nTwo small text files carry most of the answers. Both are public. Neither requires a login.
\nThe first is ads.txt, which stands for authorized digital sellers. A publisher places it at the root of its own domain. Each line names an advertising system, an account identifier within that system, and one of two words: DIRECT or RESELLER. DIRECT means the publisher itself holds the account. RESELLER means the publisher has authorized that account to sell on its behalf through another party. Lines may also carry a certification authority identifier.
\nThe second is sellers.json, which an exchange publishes about itself. For each account identifier, it lists who the seller is, the seller's type (publisher, intermediary or both), and whether the name is shown or kept confidential. Matching the two files is the core skill. You take a line from the publisher's ads.txt, find the same account identifier in the exchange's sellers.json, and check that the entity named in the second file agrees with the publisher in the first.
\nWhen they agree, the route is verified. When they do not, you have one of several situations:
\nNone of these proves fraud. Files go stale, staff change, and small publishers copy partner snippets without reading them. But a buyer who spends against unverifiable routes is betting that the paperwork is merely messy. A conservative default is to hold back spend until the route checks out, and to hold back faster when the content is aimed at families or covers a sensitive subject.
\nads.txt and sellers.json describe who is allowed to sell. The SupplyChain object, often called schain, describes who actually did, impression by impression. It rides inside the bid request as a list of nodes. Each node names one company that handled the opportunity, gives that company's account identifier, says whether it is the one that directly controls the inventory, and may add a few supporting details.
\nRead it in order. The first node should be the publisher's direct partner. The last node should be the party sending the request to you. Between them sit any resellers. A chain with two nodes means one hop. A chain with six means five companies stand between you and the publisher.
\nThree checks turn the object into a decision tool. Count the nodes. Compare the first node's account identifier against the publisher's ads.txt. Confirm that every node's seller appears in the sellers.json file of the company it names. A chain that passes all three is complete. A chain that fails the second or third is incomplete, and a chain with a missing object is silent about its own history.
\nSome buying platforms drop the field before it reaches you. Ask your agency whether it is received and logged. Without it, you audit from the outside, which still works but takes longer.
\nAuditing is only half the work. The other half is deciding where you want spend to go, which is a choice about a small number of relationships rather than hundreds of routes.
\nStart from the publishers and apps that matter most to the plan, then ask four questions of each exchange that can sell them. 1. Does the exchange hold a DIRECT line in the publisher's ads.txt? 2. Does its sellers.json confirm the seller by name? 3. Does it pass a complete chain on the bid requests you receive? 4. Does it give you log-level detail you can export, so you can check results yourself?
\nCount the exchanges that say yes to all four. For most buyers, a short list emerges, often between three and six. That is plenty. A buyer who routes spend through a dozen exchanges is not buying reach. They are buying duplicates.
\nThen rank the survivors by what they offer that others do not. One may have premium video. Another may hold the best access to a set of local publishers. Another may simply give cleaner logs. A direct path with an acceptable price beats a cheaper reseller route you cannot verify. If the direct route is costlier, ask what it buys you: clearer reporting, a named contact, or a more stable floor. Often the premium is smaller than it looks once duplicate impressions and fee stacks are subtracted.
\nA path audit is a monthly exercise, not a one-time cleanup. Run it in five steps.
\nStep one: export your bid or impression log for the last 30 days, grouped by publisher domain or app and by selling route. Step two: for the top 50 publishers by spend, download each ads.txt. Step three: download the sellers.json for each exchange in your log. Step four: match them, and mark each route as verified, partial or unverified. Step five: compute the publisher revenue share, which is the portion of each dollar of media spend that reaches the publisher, using whatever fee data your platform and agency disclose.
\nAn illustrative result might look like this. Suppose you spent 80,000 dollars last month. Verified direct routes carried 50,000 dollars, one-hop verified resellers carried 18,000 dollars, and 12,000 dollars went to routes that were partial or unverified. That last bucket is where you act first. You do not need to cancel it all. You move the largest unverified routes to a hold, ask the seller to fix its files, and shift budget to verified paths while the answers arrive.
\nRecord what you find in a dated spreadsheet. Over several months, patterns appear. Some routes get better. Others go quiet when you ask questions.
\nThen write the lessons into the insertion order, which is the contract that tells your agency or trading desk how to spend your money. Operational language you can ask counsel to review includes the following.
\nAgencies sometimes resist. They may argue that detailed path control reduces scale or limits their ability to optimize. Fair. Offer an exception process, not a loophole: they can request approval for any route outside the rules, with a stated reason, and you decide.
\nHeader bidding: A publisher-side auction method in which many partners bid at once, before the publisher's ad server makes the final choice. Direct path: A route in which the seller holds an account the publisher itself controls, shown as DIRECT in the publisher's ads.txt. Reseller: A company authorized to sell a publisher's inventory through another party, shown as RESELLER in ads.txt and adding at least one hop. Seller identifier: The account code that names a specific seller inside an advertising system, used to match ads.txt lines against sellers.json entries. Hop count: The number of companies that handle an impression between the publisher and the buyer, read from the nodes in the SupplyChain object.
\n1. Assuming the buying platform picks the best route for you, when many platforms favor the route that serves their own margin or inventory access. 2. Treating a lower cost per thousand impressions on one route as proof of efficiency, without checking whether a reseller layer, a duplicate bid or an unverifiable seller sits behind it. 3. Reading ads.txt alone and skipping sellers.json, which leaves you unable to confirm that the account identifier belongs to the entity the publisher meant to authorize. 4. Running a path audit once, filing the results, and never repeating it, even though files, partners and routes change every month. 5. Leaving path requirements out of the insertion order, so that every verified path you chose is a preference the agency may ignore.
\n1. Make the direct path the default and require a written reason for every route that adds a second reseller or more. 2. Keep an approved list of a few exchanges per major publisher group, and revisit it on a fixed schedule rather than after a problem. 3. Ask your agency for the full chain and selling account identifier in your logs, and confirm the first delivery before you rely on it. 4. Match ads.txt against sellers.json for your top publishers every month and record the status of each route. 5. Put the verification rules, fee disclosure and reporting cadence in the insertion order, with counsel reviewing the wording.
\n1. When a seller will not show its name in sellers.json, ask why before bidding, and treat a slow or vague answer as an answer. 2. Group your audit by publisher, not by exchange, because the same site can look clean through one route and cloudy through another. 3. Look for the same impression appearing in your logs more than once within a second or two, which is the clearest sign you are bidding against yourself. 4. Review path rules before large seasonal pushes such as a major sports finals or holiday window, when buyers rush and routes multiply. 5. Share your approved list with publishers you work with directly, since a quick confirmation call often fixes a stale ads.txt file faster than any support ticket.
\nPrebid.js and Prebid Server (Apache 2.0) are the open implementations of header bidding on the publisher side. Reading their documentation shows you exactly how bids are requested, compared and passed on, and if you own sites or apps you can run them yourself. For your audit, Python and Scrapy (BSD) can fetch ads.txt files and sellers.json files at scale, so you can match them in a spreadsheet or small script. One honest limit: no open source demand-side platform reaches parity with commercial platforms at full scale, so most buyers still rent that layer. These tools give you the knowledge and the verification, which is where your negotiating position starts.
\nHeader bidding lets many exchanges compete for the same impression at the same moment, which improves publisher yield and also multiplies the routes by which one opportunity reaches a buyer. The buyer cannot switch it off, but can decide which routes carry spend. Three public records make that decision possible: the publisher's ads.txt, the exchange's sellers.json, and the SupplyChain object inside each bid request. A short list of verified direct and one-hop paths, a monthly audit, and path rules written into the insertion order turn that knowledge into control over fees, duplicates and accountability.
\n1. Header bidding is a publisher tool, but its consequences, duplicate bid requests and layered resellers, land in the buyer's logs and budget. 2. A route is verified only when the publisher's ads.txt line, the exchange's sellers.json entry and the SupplyChain nodes agree with each other. 3. Every extra reseller hop takes a fee and removes accountability, so a second hop needs a written reason. 4. A short, ranked list of direct paths usually outperforms a long list of cheaper routes you cannot check. 5. Path rules only bind your agency when they appear in the insertion order, with fee disclosure and a reporting schedule.
\nThis month, export 30 days of bid logs, match your top 50 publishers' ads.txt against sellers.json, and send your agency a one-page rule set limiting spend to verified routes with at most one reseller hop.
\n1. IAB Tech Lab. ads.txt specification, authorized digital sellers, current version. 2. IAB Tech Lab. sellers.json specification, current version. 3. IAB Tech Lab. SupplyChain object specification (schain) for OpenRTB, current version. 4. IAB Tech Lab. OpenRTB specification, current version. 5. Prebid.org. Prebid.js and Prebid Server documentation. 6. Scrapy Project. Scrapy documentation.
"},{"id":"chapter-30-private-marketplaces-and-curated-deals","title":"Chapter 30: Private Marketplaces and Curated Deals","html":"By the end of this chapter, the reader will be able to: 1. Describe how a private marketplace deal moves from a deal identifier to a cleared impression. 2. Explain why curation adds a middle layer and calculate how its fees stack on top of the base media cost. 3. Apply a break-even rule that converts a price premium into the extra outcome a curated deal must deliver. 4. Evaluate a curation vendor with a structured set of diligence questions. 5. Design a holdout test that compares a curated deal to open exchange buying. 6. Judge whether a curated deal adds real value or repackages open supply under a new label.
\nStart with arithmetic. Suppose your open exchange buying clears at 10 dollars per thousand impressions. A salesperson offers a curated deal at 16 dollars. The pitch deck says \"premium,\" \"audience matched,\" \"brand safe.\" The premium is six dollars, which is 60 percent more than you pay today.
\nNow ask the only question that matters. For the curated deal to merely tie the open exchange, each impression has to produce 60 percent more of whatever you are buying. Not 10 percent more. Not \"better quality.\" Sixty percent more sales, sign-ups, store visits or lift in recall, measured the same way on both sides. If it produces 40 percent more, you lost money while feeling smart.
\nMost brand managers never run that calculation. They assume the premium label is the evidence. It is not. A label is a claim, and a claim is something you test.
\nThe second assumption worth challenging is that a private deal means you bypassed the crowd. It often does not. A private marketplace is a gated auction, but the impressions inside it may come from the same publishers, passed through the same sellers, that you could have bought openly. What changed is who stands between you and the publisher, and what that person takes. Sometimes the new party earns its fee by removing junk, matching audiences and shortening the path. Sometimes it is a toll booth on a road you already drive for free.
\nThe cost of proof is why it is worth buying. A holdout is not cheap, tidy quality reports are, and this chapter gives you the mechanics, the fee math, the diligence questions and a test plan so the premium has to pay for itself on your numbers, not the vendor's.
\nA private marketplace, usually shortened to PMP, is an auction that only invited buyers can enter. The publisher, or the seller working for the publisher, decides which inventory goes into it and who may bid. Open exchange buying works the other way: any approved buyer can see the impression and bid.
\nThree pieces define a deal. The first is the deal identifier, a string of characters your buying platform uses to find that specific package of inventory. The second is the floor, the lowest price at which the seller will accept a bid. The third is the inventory definition: which sites or apps, which placements, sometimes which audience segments. The technical language for all of this lives in the OpenRTB specification, the shared standard that lets buying and selling systems talk to each other.
\nDeals come in a few flavors, and the differences change what you can promise your finance team.
\nA preferred deal gives you first look at an impression at a fixed price. A private auction deal is the classic PMP: a floor, a limited bidder list, and a clearing price set by whoever else shows up. A guaranteed arrangement commits both sides to a volume, much like a direct buy. Know which one you are signing, because a floor with no volume commitment means the seller owes you nothing if inventory sells elsewhere.
\nOne practical point surprises people. Having a deal identifier does not mean you will spend. Your bid must clear the floor and your targeting must match the deal terms. Check bid requests received, bids submitted and wins, in that order, in the first week.
\nCuration is the newer idea, and it deserves careful reading. A curator is a party that assembles a package of supply, often by combining selected publishers or sellers with audience data, and then offers that package to buyers as a deal. The curator may be a data company, an agency group, a technology vendor or a publisher collective. The package shows up in your buying platform with a deal identifier, just like a traditional PMP.
\nThe pitch is reasonable. The open exchange holds an enormous volume of impressions, and most of it is not worth your money. Someone who sorts through that volume, removes poor sites, trims duplicate paths to the same publisher and attaches audience signals can save you work. Think of a scout in baseball: you could watch every high school game yourself, but a scout narrows the list. You pay for that judgment, and the question is whether it is good.
\nHere is the structural point. Curation puts one more party between your money and the publisher, and that party takes its own slice as an added fee, a revenue share, or a markup folded into the floor. Often you cannot see which.
\nThat is not automatically bad. Layers are acceptable when each one removes more waste than it costs. The trouble starts when the layer is invisible, because you cannot compute the break-even. This ties to a principle from earlier in the book: buyers who control the path make better decisions than buyers who shop on rate alone. Curation can help with path control only if the curator tells you the path.
\nTake the 16-dollar curated deal from the opening and open it up. This is an illustrative breakdown, not a quote from any vendor. Real structures vary, and your contracts decide what is true for you.
\n| Layer | Illustrative share of 16 dollars | |---|---| | Publisher revenue | eight dollars | | Seller or exchange fee | two dollars | | Curator fee or markup | four dollars | | Buying platform fee and data charges | two dollars |
\nOn the open exchange at 10 dollars, the same layers might look like this: publisher revenue of seven dollars, seller or exchange fee of two dollars, and buying platform fee of one dollar. There is no curator line.
\nCompare what reaches the publisher. In this example the publisher gets eight dollars on the curated deal and seven dollars on the open buy. You paid six dollars more, and the publisher received one dollar of it. The other five dollars went to intermediaries. That is a fine outcome if the curator's audience matching and filtering is worth about four dollars to you. It is a poor one if the curator mostly renamed supply you could have reached directly.
\nAsk for the publisher revenue share as a number, per deal, in writing. Then compute the effective price per useful impression. If 20 percent of open exchange impressions are unviewable, your real open cost is higher than 10 dollars, and the curated premium is smaller than 60 percent. Adjust both sides before you compare.
\nRun the break-even rule in this order. Take the curated price, divide it by the open price, and subtract one. That is the premium. Adjust both prices for waste. Then ask whether the curated deal can plausibly deliver that much extra outcome per impression. A deal that costs 30 percent more after adjustment needs 30 percent more outcome, no less.
\nCuration tends to earn its fee when your brand safety needs are narrow and a curator has already done the screening, when you need audience precision your own data cannot supply, or when a curator can consolidate dozens of resellers into a handful you can audit.
\nCuration is weakest when the package is simply a list of well-known sites that you could buy through your own platform with a site list and a few filters. Suppose a curator sells you \"premium news and lifestyle\" at 16 dollars, and your own platform, with a custom site list of the same publishers, reaches those placements at 11 dollars. The curator is selling you your own filter at a five dollar markup.
\nHow do you tell the difference? Look for these signals.
\nThe overlap test. Ask the vendor for the full publisher list behind the deal, then compare it to the placements you already buy openly. High overlap with a large price gap is a warning. Low overlap with strong quality metrics suggests the curator has access you lack.
\nThe data-origin test. If the audience layer comes from a third-party segment, ask where the data came from, how fresh it is, and what share is modeled rather than observed. Audience labels built on stale or inferred signals add cost without precision. For children, health topics, finance and other sensitive categories, set a conservative default: exclude them from curated targeting unless legal counsel and your compliance lead have approved the segment and its consent basis.
\nNone of these tests prove value. They tell you whether it is worth paying for a proper experiment.
\nBefore you sign anything, put these questions in an email and keep the replies. Vague answers are answers.
\n1. Who are the publishers behind this deal, and can you give me the full list? 2. Is your fee a separate line, a revenue share, or built into the floor? 3. What share of the price I pay reaches the publisher? 4. How many intermediaries sit on each path? 5. Where does the audience data come from, and is any of it modeled? 6. What happens if the deal underdelivers? 7. Will you support a holdout test and give me impression-level logs?
\nThe last question often separates serious vendors from the rest. A vendor confident in the product welcomes a fair test. One who resists is protecting something. Remember that these are operational questions. Contract terms, data rights and privacy obligations belong with your counsel, who decides what is acceptable.
\nQuality metrics should answer a narrow question: is this deal giving me more of what I want per dollar than my alternative? Collect them at the deal level, not the campaign level, so a good deal is not hidden by a bad one.
\nPrice metrics: effective cost per thousand impressions after fees, cost per viewable thousand, and cost per completed view for video.
\nDelivery metrics: bid rate, win rate and spend pacing against plan.
\nQuality metrics: viewability, completion, invalid traffic rate and brand suitability flags, all from the same measurement source on both sides.
\nPath metrics: number of intermediaries, share of spend reaching the publisher, and the match between declared and observed paths.
\nOutcome metrics: conversions, sign-ups or store visits. These are noisy at small volume, which is why the experiment below matters more than the dashboard.
\nBreak results out by publisher and placement. If three sites drive most of the outcome, buy those three directly and drop the curator.
\nA fair test needs a control. Without one, you are comparing a curated deal in one month to open buying in another month, and seasonality, creative fatigue and competitor activity will swamp whatever the deal contributed.
\nHere is a plan you can run with modest tools. The figures are illustrative.
\nStep one: pick one outcome and one measurement window. For example, online orders over six weeks. Decide in advance what counts, and do not change it after results arrive.
\nStep two: match your markets. Suppose you can buy by metropolitan area and you have 40 markets of similar size and history. Rank them by past sales, then pair neighbors in the ranking and randomly assign one of each pair to the curated arm and the other to the open exchange arm. Pairing keeps the groups comparable.
\nStep three: add a holdout. Set aside a smaller set of markets, say eight, that receive neither deal, only your normal baseline activity. This is the group that tells you what would have happened without either test. The curated and open arms each measure lift against it.
\nStep four: equalize everything else, including creative, frequency caps, flight dates and bid strategy. If the curated deal cannot spend its budget, record that as a result.
\nStep five: set your decision rule before launch. For example, the curated deal wins only if its incremental result per dollar, after all fees, exceeds the open arm by at least the premium you calculated. With a 60 percent premium, you need 60 percent more incremental outcome per dollar of spend, and a statistically credible margin.
\nStep six: analyze with an honest method. CausalImpact, a statistical package that builds a forecast of what the treated markets would have done without the campaign, is a good fit for comparing a treated group against matched controls over time. DuckDB, a free database that runs on a laptop, can hold your impression logs and market-level sales and let you write the comparisons as simple queries.
\nSuppose the open arm produced 90 incremental orders per 10,000 dollars and the curated arm produced 120. That is 33 percent more against a 60 percent premium, so the curated deal lost on value. Renegotiate the fee, narrow the deal to its best publishers, or walk away. If it had produced 160, that is 78 percent more, above break-even, and the deal earns a larger share of the plan.
\nSmall samples produce false winners, so if your volume cannot support the test, run a longer, simpler comparison rather than pretend to precision. And never let the vendor choose the markets or compute the result.
\nPrivate marketplace (PMP): An auction in which the seller invites selected buyers to bid on defined inventory under agreed terms, usually with a floor price and a deal identifier. Curation: The practice of assembling a package of supply, often with audience data attached, and selling it to buyers as a deal. Floor price: The lowest bid the seller will accept for an impression in a given auction or deal. Supply path: The chain of systems and companies that carry an impression from the publisher to the buyer, each of which may take a fee. Holdout: A set of markets or users that receives no test advertising, used to estimate what would have happened without it.
\n1. Treating the word \"premium\" as evidence and skipping the break-even calculation, so a 60 percent price increase is accepted without proof of 60 percent more outcome. 2. Comparing a curated deal to an open buy on headline CPM alone, ignoring waste, viewability and the fee line that the curator quietly added. 3. Letting the vendor design the test, pick the markets or report the lift, which hands the referee's whistle to the team selling the product. 4. Signing a minimum commitment before running a test, which removes your bargaining power and turns diligence into a formality. 5. Reading deal-level averages without breaking them out by publisher, so one strong site hides a long tail of weak placements.
\n1. Calculate the premium and the break-even outcome before the first meeting, and bring the number to the table. 2. Request the full publisher list, the curator fee treatment and the publisher revenue share in writing for every deal. 3. Compare curated supply against your own site list and filters on the open exchange to see what the curator actually adds. 4. Run a matched-market test with a true holdout, with the decision rule written down before launch. 5. Keep a deal log with price, premium, incremental result and decision, and review it before every renewal.
\n1. Start with a small commitment and a short term. A vendor who believes in the product will accept a test period. 2. Ask for impression-level logs. If a vendor can only provide summaries, you cannot audit the path or the quality. 3. Negotiate the fee structure as aggressively as the floor. A flat fee or declared share is easier to compare than a markup hidden in the price. 4. Rerun the comparison when a deal renews. Supply quality, data freshness and competitor behavior drift over time. 5. If a curated package is mostly a list of recognizable publishers, buy those publishers directly through your own site list and keep the fee.
\nTwo free tools cover the analysis side of this chapter. DuckDB (MIT license) is an embedded analytical database that runs on a laptop, so you can load impression logs, deal reports and market sales files and query them with plain SQL without standing up a server. CausalImpact (Apache 2.0 license) is a statistical package that estimates what a treated market would have done without the campaign, using matched control markets, and reports the lift with a credible interval. Used together, they let a small team run a holdout analysis and show its work. Neither one buys media for you. At full scale, no open source demand-side platform reaches parity with commercial buying systems, so the deal itself will still run through whichever buying platform you already use.
\nA private marketplace is a gated auction with a floor, a deal identifier and a defined inventory list, and curation adds a middle party that packages supply and data for a fee. Because each layer takes a share, a curated deal must earn its premium through extra outcome per impression, and the break-even is the price premium itself. Diligence questions expose the fee treatment, the publisher list, the path and the data origin. Deal quality metrics should be read per deal and per publisher, not as a blended average. A matched-market test with a holdout, a written decision rule and an independent analysis is the fair way to decide whether a deal is value or repackaged open supply.
\n1. A price premium creates a matching outcome requirement: a 60 percent higher price needs 60 percent more incremental result per impression to break even. 2. Curation adds a layer, and every layer takes a share, so ask for the publisher revenue share in writing. 3. Compare curated deals against your own site list and filters on the open exchange to learn what the curator really adds. 4. Judge quality by deal and by publisher, using the same measurement source on both sides of the comparison. 5. Run a matched-market test with a holdout and set the decision rule before launch, so the data cannot be bent to fit the story.
\nBefore you accept any curated deal, divide its price by your open exchange price, subtract one, and write that premium on the contract folder. Then run a matched-market holdout and renew only if incremental results per dollar beat that number.
\n1. Interactive Advertising Bureau Tech Lab. OpenRTB specification, current version. Primary reference for auction mechanics and deal fields. 2. Interactive Advertising Bureau Tech Lab. ads.txt and sellers.json specifications. Reference for declaring authorized sellers and tracing supply paths. 3. Brodersen, K. H., Gallusser, F., Koehler, J., Remy, N., and Scott, S. L. Inferring causal impact using Bayesian structural time-series models. Annals of Applied Statistics. 4. DuckDB documentation. Embedded analytical database reference. 5. Google. CausalImpact package documentation. Reference for the open source R package.
"},{"id":"chapter-31-native-and-content-distribution","title":"Chapter 31: Native and Content Distribution","html":"By the end of this chapter, the reader will be able to: 1. Describe the parts of a native ad unit and explain how each part is controlled by the advertiser or the publisher. 2. Compare content recommendation widgets, in-feed placements and publisher partnerships on quality, control and cost. 3. Apply FTC disclosure guidance to native ads so that a reasonable reader can tell paid content from editorial. 4. Evaluate a landing page for content traffic using speed, clarity and continuity checks. 5. Design a measurement plan that credits assisted outcomes without inflating them. 6. Build a content distribution plan with budget splits, stop rules and review dates.
\nMost brand managers I meet believe native advertising works because it hides. The ad blends in, readers do not notice it is an ad, and so they click. That belief sounds clever, and it fails in three ways.
\nFirst, it fails legally and ethically. An ad that works only when the reader is confused is an ad built on a problem waiting to happen. Regulators have said plainly that paid content must not mislead people about what it is. Second, it fails commercially. A reader who clicks believing they are headed to a news story, and lands on a sales page, leaves within seconds and remembers the brand as the one that tricked them. You paid for the click and bought a grudge. Third, it fails operationally, because the thing that makes native placements useful has nothing to do with disguise.
\nNative works when the ad is genuinely at home next to the content around it. A useful guide placed beside a story about the same problem is welcome. A sales pitch dressed up as a story is not. The difference is not the visual styling. It is whether the thing you are promoting deserves the reader's next five minutes.
\nI have watched budgets drain into recommendation widgets because the cost per click looked wonderful, only to find that the traffic never read a word. I have also seen small budgets earn real attention because the content was good and the placement was honest. This chapter helps you tell those two situations apart before you spend the money, and it ends with a plan template you can fill in this week.
\nThe ad is assembled from parts: a headline, a short description, a main image, a small logo or icon, a sponsor name, and a call to action label. The publisher's page template decides how those parts are arranged, so the same ad can look like a card in one feed and a line of text in a sidebar module. You supply the parts. The publisher supplies the frame.
\nThat division changes your creative process. Display banners are finished pictures. Native ads are kits. A long headline may be cut off by one publisher and shown in full by another, and an image with text baked in may be rejected outright. Build each kit with a headline that still works when trimmed, an image that survives a tight crop, and a description that adds a fact rather than repeating the headline.
\nNative inventory is sold through the same programmatic plumbing covered earlier in this part, and the practical lesson is simple: the more of the asset kit you provide, the more placements you qualify for. Missing assets quietly shrink your reach.
\nThree types behave differently. In-feed native appears inside a stream of content, such as a news list or a social-style scroll. Recommendation widgets appear below or beside an article under labels like \"you may also like\" or \"from around the web.\" Publisher-sponsored content is a full article or video produced with or for the publisher.
\nWidgets are the cheapest way to buy attention on the open web. They are also where most native budgets go to be disappointed.
\nThe mechanics explain why. A widget network places a module on thousands of publisher pages, advertisers bid on clicks, and publishers earn a share of each click. That rewards one thing: getting the click. Teasing headlines and tempting thumbnails perform well on click rate because the system pays for exactly that. The reader, meanwhile, is often at the bottom of the page, finished with the article and scrolling past.
\nHere is what that looks like in a plain example. Suppose a fictional outdoor gear brand, Birchwater Outfitters, spends 12,000 dollars on a widget campaign promoting a trail planning guide. The click report looks excellent, with a low cost per click and thousands of visitors. Then the brand opens its analytics and finds that most sessions lasted a few seconds, scrolled almost nowhere and never touched the guide. The clicks were real. The attention was not. The figures are illustrative.
\nQuality varies enormously inside one network. A good one shows which sites sent traffic and lets you block poor ones. A weak one gives you a blended number and a promise. Before committing budget, ask in writing whether you get a placement report by publisher domain, whether exclusions hold across all campaigns, and which third-party verification you can use for viewability and invalid traffic. A vendor that cannot show domain-level reporting is asking you to buy blind.
\nStart small. A test of a few thousand dollars with strict domain controls beats a large launch. Judge it on engaged sessions, not clicks, defined before you start (for example 30 seconds or more plus a second page view). In many accounts a few domains produce nearly all the engaged traffic. Cut the tail and move the money.
\nNetwork buying is fast, scalable and thin. Publisher partnerships are slow, limited and deep. Neither is better in general. The choice depends on what you need the content to do.
\nA network buy suits a content piece that stands alone and needs volume: a calculator, a checklist, a short video. You control targeting, creative and stop rules, and you inherit whatever inventory the network sells.
\nA publisher partnership suits a brand that needs borrowed credibility or a specific audience, such as a regional sports outlet your customers actually read. The publisher may produce a sponsored article, label it, and promote it on its pages and newsletters for a flat fee. The cost per visitor is higher. The visitor is usually warmer.
\nA fair comparison uses the same yardstick for both. Suppose a fictional home services company, Lantern Ridge Roofing, has 30,000 dollars. A network buy might send many cheap visits with a thin slice engaging. A publisher deal might send far fewer visits, with a much larger share reading to the end. Divide the budget by engaged sessions, not by visits, and the gap between the two often narrows or reverses. That is the number to bring into the planning meeting.
\nPartnership terms deserve close reading. Ask who owns the content after the campaign, whether you can reuse it on your own channels, how long it stays live, and what the publisher will report. Contract terms vary and counsel should review anything binding. This is operational guidance, not legal advice.
\nMany mature programs use both: the partnership produces the anchor piece, and the network extends it at a controlled cost.
\nDisclosure is not a legal footnote you hand to someone else. It is part of the creative brief.
\nThe Federal Trade Commission has published guidance on native advertising that centers on a simple principle: an ad should not mislead a reasonable consumer about whether it is an ad. If content looks like editorial but was paid for, the format must make that clear. The guidance looks at the overall impression, covering the headline, the layout, the label, the imagery and the context.
\nIn practice, that translates into habits you can adopt today:
\nEven when the publisher controls labeling, the advertiser is accountable for how its ad presents. Ask for screenshots of the live unit on phone and desktop at launch.
\nBe conservative where audiences are sensitive. Content aimed at children, or touching health and money decisions, calls for the clearest labels and the most careful claims. This is operational guidance only. Your counsel decides what the law requires in your situation.
\nContent traffic arrives in a different mood than search traffic. A searcher typed a need into a box. A native reader was browsing, and you interrupted with a good idea.
\nStart with continuity. The headline on the page should echo the promise on the ad. If the ad says \"A free trail planning worksheet,\" the page should open with that worksheet, not a homepage banner. Every extra click between promise and delivery loses people. Send native traffic to a dedicated page, never the homepage.
\nThen speed. Phone readers lose patience quickly. Run each content landing page through Lighthouse, a free auditing tool built into the Chrome browser developer tools and available as a command line package. It reports load behavior, layout shifts, accessibility issues and basic search hygiene. Do not chase a perfect score. Fix the warnings that cost readers: oversized images, scripts that block rendering and text that jumps as the page loads.
\nNext, clarity of the ask. A content page can carry one primary action and one soft action. The primary action might be downloading a guide in exchange for an email address. The soft action might be reading a related article.
\nBefore launch, confirm that no pop-up fights for attention in the first seconds, that tracking tags fire and respect consent choices, and that the page names its publisher.
\nFinally, check the traffic itself. Sessions that leave in under five seconds point to the source. Sessions that linger and do not act point to the page or the offer.
\nNative rarely closes the sale. It starts a conversation, adds a second touch, or reminds a reader who already knew you. If you judge it by last-click conversions alone, it will look weak, and you will cut it.
\nThe opposite error is just as costly: credit every viewer with every later sale and you will conclude it is a miracle.
\nBegin with engaged sessions, defined before the campaign starts. Then track the second visit: how many readers come back within a few weeks? A reader who returns is worth more than one who clicked once. Next, track micro-conversions, the small actions that signal interest, such as a guide download, a newsletter signup or a pricing page view. Finally, look at assisted conversions, meaning sales where a native visit appeared earlier in the path but was not the last touch.
\nMatomo, a free open source analytics platform you host yourself, can record all of this. It lets you tag native campaigns with consistent parameters, view visitor paths, set goals for micro-conversions and keep the data on your own servers. Agree on a naming pattern for source, medium and campaign before launch. One misspelled name can split a week of data into two rows no one reconciles.
\nAssisted numbers are still correlational, so test where you can. A simple test holds back one comparable region or audience segment from native spend for a few weeks, then compares downstream results with the exposed group. Treat small tests as direction, not proof. One rule: report assisted outcomes beside last-click results, label each clearly, and never add them together.
\nA plan fits on one page. If it does not, the plan is not yet clear. Here is a template with a worked illustration.
\nThe fields to fill in:
\nNow an illustration. A fictional regional fitness studio, Cedar Court Training, plans a six-week push promoting a free guide for people returning to exercise after a long break. The budget is 18,000 dollars. The plan puts roughly half into one sponsored article on a local lifestyle publisher, a third into in-feed native placements limited to a short list of approved domains, and the remainder into a widget test capped at a small ceiling. The disclosure label reads \"Sponsored\" on every unit and again above the article headline. The landing page offers the guide as the primary action and a class schedule as the soft action. An engaged session means 30 seconds or more plus a second page view. Stop rules end any domain whose engaged rate sits well below the plan average after a full week of delivery, and pause the widget test if it fails to produce a meaningful share of engaged sessions by week three.
\nReview dates are fixed at the end of each week. At the midpoint the team compares cost per engaged session across the three channels and shifts money toward the cheapest. At the end it records returning readers, guide downloads and assisted signups, plus one lesson for the next plan.
\nThe template prevents a launch with no definition of success, and it keeps the disclosure decision in the brief where it belongs.
\nNative ad unit: An ad assembled from parts such as headline, image and description, then displayed in a frame styled by the publisher's page template. Content recommendation widget: A module under or beside an article that lists links to other content, some of which are paid placements bought through a network. Sponsored content: An article, video or guide produced by or for an advertiser and published in a labeled section of a publisher's property. Engaged session: A visit that meets a threshold you define in advance, such as a minimum time on page plus a second page view or a meaningful scroll. Assisted conversion: A sale or signup where a given channel appeared earlier in the visitor's path but was not the final touch before the action.
\n1. Judging widget campaigns by click rate and cost per click, which the system is built to flatter, instead of by engaged sessions and returning readers. 2. Sending native traffic to the homepage, where the promise made in the ad is nowhere to be found and readers leave within seconds. 3. Using a banner design in a native slot, so text baked into the image gets cropped, rejected or ignored. 4. Treating disclosure as the publisher's problem, then discovering on a phone screenshot that the label is missing or too faint to read. 5. Adding assisted conversions to last-click conversions in one report, which counts the same sale twice and overstates the channel.
\n1. Write the engaged session definition in numbers before any money is spent, and judge every placement against it. 2. Demand domain-level reporting and exclusion controls from any network, and walk away from vendors that cannot provide them. 3. Build native creative as a kit with a trimmable headline, a cropping-safe image and a description that adds a new fact. 4. Place a plain label on the unit and again on the landing article, and verify both on phone and desktop at launch. 5. Tag every campaign with an agreed naming pattern in Matomo, and review the spelling before launch day.
\n1. Ask a publisher for its last three sponsored pieces and the reader response to each. That beats any rate card. 2. Test headlines against each other in small batches, and keep the honest version even when a teasing version wins on click rate. 3. Start in-feed buying on a short approved domain list, and expand only when a domain proves its engaged rate. 4. Plan a second touch for readers who download your guide, such as a short email series, because the first visit rarely closes anything. 5. Keep a file of every placement you cut and why, so the next team skips the same experiment.
\nTwo free, open source tools cover most of what this chapter needs. Matomo (GPL license) is an analytics platform you can host yourself. Use it to tag native campaigns, record visitor paths, define goals for micro-conversions and keep your visitor data under your control. Lighthouse (Apache 2.0 license) audits a landing page for load speed, layout stability, accessibility and basic search hygiene, and it runs inside the Chrome developer tools or from the command line. Run Lighthouse on each content page before launch and again after any template change. No open source tool matches a commercial recommendation network for reach, so buying still happens in marketplaces. These two protect the page and the data.
\nNative advertising succeeds when the promoted content deserves its place beside the content around it, not when the ad hides. The format is a kit of parts that each publisher frames in its own way, recommendation widgets are cheap but frequently poor in attention quality, and publisher partnerships cost more per visit but often return warmer readers. Disclosure under FTC guidance belongs in the brief, with plain labels placed before the click and again on the landing page. Content traffic needs a dedicated, fast and focused landing page. Measurement should rely on engaged sessions, returning readers, micro-conversions and assisted outcomes, reported beside last-click results and never added to them. A one-page distribution plan with stop rules and fixed review dates keeps the spend honest.
\n1. Native works through relevance and honesty, and disguise creates legal, reputational and performance risk. 2. Judge widget traffic by engaged sessions and returning readers, because click rate rewards teasing headlines. 3. Compare network buys and publisher partnerships by cost per engaged session, not cost per visit. 4. Plain, visible labels belong on the unit and on the landing article, and the advertiser answers for both. 5. Report assisted outcomes beside last-click results, labeled separately, and test with a holdout where you can.
\nEvery native buy needs three written lines, namely what counts as an engaged session, the exact disclosure label and the date you will cut weak domains, and the campaign should be held to them.
\n1. Federal Trade Commission. Native Advertising: A Guide for Businesses. Staff guidance on when paid content must be identified as advertising. 2. Federal Trade Commission. Enforcement Policy Statement on Deceptively Formatted Advertisements. Policy statement on ads that resemble editorial content. 3. IAB Tech Lab. OpenRTB Specification, Native Ads Specification. Asset definitions for native bid requests and responses. 4. Matomo documentation. Campaign tracking, goals and visitor path reports. Open source analytics, GPL license. 5. Google Chrome Developers. Lighthouse documentation. Auditing performance, accessibility and best-practice checks, Apache 2.0 license.
"},{"id":"chapter-32-contextual-targeting-in-a-privacy-first-market","title":"Chapter 32: Contextual Targeting in a Privacy-First Market","html":"By the end of this chapter, the reader will be able to: 1. Distinguish keyword, category and semantic contextual methods and match each to a campaign need. 2. Map a page or video to the IAB Content Taxonomy and explain why a shared vocabulary matters. 3. Apply brand suitability tiers to a real placement list and justify where each tier draws its line. 4. Calculate the cost of classification errors using a simple worked model. 5. Design a small in-house classifier with open tools and judge when it is good enough to trust. 6. Evaluate a contextual buy against an audience buy with a controlled test and a defensible decision rule.
\nA home-improvement client's ads once ran beside a celebrity divorce story, and I had approved the contextual package because the keyword list looked airtight. Words like \"deck,\" \"stain\" and \"weatherproof\" were everywhere in the reports. The ads ran beside articles about a celebrity's divorce settlement, a story on a sports team's \"deck\" of playoff cards, and a recap of a storm that had flattened a neighborhood. Every one of those pages contained a matching word. None of them contained a person thinking about a weekend project.
\nThat was the day I stopped believing the sentence that most of the industry repeats: contextual targeting is the simple, safe, cheap option you fall back on when you lose the cookie. It is none of those three things. It is simple to explain and hard to execute. It is safe from a privacy standpoint and exposed on a brand standpoint. It is cheap per thousand impressions and expensive when it goes wrong at volume.
\nHere is the honest framing. Contextual targeting places ads based on what a page, video or audio program is about, not on who is looking at it. Because it needs no personal profile, it holds up well under privacy rules and consent limits. That part is real. The other part is that a machine has to read millions of pages and decide what each one means, and machines misread. Every misread page is a dollar spent on the wrong moment, or worse, a brand sitting next to something its leadership would never approve.
\nThis chapter treats contextual as a craft with measurable error rates. You will learn how the methods differ, how the taxonomy and suitability tiers give you control, how to build a modest classifier yourself with free software, and how to test contextual against audience buying so the decision rests on your data instead of a vendor's slide.
\nKeyword targeting is the oldest method. You supply a list of words, and the system serves ads on pages containing them. It is fast to set up and easy to audit. It is also literal. The word \"jaguar\" does not know whether it is describing a cat, a car or a football team. Negative keywords, the words that block a page, help, but a long negative list is a sign that your positive list is doing too little work.
\nCategory targeting steps up one level. Instead of words, you pick topics from a predefined list such as automotive, personal finance or youth sports. A classifier has already assigned each page to one or more topics, so you buy the topic. This is easier to scale and less brittle than raw keywords. The weakness is that you inherit someone else's classifier, and its accuracy is rarely disclosed in a way you can verify.
\nSemantic targeting tries to understand meaning. Instead of counting words, a language model converts the page into a mathematical representation of its subject, tone and intent, then compares that to the profile of what you want. In plain terms, it asks whether this page behaves like the pages your best customers read, not just whether it shares vocabulary. It can tell a review of a pickup truck from a story about a truck crash. It can separate a page about debt relief that helps a reader from one that exploits fear.
\nNone of the three is superior in every case. Keywords suit narrow product launches where exact terms matter, such as a specific model name. Categories suit broad reach with predictable inventory. Semantic methods suit situations where intent and tone matter, like financial services, health-adjacent products or anything where the wrong mood is damaging. Many accounts run all three in separate line items so results can be compared instead of blended.
\nBuyers, publishers and technology vendors needed a common way to describe content, so the Interactive Advertising Bureau (IAB) Tech Lab maintains the IAB Content Taxonomy. It is a hierarchical list of content categories, with broad top-level topics and finer subcategories beneath them. A page about trail running shoes might sit under sports, then running, and also under shopping. Newer versions of the taxonomy also describe things beyond topic, such as the type of content and the audience it suits.
\nFirst, it makes deals portable. If your agency, your publisher and your ad server all say \"category sixteen under sports,\" you are not translating between dialects. A deal written in taxonomy terms can move between sellers with less friction.
\nThird, it gives your in-house classifier a ready topic list, so your output speaks the same language as everyone you buy from.
\nA caution: a taxonomy label is a claim about a page, not a fact. A long article on retirement that mentions a market crash in two paragraphs may be labeled finance, news and economy, and which label wins matters when you also block negative news.
\nSafety and suitability are different ideas. Safety is the floor: content nobody should fund, such as material involving illegal activity or graphic harm. Suitability is the brand-specific ceiling: content that is legal and acceptable but wrong for your brand. A beer company and a children's cereal brand share the same floor and have very different ceilings.
\nA practical way to manage this is a tier system. I recommend three tiers, written down and signed off by someone with authority before the first dollar moves.
\n| Tier | Meaning | Typical action | |---|---|---| | Block | Content that conflicts with the brand or the floor | Exclude always | | Caution | Sensitive but sometimes acceptable, such as hard news or crime coverage that is factual | Exclude by default, allow by exception with review | | Allow | Content that fits the brand | Buy freely |
\nThe middle tier is where the work lives. Take a pharmaceutical-adjacent wellness brand. A page about managing a chronic condition may be exactly the audience, yet the same page might include distressing patient stories that the brand does not want beside a cheerful offer. Caution does not mean forbidden. It means a person looks before the budget flows.
\nFor anything touching children, health, money stress, politics or religion, set a conservative default. If an audience might include minors, the tier should lean toward Block until counsel and your compliance lead agree otherwise. That is operating guidance, not legal advice; counsel decides what the rules require in your situation.
\nClassification errors come in two kinds, and they cost different things.
\nA false positive is a page the system says is relevant that is not. You pay for an impression that cannot work. A false negative is a relevant page the system missed. You lose reach you never see, which makes it a silent cost, and silent costs rarely get fixed.
\nHere is a labeled, illustrative example. Suppose a regional lawn-care brand buys two million impressions at a six dollar cost per thousand, so the spend is 12,000 dollars. A quick audit of a random sample finds that 15 percent of delivered pages were misclassified, meaning the page had nothing to do with lawns or home care. That is eighteen hundred dollars spent on impressions that had almost no chance of working.
\nNow add the suitability problem. Suppose one in 200 impressions lands on a page that falls in the Block tier. That is 10,000 impressions. The dollar waste is small, sixty dollars. The reputational cost is not. One screenshot of an ad beside the wrong story, sent to a marketing executive or posted online, can cost more than the entire campaign.
\nWhat follows from this is a simple discipline. Measure error rate as a standing metric. Audit a random sample of delivered pages each week. Record how many were wrong on topic, wrong on tone, and wrong on suitability. Any vendor who resists a sample audit is asking you to trust a number you cannot check. When the error rate is high, the cost per thousand is misleading, because the effective cost is the price divided by the share of impressions that were actually right. A six dollar buy with 15 percent waste behaves like a seven dollar buy.
\nYou do not need a large data science team to build something useful. You need patience and a few free libraries. What follows is a sketch of how a small team can do it, and a statement of its limits.
\nStart with collection. A crawler such as Scrapy can fetch pages from publishers where you have a deal or permission, and respect each site's robots rules. Keep the raw text, the URL, the date fetched and the title. Store it in a PostgreSQL database so you can query it later.
\nNext, prepare the text. spaCy handles the language work: splitting text into sentences and words, reducing words to base forms, and tagging names of people, organizations and places.
\nThen label a training set by hand. This is the part people skip, and it is the part that decides quality. Pick several hundred pages, assign them to your chosen taxonomy nodes and your suitability tiers, and have a second person label a portion independently. Where the two of you disagree, you have found ambiguity that a machine will also struggle with.
\nFinally, train a model. scikit-learn provides standard text classifiers, including simple, fast ones like logistic regression and linear support vector machines, built on word-weight features. These are easy to explain, quick to train and often good enough for a first pass. Hold back a portion of your labeled pages for testing and report two numbers: precision, the share of pages labeled relevant that truly are, and recall, the share of truly relevant pages that were found.
\nIt will not match a vendor that scans the open web at massive scale, and no open-source option reaches full parity with a large commercial contextual service across all inventory. The value of building one is different. It teaches your team what errors look like, gives you a yardstick to audit vendors, and lets you control a private allow list for the placements you care about most.
\nAudience targeting aims at people: segments built from past behavior, first-party lists, or modeled profiles. Contextual aims at moments. The tradeoffs are real, and neither wins everywhere.
\nContextual tends to do well when the purchase decision is tied to what the person is reading right now. Someone reading a comparison of ovens is in the market. Someone reading a recipe is not, yet an audience segment might label both as \"home cooks.\" Contextual also holds up when identifiers are scarce or consent is withheld.
\nAudience targeting tends to do better when the product fits a stable customer profile that is not visible in any single page: a premium subscription, a repeat-purchase product, a service with a long consideration cycle. It also supports frequency control across sites, which contextual alone cannot do, since without an identifier you cannot easily stop the same person from seeing your ad nine times on nine pages.
\nA sensible plan treats them as partners with different jobs. Use contextual for reach, discovery and moments of intent. Use audience methods, built from your own first-party consented data, for people you already know. Then measure which one actually produces the result you care about, instead of assuming.
\nA good test is boring and strict. Here is a design that fits a modest budget.
\nDefine one goal and one measure before launch. Suppose the goal is new customers and the measure is cost per new customer, tracked through your own site and a coupon or code. Choose two cells: one contextual, one audience. Match budget, creative, bid approach, dates and landing page. Only the targeting method changes.
\nSuppose each cell gets 30,000 dollars over four weeks. Add a third cell, a holdout, where a matched group of regions or a randomly chosen share of eligible traffic sees none of your ads. The holdout lets you estimate how many customers would have arrived anyway. Without it, you may credit your ads for purchases that were already coming.
\nSet the decision rule before you look at results. For example: contextual wins if its cost per new customer is at least 10 percent lower than the audience cell and the audited error rate stays under a ceiling you chose in advance. If the results sit within a narrow gap, call it a tie and consider a blended plan, because the sample may be too small to separate them.
\nRead outcomes by placement too, since a few categories often do most of the work. Cut the weak ones, then retest.
\nContextual targeting feels privacy-friendly, and in terms of personal data it is. That feeling can lull people into relaxing on content judgment, which is the wrong response. Pages where children read, where people grieve, where they seek help for money trouble or illness, deserve extra care.
\nSet conservative defaults. Exclude content aimed at children from any campaign not designed for them, and be wary of pages where age is uncertain, such as gaming forums or video sites. For sensitive categories, such as health conditions, finances under strain and religion, avoid targeting the sensitive topic directly unless counsel has reviewed the practice, and keep the creative respectful when your ad lands beside it. Operational guidance here is not legal advice; counsel decides what applicable rules require.
\nContextual targeting: Placing ads based on the subject, tone and intent of the content, not on a profile of the viewer. IAB Content Taxonomy: A hierarchical list of content categories maintained by the IAB Tech Lab, used so buyers, sellers and vendors describe pages in the same terms. Brand suitability: A brand-specific judgment of whether acceptable content fits that brand, as opposed to brand safety, which is the shared floor of content nobody should fund. Precision and recall: Precision is the share of pages labeled relevant that really are; recall is the share of truly relevant pages the system managed to find. Holdout: A matched group that sees none of the ads, used to estimate how many results would have happened without the campaign.
\n1. Trusting a keyword list without reading delivered pages, which lets literal word matches pass as relevance. 2. Accepting a vendor's tier name, such as \"moderate,\" as a specification instead of requesting sample pages at each level. 3. Judging cost per thousand without adjusting for the error rate, so a cheap buy that is wrong one time in six looks like a bargain. 4. Skipping the human-labeled training set when building a classifier, then wondering why the model mirrors its own mistakes. 5. Running a contextual test without a holdout or a pre-written decision rule, which turns the result into an argument.
\n1. Write brand suitability tiers down per category and tone, and get sign-off from a decision maker before launch. 2. Audit a random sample of delivered pages every week and log topic, tone and suitability errors separately. 3. Run keyword, category and semantic methods in separate line items so you can see which one earns its budget. 4. Use the IAB Content Taxonomy labels in your deal terms and internal reports so every partner speaks the same vocabulary. 5. Keep an escalation path with a named person who can remove a placement within one hour.
\n1. Ask any contextual vendor which taxonomy version and which node a campaign targeted; a clear answer signals a disciplined seller. 2. Review the first 50 delivered URLs by hand within three days of launch, before the budget scales. 3. Pair a contextual buy with a narrow private allow list of publishers you have checked yourself. 4. Treat breaking news as its own risk class; classifiers lag events by hours, and those hours can be costly. 5. When contextual and audience results tie, split the budget and keep testing instead of declaring a winner.
\nA small in-house contextual program can run entirely on free, open-source software. Scrapy (BSD license) collects page text from publishers you have permission to read. spaCy (MIT license) splits the text into sentences, words and named entities. scikit-learn (BSD license) trains and scores the classifier, including precision and recall on held-back pages. PostgreSQL (PostgreSQL license) stores pages, labels, tiers and audit results so you can query them. These tools will not match a large commercial service that scans the whole web, but they give you control and a yardstick.
\nContextual targeting is not a fallback. It places ads by the meaning of content, using keyword, category or semantic methods, each with its own strengths. The IAB Content Taxonomy gives the market a shared vocabulary, and written suitability tiers turn brand judgment into rules a team can follow. Classification errors cost money in wasted impressions, missed reach and brand exposure, so error rate should be audited every week. A modest classifier built with open tools helps a team understand and verify the work, though it cannot match large commercial scale. Contextual and audience buying fill different roles, and a controlled test with a holdout and a pre-set decision rule shows which one deserves budget. Conservative defaults protect families and sensitive audiences.
\n1. Keyword, category and semantic methods suit different jobs, and running them as separate line items shows which one works. 2. The IAB Content Taxonomy lets you write, compare and audit deals in a shared vocabulary. 3. Brand suitability tiers need to be written, specific and approved before spending begins. 4. Misclassification raises your effective cost, so a weekly sample audit belongs in every contextual plan. 5. A fair test of contextual against audience buying needs matched cells, a holdout and a decision rule set in advance.
\nPull 50 delivered URLs before your next contextual buy at random, read them, and record how many were wrong on topic, tone or suitability. Then price the buy by what survives the audit, not by the rate card.
\n1. Interactive Advertising Bureau Tech Lab. Content Taxonomy, current version. Hierarchical content categories for describing pages and media. 2. Interactive Advertising Bureau Tech Lab. OpenRTB specification, current version. Fields used to pass content and category signals in programmatic auctions. 3. Honnibal, M., Montani, I. spaCy documentation. Natural language processing library, MIT licensed. 4. Pedregosa, F. et al. Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research. BSD licensed software and published methodology. 5. Scrapy project documentation. Web crawling framework, BSD licensed. 6. Manning, C., Raghavan, P., Schutze, H. Introduction to Information Retrieval. Cambridge University Press. Text classification, precision and recall.
"},{"id":"chapter-33-retargeting-discipline","title":"Chapter 33: Retargeting Discipline","html":"By the end of this chapter, the reader will be able to: 1. Explain why retargeting reports tend to overstate its contribution and identify which audiences drive that distortion. 2. Design audience segments, recency windows and suppression lists that match shopper intent. 3. Apply frequency caps and sequential messaging that remind a shopper without wearing out the welcome. 4. Construct a holdout test and calculate true incremental cost from its results. 5. Evaluate consent, privacy and family-safety constraints that limit who may be retargeted. 6. Decide when a retargeting program should be reduced, rebuilt or switched off.
\nHere is a question I put to brand managers whenever retargeting comes up. Of the shoppers your retargeting campaign says it converted last month, how many would have bought anyway?
\nMost people answer with a shrug, or with the dashboard number. Both are wrong in the same way. The dashboard counts a win whenever a shopper saw your ad and then bought. It cannot see the shoppers already walking to the register when the ad appeared.
\nThat is the uncomfortable fact about retargeting. You are not choosing strangers. You are choosing people who already visited your site, already looked at a product, and in many cases already decided. Showing an ad to someone who was going to buy is like paying a ticket scalper to stand beside a fan who already holds a seat. The scalper will happily report that every fan he stood next to got into the stadium.
\nI am not telling you to abandon retargeting. In some accounts I have reviewed it is among the most efficient lines in the plan, and in others it quietly consumes a large share of budget while contributing almost nothing. The difference is discipline. Discipline means you know which shoppers need a nudge and which do not, you stop spending on the second group, and you prove the difference with a test instead of a report.
\nChapter 1 explained why a trackable result is not the same as a measured one. This chapter applies that argument to the channel where the gap is widest, and ends with a holdout design you can run next month.
\nRetargeting means showing ads to people who have already interacted with your brand, usually identified by a browser cookie, a mobile identifier or a hashed email address.
\nBecause these shoppers are far more likely to buy than a cold audience, cost per acquisition on a retargeting line often appears several times better than on any other line in the plan.
\nHere is the trap. Part of that efficiency belongs to the shopper, not the ad. Intent was created earlier, by search, a friend, a prospecting ad or the product itself. Retargeting harvests demand that other work planted. The harvest is real, but crediting the harvester for the whole field is bookkeeping, not measurement.
\nSo the useful question changes: not \"how many buyers did retargeting touch?\" but \"how many extra buyers exist because retargeting ran?\" The first number is flattering. The second takes a test and is often humbling.
\nA single audience called \"all site visitors\" is the most common retargeting setup and the weakest. A person who landed on your homepage by accident and a person who sat through your shipping page with three items in a cart are not the same prospect. Treating them alike wastes money on the first and underserves the second.
\nStart with a short ladder of intent, ordered from coolest to warmest:
\n| Segment | Typical signal | Starting posture | |---|---|---| | Shallow visitors | One page, short visit | Lowest priority, often excluded | | Browsers | Product or category pages, repeat visits | Light reminders, tight window | | Cart and checkout abandoners | Item added, purchase not finished | Highest bid, shortest window | | Lapsed customers | Bought before, quiet for a long stretch | Replenishment or return offers |
\nFour segments are plenty. Build only what you can feed with distinct messages.
\nSegment by behavior, not by guesses about who the person is. A cart abandoner may have been distracted by a phone call, shocked by shipping, or comparing a rival store. Address the likely barrier, and confirm it with testing.
\nIntent fades. A shopper who looked at a pair of running shoes this morning is a different prospect from the same shopper six weeks later, who has probably bought elsewhere, changed plans or forgotten you.
\nA recency window is the number of days after an interaction during which a person stays in the audience. It is one of the cheapest levers you own and one of the most neglected. Many accounts leave the window at the platform default, often 30 days or longer for everyone, and then wonder why the spend feels sticky.
\nA sensible starting rule is to match the window to your purchase cycle, with warmer segments getting shorter windows. Cart abandoners might stay in for a few days to a week, because the decision is nearly made and either completes quickly or stalls. Browsers might get two weeks. Lapsed customers follow the natural replenishment rhythm of the product: a consumable that runs out in 30 days deserves a reminder near day 25, not a daily nag from day two.
\nThen check your own data: how many days pass between a first visit and a purchase for people who eventually bought? In many shops most purchases land in a short span, and spending in the long tail tends to show poor returns once tested. The holdout later in this chapter can be run by window too.
\nSuppression means deliberately excluding people from an audience. Of all the retargeting disciplines, it returns the most for the least effort.
\nThe first and most obvious exclusion is recent buyers. Someone who bought a blender yesterday does not need to see that blender for the next three weeks. Yet I still see accounts spending real money chasing people who finished checkout days ago, because the conversion pixel and the audience list sync on different schedules. Tie your purchase event to your audience tool so a buyer drops out within hours, not days.
\nThe second is shallow visitors who match nothing of value. Bounces, bot traffic, employees and agency staff inflate the pool and distort frequency.
\nThe third is people who asked not to hear from you. Someone who withdrew consent, opted out or complained must leave every audience across every channel. This is not optional courtesy. It is a legal and ethical line, and the section on privacy returns to it.
\nFrequency is how many times one person sees your ad in a period. Retargeting is unusually prone to excess because the audience is small and the bids are high, so a handful of people can absorb enormous impression counts.
\nSet caps and then watch the distribution, not just the average. An average frequency of six can hide a few people at 40. Those people are not more likely to buy. They are the ones most likely to feel followed and to block your ads. A good starting cap for warm segments is a few impressions per day with a weekly ceiling, tightened for shallow segments. Treat these as starting points to test, not universal truths.
\nSequential messaging solves part of the problem. Instead of repeating one ad, you tell a short story in order. Suppose a shopper views a product page. The first ad shows the product and one clear reason to pick it. If no return visit follows in a few days, the second ad answers an objection: delivery speed, returns, sizing. A third ad, only if needed, offers the final nudge, whether that is a modest incentive or a deadline that is genuine. After that, the shopper leaves the audience and gets a rest.
\nA sequence also shows you which step does the work, because you can compare conversion at each stage.
\nA caution about incentives. Offering a discount to every cart abandoner trains shoppers to abandon carts on purpose, and it hands margin to people who would have paid full price. Reserve incentives for the later step of a sequence and for segments where the holdout shows real lift.
\nDynamic creative means a template that automatically fills in the product a shopper looked at, usually pulled from your catalog. It is a useful tool, and it is also where retargeting most often becomes tedious. A feed of the exact item you already viewed, shown 15 times, is not persuasion. It is an echo.
\nWork at the concept level instead. A concept is the reason to buy, not the product tile. For one running shoe, the concepts might be comfort for long distances, a return policy that removes the risk, or a limited colorway that sells out. Build three or four concepts, then let the dynamic template drop the correct product into each one. You get relevance from the feed and variety from the concepts.
\nTest concepts against each other rather than against the old default. Small retargeting pools rarely support many variants.
\nRespect the shopper's context too. A banner for a gift, a medical item or a financial product that follows someone across a shared family tablet can embarrass them. Set a conservative default: no dynamic product display for sensitive categories, and none at all on content aimed at children.
\nRetargeting depends on recognizing a person across sites, which is exactly what privacy rules and browsers increasingly restrict. Several browsers limit or block third-party cookies by default, and mobile systems ask users whether an app may track them. Your retargeting pool today is smaller than the raw count of site visits suggests, and it will likely keep shrinking.
\nConsent is the other pillar. In regions with strict privacy law, you generally need a person's permission before placing the identifiers that power retargeting, and that permission must be real: clearly explained, freely given and easy to withdraw. A banner that nudges everyone toward \"accept\" while hiding the \"decline\" button is a liability, not a solution. Where the rules apply, build your audience only from people who agreed, and remove people who change their mind. This is operational guidance only. Your counsel decides what the law requires in your markets.
\nUse a consent manager that records each person's choice and blocks tags until permission is given, and keep that log. Never build audiences from pages that reveal sensitive facts about a person, such as health, finances or children's activities. A smaller, consent-based audience is usually a more engaged one anyway.
\nCannibalization happens when retargeting takes credit for sales that other channels or the shopper's own intent would have delivered. It is common, and it is invisible in platform reports, because those reports only count what happened after an ad was shown.
\nThe cure is a holdout test: randomly keep a slice of your retargeting audience from seeing ads, then compare their purchases against the group that did see ads. Because the split is random, any gap in buying between the groups is caused by the ads.
\nHere is a design you can run. Take one segment, say cart abandoners. Randomly assign 90 percent to the test group and 10 percent to the holdout, before any ad is served. Serve normal ads to the test group. For the holdout, serve nothing, or serve a neutral public service ad if your platform needs a placeholder for fairness in the auction. Run it long enough to cover at least one full purchase cycle, and do not change bids, creative or caps midway. Record purchases for both groups over the same period, including purchases completed through any channel.
\nAn illustrative example, with round numbers. Suppose the test group holds 900,000 people and the holdout holds 100,000. In the test group, 28,800 people buy, a rate of 3.2 percent. The holdout buys at a rate of 2.0 percent. The gap is 1.2 points. Applied to 900,000 people, that is about 10,800 extra purchases caused by the ads. If you spent 972,000 dollars and the platform credited 54,000 conversions, the dashboard shows an 18 dollar cost per purchase. The true incremental cost is 972,000 dollars over 10,800 purchases, or 90 dollars each.
\nSame spend, same campaign, and a cost per result five times higher. Whether 90 dollars is acceptable depends on your margin and customer value. You now decide with the real number.
\nWith small samples a gap this size can be noise, so check it against chance. If you cannot run a clean user-level split, a geographic holdout or a time series analysis can approximate it, with more caution.
\nSome retargeting should end. The holdout will tell you when.
\nStop or sharply cut a segment when its incremental cost exceeds the contribution margin of the sale, and no later repeat purchase makes up the gap. Stop when the holdout gap disappears into noise across two clean tests. Stop when consent rates fall so far that the audience cannot hold a test.
\nStopping is rarely all or nothing. A common result is strong lift for cart abandoners, weak lift for browsers and none for shallow visitors. Fund the first, trim the second, drop the third, and move the freed budget where tests say there is room.
\nRun the test again every quarter. Audiences age, rivals change their bids and your site changes. Retargeting is conditioning, like a training camp: skip it for a season and you notice in the standings.
\nRetargeting: Showing ads to people who have already interacted with a brand, such as by visiting its site or abandoning a cart.
\nHoldout group: A randomly selected slice of an audience that is deliberately not shown ads, used as the baseline to measure what the ads caused.
\nIncrementality: The extra conversions that happened only because the ads ran, as opposed to conversions that would have happened anyway.
\nSuppression list: A list of people excluded from an audience, such as recent buyers or people who withdrew consent.
\nRecency window: The number of days after an interaction during which a person remains eligible to see retargeting ads.
\n1. Judging retargeting by the platform's conversion count, which includes buyers who were already on their way and makes the channel look several times better than it is. 2. Leaving recent buyers in the audience for days because the purchase event and the audience list do not sync, so the brand pays to advertise to customers who just checked out. 3. Using one broad \"all visitors\" audience with one default window, which spends the same on a bounce as on an abandoned cart. 4. Letting the average frequency hide a small group of shoppers seeing the same ad dozens of times, which drives blocking and annoyance. 5. Offering a discount in the first retargeting ad to everyone, which teaches shoppers to abandon carts and gives away margin on sales that were already certain.
\n1. Run a holdout test on each major segment before setting budgets, and repeat it every quarter so decisions rest on current lift instead of old reports. 2. Match recency windows to your actual purchase cycle, with the shortest windows on the warmest segments, and test the tail to find where value ends. 3. Sync purchase events to audience lists so buyers leave within hours, and audit a sample of recent orders each month to confirm it works. 4. Write three or four reasons to buy and let dynamic templates insert the product, so shoppers see variety instead of an echo of the item they already viewed. 5. Build audiences only from consented visitors, keep a record of each choice, and remove withdrawn permissions from every channel at once.
\n1. Look at the frequency distribution, not the mean. The top few percent of viewers usually tell you where your complaints will come from. 2. Compare display against your email program. A cart reminder by email costs little, and a holdout often shows display adds less on top than expected. 3. Run holdouts by segment. A blended result can hide a strong cart abandoner lift under a weak browser result, and you will cut the wrong budget. 4. Strip internal traffic and agency staff from every pool. A handful of employees clicking daily can make a segment look healthier than it is. 5. If the audience is too small to test, say so and lower the budget. A line you cannot measure should not receive money you cannot defend.
\nThree free, open source tools cover the core of this chapter. Klaro (BSD license) is a consent manager that records each visitor's choice and blocks tracking tags until permission is given, which keeps your retargeting audiences consent-based. Matomo (GPL license) is self-hosted web analytics that lets you see segment behavior, recency patterns and purchase timing on data you own. CausalImpact (Apache 2.0 license) is a statistical package that estimates the effect of a campaign from a time series, which is useful when you cannot run a clean user-level holdout. A full demand-side platform with retargeting built in has no open source option that reaches parity, so plan to buy that media through your chosen ad marketplace and use these tools to govern and measure it.
\nRetargeting reaches people who already know your brand, so its reported efficiency is inflated by buyers who would have purchased anyway. Discipline starts with segments built around intent, recency windows matched to the purchase cycle, and suppression lists that remove buyers, internal traffic and anyone who withdrew consent. Frequency caps and short sequential stories keep the program a reminder rather than a pursuit, and dynamic creative works best when it fills concepts instead of repeating a product tile. Consent rules and browser limits shrink the audience, and a conservative default protects sensitive categories and children. A randomized holdout turns the question of cannibalization into a number, and the same test tells you where to trim and when to stop.
\n1. Retargeting harvests demand created elsewhere, so a platform's conversion count overstates what the ads caused. 2. Segment by behavior and keep only as many segments as you can feed with distinct messages. 3. Suppression of recent buyers, internal traffic and withdrawn consent is the cheapest improvement available. 4. Short sequences of different messages, with caps checked by distribution, protect the shopper and the brand. 5. A randomized holdout by segment gives you the incremental cost, and it tells you when to cut or end a line.
\nBefore next month's budget meeting, carve a random 10 percent holdout out of your cart abandoner audience and run it for one full purchase cycle. Fund retargeting at the incremental cost that test reveals, not at the cost the dashboard reports.
\n1. Brodersen, K.H., Gallusser, F., Koehler, J., Remy, N., and Scott, S.L. Inferring causal impact using Bayesian structural time-series models. Annals of Applied Statistics. Published methodology behind the CausalImpact package. 2. Interactive Advertising Bureau Tech Lab. OpenRTB specification, current version. Reference for how programmatic auctions pass audience and bid data. 3. Klaro. Consent manager documentation and project repository. Reference for recording and honoring visitor consent choices. 4. Matomo. Self-hosted analytics documentation. Reference for segments, funnels and consent-aware tracking. 5. General Data Protection Regulation, provisions on conditions for consent. Reference for the standard of freely given, withdrawable permission.
"},{"id":"chapter-34-brand-safety-suitability-and-invalid-traffic","title":"Chapter 34: Brand Safety, Suitability, and Invalid Traffic","html":"By the end of this chapter, the reader will be able to: 1. Define brand safety, brand suitability, and made-for-advertising inventory, and separate each from the others. 2. Classify invalid traffic as general or sophisticated using the Media Rating Council framing. 3. Compare verification vendor reports with log-level review and decide what each can and cannot prove. 4. Calculate the cost of over-blocking for a campaign using a simple worked model. 5. Design a written threshold profile that reflects a family-first default. 6. Evaluate a campaign against a quarterly control checklist and assign an owner to every gap.
\nBrand safety tooling is a tax, and plenty of buyers resent paying it. You pay it on every thousand impressions, and what you get back is a report that says the tax worked. Most brand managers have never seen a client lose a quarter because an ad ran next to a bad page. Plenty have watched a campaign starve because the blocklist grew until nothing was left to buy. A skeptic would say the whole discipline is insurance sold by the people who profit when you feel afraid.
\nThe skeptic is partly right. Fear sells filters, and filters have a price in reach, in cost per result, and in the quality of the audience you end up with. A brand that blocks every news page because one headline was ugly has not protected anything. It has bought a smaller, more expensive campaign.
\nHere is where the skeptic goes wrong. The risk is real, it is just not evenly spread. Some of it is reputational, where your ad lands beside content that contradicts what you stand for. Some of it is financial, where your money pays for bots, scripts, and sites built only to collect ad revenue. The first risk needs judgment. The second needs arithmetic. Treating them as one problem called \"safety\" is how brands end up overpaying for both.
\nThink of a stadium. You do not ban fans because someone once threw a bottle. You set rules at the gate, you train the staff, you watch the cameras, and you review the incident log afterward. That is the posture of this chapter: written thresholds, measured outcomes, and a refusal to pay for protection you cannot verify. Operational guidance only here; your counsel and your contracts decide the legal questions.
\nBrand safety is the floor. It covers content almost no legitimate brand wants to sit beside: material that is illegal, graphic, hateful, or built to exploit. Because the floor applies to everyone, you can set it once, write it down, and apply it to every campaign without a long debate.
\nSuitability is a dial, and each brand turns it differently. A funeral home, a snack company, a bank, and a children's toy maker all have different comfort levels with the same news story or the same comedy clip. Nothing about the page is unsafe in the universal sense. It is simply wrong for you, or right for you at one setting and wrong at another.
\nThe practical difference is who decides. The floor can be delegated to a standard list. The dial cannot. If your team never sits down and writes what your brand will and will not appear beside, the dial gets set by default, usually by whoever configured the platform first.
\nA useful way to write the dial is in three tiers per content topic. Avoid means never buy. Limit means buy only in contexts you have reviewed, such as a named list of publishers. Accept means buy freely. Most topics land in Limit, which is the honest answer: context matters more than topic. A story about a medical condition on a respected health publisher is a different placement than the same words in a comment thread on a page full of ads.
\nTopic labels come from classifiers, and classifiers err in both directions. Have someone read actual pages, not only category names.
\nMade-for-advertising inventory, often shortened to MFA, describes sites built primarily to host ad slots rather than to serve readers. They tend to have thin or recycled content, heavy ad density, slideshow formats that force many page loads, and traffic bought from cheap sources to fill the slots. The visitors are often real humans, which is exactly why MFA is slippery. It may pass an invalid traffic check and still be a poor place for your money.
\nSo treat MFA as a quality question, not a fraud question. Nothing on the page is necessarily illegal. The issue is that attention there is shallow and your brand is sharing a page with a wall of other ads.
\nSignals worth reviewing in your own logs: unusually high impressions per visit, very short time between page load and the next ad request, domains whose traffic arrives mostly through a single reseller, and sites whose share of your spend grew fast while conversions did not. None of these proves anything alone. Two or three together justify putting a domain on a watch list and reviewing it by hand.
\nA rule I use: if you cannot describe in one sentence why a reader would visit the site on purpose, hold your spend until you can. That is not scientific. It is a quick filter that catches many of the worst cases before any report does.
\nThe Media Rating Council separates invalid traffic into two groups, and the split is useful because the groups need different tools.
\nGeneral invalid traffic, or GIVT, is traffic that can be caught by routine filtering: declared bots and crawlers, known data center addresses, and other sources on maintained lists. If a visit announces itself as a search engine crawler, nobody is fooled. Filtering GIVT is table stakes, and most reputable platforms remove much of it before you are billed. Your job is to confirm they did.
\nSophisticated invalid traffic, or SIVT, is built to pass as human. Examples include hijacked devices running ads in the background, malware that loads ads nobody sees, falsified site or app identity so that low-value inventory is sold as premium, and click or conversion manipulation. Catching it takes analysis across many data points, not a lookup table. You will not find SIVT by reading one line of a log. You find it by comparing patterns across thousands of lines.
\nWhy does the split matter to you? Because the cost of each is different. GIVT is mostly a billing and reporting issue: you should not pay for it and your numbers should exclude it. SIVT is a decision issue: it can distort your results so that a bad source looks like a good one. If a fraudulent site produces cheap clicks and a few fake conversions, your optimization tools will happily send it more budget.
\nOne more practical point. Invalid traffic rates are reported after the fact, and credits for invalid impressions depend on your contract and on the platform's own policy. Read those terms before you spend, not after. I am not giving contract advice here, only telling you where to look.
\nVerification vendors place measurement code or use pre-bid signals to classify impressions for safety, suitability, viewability, and invalid traffic. They are useful. They see across many campaigns and many publishers, which no single brand can. They also have limits you should understand.
\nA vendor classifies; it does not explain. When a report says a share of your impressions were blocked or flagged, you usually cannot see which pages, which sellers, or which supply paths made up that share. You are trusting a methodology. Methodologies differ, and so do the results from two vendors measuring the same campaign. That is not a scandal. It is what happens when different classifiers make different judgment calls.
\nLog-level review is the counterweight. Log-level data means one row per impression or event, with fields such as domain or app, seller identifier, timestamp, price paid, and outcome. If your platform exports it, you own a record you can query yourself. Where it does not, ask for the closest available substitute, such as placement-level reports with seller detail, and note that gap in your plan.
\nWhat can you do with the rows? Compare cost per result by domain. Find sellers whose traffic spikes at odd hours. Check whether a flagged domain from the vendor shows up in your top spend. Count impressions per user per day and look for impossible values. Look for domains that appear in your logs but nowhere in your own approved list.
\nThe two approaches answer different questions. Vendors tell you what a trained classifier thinks of the inventory. Logs tell you where your money actually went. A mature program uses both and treats disagreement between them as the interesting part. When the vendor says clean and the logs show a domain with huge spend and zero conversions, you have found something worth a phone call.
\nEvery block removes inventory, and removing inventory raises the price of what is left. This is the part the safety conversation tends to skip.
\nConsider an illustrative case. A fictional footwear brand, Alder Quay Outfitters, plans a display campaign with a budget of 60,000 dollars. At a baseline of six dollars per thousand impressions, that buys 10 million impressions. The team adds a long keyword blocklist covering anything touching crime, politics, health, or tragedy. The eligible pool shrinks, competition for the remaining pages rises, and suppose the effective price climbs to eight dollars per thousand. The same budget now buys 7.5 million impressions. That is a 25 percent loss in reach, paid for in exchange for protection against pages that were mostly harmless.
\nThere is a second cost. Keyword blocking is blunt. A blocklist term like \"shooting\" removes a local high school basketball recap along with the crime story. A brand that wants to reach sports fans just blocked a chunk of sports content. Over time the campaign drifts toward a narrow set of safe, crowded, expensive pages, and every rival is bidding on the same ones.
\nA decision rule helps: before adding any broad term to a blocklist, pull a sample of 50 recent pages that match it and read them. If most are harmless, do not block the term. Use a narrower rule, a named list of publishers, or a limit tier instead.
\nMeasure the effect, too. Compare reach and cost per result before and after any large list change. If the cost per result worsens and the share of flagged impressions barely moved, the block bought nothing. Protection has to be bought like anything else, at a price you can defend.
\nSuppose Alder Quay exports a week of impression rows and loads them into a local analytical database. The first query groups spend by domain and sorts descending. The second joins that result to conversion events and calculates cost per result for each domain with at least 500 impressions. The third counts impressions per device identifier per day.
\nThe findings are plain. Twelve domains account for 30 percent of spend. Three of them have impression counts far above what the traffic volume would justify and no conversions at all. One seller identifier appears on a large share of the suspicious rows. A handful of device identifiers show hundreds of impressions in a day.
\nNone of that is proof of fraud. It is a shortlist. The team pulls the three domains, reads the pages, and finds slideshow layouts with repeated ad reloads. Those go on the watch list and then the blocklist. The seller identifier goes to the buying platform with a question and a request for clarification. The device pattern goes to the vendor, who confirms some of it as invalid traffic and not the rest.
\nNotice the order. Logs found the questions. The vendor helped answer part of them. Human reading closed the rest. If the team had only looked at the vendor dashboard, it would have seen a single number and moved on.
\nThe platform defaults reflect a general audience. Your brand may have a narrower one, and where children, teens, or families are part of your audience, you should set the dial more conservatively than any default.
\nA family-first profile starts with a short written standard, something like: we do not want our advertising near content that is sexual, graphic, exploitative, or aimed at children when our product is not for children. Then it translates that standard into tiers, lists, and checks.
\nThree practical rules. First, if the product is not meant for minors, avoid inventory known to skew toward children, such as kids' games and video channels. Advertising to children carries legal and platform restrictions that vary by place, so counsel should review any plan that might touch that audience. Second, if the product is meant for families, use named publisher lists for the most sensitive placements, not open exchange buying. Third, treat gambling, alcohol, and other age-restricted categories as the same problem in reverse: keep your ads away from environments where minors are likely.
\nThresholds belong in writing, signed by someone with authority, and revisited on a schedule. When a news event changes what is acceptable overnight, the written profile gives your team a basis for a fast decision, not a panicked one. It also gives agencies and partners a standard to follow and be held to.
\nRun this at the start of each quarter and after any large change in spend or platform.
\n1. Is the written safety floor current, and does every campaign apply it? 2. Is the suitability profile written with avoid, limit, and accept tiers, and does it have a named owner? 3. Does each platform export domain or app plus seller detail, and is it being stored? 4. Has someone read a sample of pages from the top 20 spend domains this quarter? 5. Are invalid traffic credits and reporting terms understood for each buying source? 6. Is the vendor report compared against your own logs, with disagreements listed? 7. Have the last three large blocklist changes been measured for reach and cost effect? 8. Is MFA review on the calendar, with a watch list and a decision date? 9. Does the family-first standard cover audience, placement, and creative? 10. Is there a record of what changed and why?
\nAny item answered no gets an owner and a date. The list is deliberately short. A checklist nobody completes is worse than none.
\nBrand safety: The practice of keeping ads away from content that is nearly universally unacceptable for advertisers, such as illegal or graphic material. Brand suitability: The brand-specific judgment of which acceptable content fits a particular brand, audience, and message. Made-for-advertising inventory: Sites or apps built mainly to host ad slots, often with thin content and heavy ad density, regardless of whether visitors are human. General invalid traffic (GIVT): Non-human or non-qualifying traffic that routine filtering, such as bot lists and data center lists, can identify. Sophisticated invalid traffic (SIVT): Invalid traffic designed to look human, needing multi-point analysis to detect, such as hijacked devices or falsified identity.
\n1. Treating safety and suitability as the same setting, so a universal floor is applied where a brand-specific judgment was needed. 2. Adding broad keywords to a blocklist without reading sample pages, which removes good inventory and raises the price of the rest. 3. Accepting a single vendor percentage as proof, without asking which domains, sellers, and methods sit behind it. 4. Reading only the GIVT line in a report and assuming the sophisticated cases were handled too. 5. Buying a \"clean inventory\" promise without checking your own logs for domains with high spend and no outcomes.
\n1. Write the safety floor once, apply it to every campaign, and store it where the whole team can see it. 2. Write the suitability dial as avoid, limit, and accept tiers, reviewed against real pages by a named owner. 3. Export and keep log-level data with domain or app and seller detail wherever the platform allows. 4. Measure reach and cost per result before and after every major list change. 5. Review the written family-first standard each quarter and after any major news event affecting your categories.
\n1. Sort your own spend by domain every month; the top 20 rows usually teach more than a dashboard. 2. Pair every vendor flag with a log lookup, because agreement builds confidence and disagreement shows where to dig. 3. Prefer a named publisher list over a long blocklist when a topic matters a great deal and the pool is small enough to manage. 4. Ask each buying source how and when invalid traffic credits are issued before you commit budget. 5. Keep a one-page change log for lists and thresholds so decisions survive staff turnover.
\nPython and DuckDB (MIT license) can carry the whole log review on a laptop. Python cleans and joins exported impression files, and DuckDB runs fast analytical queries against those files directly, which makes the spend by domain, cost per result, and impressions per device checks simple to repeat. Metabase (AGPL license) sits on top and turns saved queries into dashboards your team can open without writing code, such as a weekly domain spend table with a watch list column. Be plain with yourself about scope: these tools review and report. They do not replace a verification vendor's classification at scale, and no open source option reaches parity there.
\nBrand safety is a universal floor, brand suitability is a dial each brand sets for itself, and invalid traffic is a financial problem that comes in a routine form and a sophisticated one. Made-for-advertising inventory sits between them as a quality concern. Vendors classify at scale but explain little, while log-level review shows where money actually went, so a sound program uses both and studies where they disagree. Over-blocking has a real price in reach and cost, which is why list changes should be measured. A written family-first standard, a short quarterly checklist, and a simple change log keep the program honest.
\n1. Safety is a single floor for everyone, while suitability is a brand-specific dial that must be written down. 2. Made-for-advertising inventory is a quality question that your own logs can help answer. 3. GIVT is caught by routine filtering, and SIVT needs analysis across many data points. 4. Vendor reports classify, and log-level review explains, so use both. 5. Every broad block has a cost in reach and price, so sample pages and measure before and after.
\nThis week, export one month of impression rows with domain and seller detail, sort spend by domain, and read the pages behind your top 20 before you change a single blocklist term.
\n1. Media Rating Council. Invalid Traffic Detection and Filtration Guidelines Addendum. General and sophisticated invalid traffic definitions. 2. IAB Tech Lab. Content Taxonomy. Standard content category definitions used for classification. 3. IAB Tech Lab. ads.txt and sellers.json specifications. Seller authorization and supply path disclosure. 4. World Federation of Advertisers. Global Alliance for Responsible Media Brand Safety Floor and Suitability Framework. Archived reference; the initiative was discontinued in 2024. 5. DuckDB Documentation. Querying CSV and Parquet files directly. Open source analytical database, MIT license.
"}]},{"title":"Part VI: Owned, Affiliate, and Commerce Channels","items":[{"id":"chapter-35-email-as-the-highest-control-channel","title":"Chapter 35: Email as the Highest-Control Channel","html":"By the end of this chapter, the reader will be able to: 1. Explain in plain language what SPF, DKIM and DMARC records do and why mailbox providers check them. 2. Design a permission-based list-building system with a clear value exchange and a confirmed opt-in. 3. Build segments from behavior and recency rather than from guesses about demographics. 4. Map a welcome, abandonment, post-purchase and win-back flow and assign each one a measurable job. 5. Evaluate email results while accounting for the distortion that mailbox privacy features add to open data. 6. Compare self-hosting a mail server with renting sending infrastructure and choose based on honest operational capacity.
\nHere is a scene I have watched play out in several forms. Illustrative composite: a regional outdoor gear brand, call it Marrow Creek Supply, has a list of about 40,000 addresses collected over several years from checkout, contests and a trade show sign-up sheet. The marketing lead is sensible. Social reach has been shrinking, so she decides to lean on the one channel the brand owns. She schedules a big promotion to the whole list on a single Thursday, from a new sending address, with a heavy image banner.
\nWithin a few hours, the numbers look strange. Delivery is fine on paper. Almost nothing is opening. By Friday the replies stop, and the promotion earns a fraction of what a small paid test would have. Nothing she did was careless. The problem is that email's reputation as the safe, owned, always-available channel hides a hard truth: you own the list, but you do not own the inbox.
\nA mailbox provider decides whether your message lands in the inbox, the spam folder or nowhere. It decides based on your history, your authentication, your complaint rate and how recipients behave toward you. The old trade show names were cold. The new address had no history. The sudden volume looked like a spammer waking up.
\nSo let me challenge the assumption straight away. Email is not the highest-control channel because it is easy. It is the highest-control channel because the levers that decide its results are ones you can actually pull: who you invite, what you promise, how you prove your identity, how often you send and whom you stop sending to. Those levers are unglamorous. They are also the whole game. This chapter is about pulling them in the right order.
\nA list is only as valuable as the share of it that wants to hear from you. A large list of reluctant people is worse than a small list of willing ones, because reluctant people click the spam button, and the spam button is the loudest signal a mailbox provider receives.
\nStart with a value exchange that is stated plainly at the point of capture. A form that says \"Join our newsletter\" asks for trust and offers nothing. A form that says \"Get the trail conditions digest every Friday, plus first access to restocks\" tells the person what they will receive and how often. Specificity filters out people who would never open the mail anyway.
\nAvoid pre-checked boxes. If someone buys a pair of boots, that is permission to send the receipt and shipping updates. It is not automatically permission to send weekly promotions, and under some state and international rules it is not enough at all. Operational guidance only here: your counsel decides what consent standard applies to your audiences.
\nUse confirmed opt-in whenever the list will carry real weight. Confirmed opt-in means a new subscriber receives one message with a link and is added only after clicking. It shrinks the list, but it screens out typos, bots and prank sign-ups before they become bounces, and it leaves a timestamped record of who agreed.
\nNever buy, rent or scrape a list. The recipients never met you, and some addresses are spam traps, which mailbox providers keep specifically to catch careless collectors. One hit can damage a sending domain for weeks.
\nDeliverability is the share of your messages that reach the inbox, and it depends on three things: who you prove you are, how recipients treat you, and how consistently you behave.
\nStart with identity. Three records sit in your domain's settings, and each answers a simple question.
\nSPF answers: which servers are allowed to send mail using my domain name? Think of a guest list at a venue door. The record lists the approved senders, and a message from an unlisted server looks suspicious.
\nDKIM answers: was this message altered after I sent it, and did it really come from me? Your sending system adds a digital signature to each message, and the matching public key sits in your domain settings. It works like a tamper-evident seal on a package.
\nDMARC answers: what should a mailbox provider do when a message fails the first two checks, and who should hear about it? It also requires that the domain shown in the From line line up with the domains that passed. You can start with a monitoring-only policy that simply sends you reports, then tighten to quarantine and finally reject once you understand every legitimate system that sends mail as you.
\nBrand managers rarely type these records themselves. Ask your technical contact three questions: are all three in place, does DMARC alignment pass for every tool that sends as us, and who reads the DMARC reports? That last one usually gets a long pause.
\nNext comes behavior. Mailbox providers watch complaint rates, bounce rates, spam trap hits and whether people open, click, reply or delete without reading. Keep complaints very low, remove hard bounces immediately and stop mailing people who show no sign of life.
\nThen consistency. A new domain or a new sending address is a stranger. Warm it up by starting with your most engaged subscribers, sending modest volume and increasing it in steps over several weeks while watching complaints and bounces. Marrow Creek's mistake was not the promotion. It was sending the largest volume the address had ever handled, to the coldest contacts, on day one.
\nSegmentation is the practice of sending different messages to different groups. Most brand managers overbuild it. They create 30 segments on age band and region, and then discover that nobody has the creative capacity to write 30 different emails.
\nStart with a few segments built on what people did, because behavior predicts behavior better than demographics do. Recency (opened or clicked in the last 30 days, 31 to 90, 91 to 180, longer) and customer status (never purchased, purchased once, repeat) cover most of the value. Add product interest only when you have different things to say.
\nRecency is the segment that protects deliverability. Mail your active group first and most often. Mail the cooling group less often. Put the long-silent group into a win-back flow, covered below, and remove them if they do not respond.
\nHere is a quick worked example, illustrative only. Suppose a list has 50,000 addresses. Active in the past 90 days: 12,000. Cooling, 91 to 180 days: 10,000. Silent beyond 180 days: 28,000. Many brands in this position mail everyone weekly. A better plan mails the 12,000 weekly, the 10,000 twice a month, and puts the 28,000 into a three-message win-back sequence. Complaint rates usually fall because the least interested people stop getting the same ask every week.
\nA flow is a set of automated messages triggered by what a person does. A campaign goes to many people at once, on your schedule. A flow goes to one person at a time, on theirs. For most brands, a handful of flows do more steady work than any amount of weekly newsletters.
\nWelcome flow. Triggered by sign-up. Its job is to deliver what you promised, set expectations about frequency and show the person why you are worth opening. Three to four messages across about two weeks is a sensible start. The first message deserves special care because welcome emails tend to be opened more than almost anything else you send, and early engagement shapes how mailbox providers treat you.
\nBrowse or cart abandonment flow. Triggered when someone shows purchase intent and leaves. Its job is to remove friction, not to throw a discount at everyone. Send a reminder, answer the likeliest objection such as shipping time or returns, and only then consider an incentive. If you open with a discount, you train customers to abandon on purpose.
\nWin-back flow. Triggered when engagement stops. Its job is a clean decision: either the person re-engages or they leave the active list. Two or three messages, the last one honest: \"We will stop sending unless you tell us to keep going.\" Then stop.
\nAssign each flow one metric that matters. Welcome: first-30-day purchase rate. Abandonment: recovered orders per thousand triggered messages. Post-purchase: repeat purchase rate. Win-back: reactivation share and the complaint rate during the sequence.
\nOperational guidance only, not legal advice. Your counsel decides how these rules apply to your business, and rules outside the United States can be stricter.
\nIn the United States, the main federal rule for commercial email is the CAN-SPAM Act. In ordinary language, it asks senders to do roughly these things:
\nA working unsubscribe is also a deliverability asset, because a person who cannot find the link will find the spam button. Keep a suppression list, a permanent record of everyone who opted out, and check every outgoing send against it. The common failure is an export loaded into another system without the opt-outs attached.
\nIf you mail audiences in other countries, assume you need a more careful look at consent. A conservative default is explicit permission, a clear record of it and an easy way to withdraw. Ask counsel before you scale.
\nOpen rate used to be the headline metric. It no longer deserves that role. Several mailbox privacy features load images automatically or through a proxy, which can register an open when the person never looked at the message. Security scanners can also trigger opens and clicks. The practical effect, stated qualitatively, is that open rates are inflated and noisy, and the amount of inflation varies by audience and by mailbox provider. A message can show strong opens and produce no sales.
\nDo not throw opens away entirely. Use them only for rough comparisons inside one audience and period, such as two subject lines against equal halves of one segment.
\nLean on signals closer to money and closer to a human decision:
\nThe holdout is the most honest tool you have. Suppose you mail a segment and hold out a random tenth of it. If the holdout group buys at nearly the same rate as the mailed group, email is claiming credit for purchases that would have happened anyway. If the mailed group buys clearly more, you have measured real lift. Attribution reports cannot do that for you, since they only count what touched what.
\nThis is where the open source path needs the most honesty. There are two separate jobs hidden in the word \"email system.\" One is the application: the list, the segments, the templates, the flows and the reporting. The other is the delivery: the mail server that connects to mailbox providers and hands over each message.
\nThe application layer is very doable with open source software. A small team can run a newsletter manager or marketing automation platform on a modest server and keep full control of its data.
\nThe delivery layer is a different animal. Running your own mail server means you manage the server's network address and its reputation, reverse DNS records, authentication, bounce handling, feedback loops, blocklist monitoring and queue behavior under load. A brand new server address has no history, and large mailbox providers are slow to trust strangers.
\nHere is the decision rule I give people. If you have someone who has run mail infrastructure before, who will own it for the long term and who will read the logs when a provider starts rejecting you, self-hosting can protect margin and keep your data inside your control. If you do not, you will spend more in lost sales and staff time than you save. In that case, point your open source application at a commercial relay service for delivery and keep the list, the data and the logic in your own hands. That is not a free option, and I will not pretend it is, because relay services bill for volume. Budget for it as a cost of doing business.
\nFor a very large sender, no open source setup matches the reputation and delivery tooling that dedicated providers have built over many years, so say so plainly in your planning. Pick the setup you can operate on a bad day.
\nSender reputation: The track record mailbox providers assign to your sending domain and network address based on complaints, bounces, engagement and consistency. Suppression list: A permanent record of addresses that must never be mailed again, including everyone who unsubscribed, complained or hard bounced. Spam trap: An address kept by mailbox providers or blocklist operators to catch senders who collect or buy addresses carelessly. Warm-up: The practice of raising sending volume gradually on a new domain or address so mailbox providers learn that your mail is wanted. Holdout group: A random slice of an audience that deliberately receives no message so you can measure true lift against those who were mailed.
\n1. Mailing an old, cold list at full volume from a new address on day one, which looks to a mailbox provider like a spammer arriving with a megaphone. 2. Treating open rate as proof of success when privacy features and security scanners inflate it, and then optimizing subject lines against noise. 3. Loading an export from one system into another without the unsubscribe and complaint records attached, so people who left are mailed again. 4. Building 30 segments before building a single working flow, and then lacking the creative capacity to say anything different to any of them. 5. Choosing to self-host the mail server because licensing is free, without anyone who will own reputation, logs and blocklist cleanup for years.
\n1. Confirm every new subscriber with a link click and store the source, date and the exact wording they saw at sign-up. 2. Keep DMARC reports flowing to a named person who reads them monthly, and tighten the policy in steps as legitimate senders are identified. 3. Segment first on recency and purchase status, mail the active group most often, and put the silent group through a win-back flow before removing them. 4. Give every lifecycle flow one owner, one job and one metric, and review each flow every quarter for stale offers and broken links. 5. Run a holdout test on your largest flow or campaign at least twice a year to measure what email adds beyond what would have happened anyway.
\n1. Send your most important new campaign to your most engaged segment first and watch complaints for a few hours before releasing the rest. 2. Use a separate subdomain for marketing mail, so a rough promotion cannot slow the receipts and shipping notices your customers depend on. 3. Test the unsubscribe link yourself in every template, from a real inbox, before each major send. 4. Keep a clean plain text version of each template, since it looks less mass produced. 5. Ask replies to come to a monitored address and answer them. A conversation is the strongest engagement signal a mailbox provider sees, and it tells you things no dashboard will.
\nThree tools cover the owned side of this chapter. Listmonk (AGPL) is a self-hosted newsletter and mailing list manager, a good starting point for list management, templates and campaign sends at modest scale. Mautic (GPL) is a marketing automation platform that handles segmentation, forms and lifecycle flows, and it asks more of your technical team to run well. Postal (MIT) is a mail delivery server for teams ready to operate their own sending, with the duties described above. None of them replaces authentication, list hygiene and warm-up, and none gives you a sender reputation.
\nEmail is the highest-control channel because the levers that decide its results are ones you can pull: permission, authentication, segmentation, frequency and removal of people who have gone quiet. You own the list, but mailbox providers own the inbox, so reputation is the asset you manage daily. SPF, DKIM and DMARC prove your identity and are now the entry fee. Behavioral segments and a few lifecycle flows beat elaborate demographic schemes, and CAN-SPAM duties are mostly about honesty, a working unsubscribe and a physical address, with counsel deciding the details. Open rates are distorted by privacy features, so rely on conversions, replies, list health and holdout tests. The open source applications are very workable, but self-hosting delivery demands real skill and long-term ownership.
\n1. A large list of reluctant people hurts you, because complaints are the loudest signal a mailbox provider reads. 2. SPF says who may send as you, DKIM says the message was not altered, and DMARC says what to do when the checks fail and who gets the report. 3. Behavior and recency make better segments than demographics, and a segment is only worth building if it receives something different. 4. Open rates are inflated and noisy, so use holdout tests, conversions and reply rates to learn what email truly adds. 5. Own the list and the application with open source tools, and be honest about whether you have the skill to own delivery too.
\nYour big send starts with a list split by recency, with only the active group mailed first and a random holdout slice reserved so you can see what email adds. Then ask your technical contact for one thing: the name of the person who reads your DMARC reports.
\n1. Federal Trade Commission. CAN-SPAM Act: A Compliance Guide for Business. Plain-language summary of commercial email requirements. 2. Internet Engineering Task Force. RFC 7208, Sender Policy Framework (SPF) for Authorizing Use of Domains in Email,. 3. Internet Engineering Task Force. RFC 6376, DomainKeys Identified Mail (DKIM) Signatures. 4. Internet Engineering Task Force. RFC 7489, Domain-based Message Authentication, Reporting, and Conformance (DMARC). 5. Listmonk, Mautic and Postal project documentation, covering installation, licensing and operating guidance for each tool.
"},{"id":"chapter-36-affiliate-and-partner-programs","title":"Chapter 36: Affiliate and Partner Programs","html":"By the end of this chapter, the reader will be able to: 1. Explain why an affiliate dashboard total and a finance-approved total of caused sales usually disagree. 2. Compare flat, tiered, and hybrid commission designs and match each to a partner type. 3. Identify coupon leakage, cookie stuffing, and affiliate brand bidding in raw click and order records. 4. Design a first-party tracking path that logs clicks, orders, and refunds without relying on a network. 5. Evaluate an affiliate program's incremental cost per order using a holdout and a simple ledger. 6. Draft partner onboarding rules that include disclosure duties, prohibited tactics, and enforcement steps.
\nOpen your affiliate dashboard. It says the program drove 600 orders last month. Now walk that number into a meeting with your finance partner and watch what happens. She will ask one question: how many of those orders would have happened if the program did not exist?
\nYou probably cannot answer. Neither can the dashboard.
\nHere is the assumption I want to challenge: that paying only for outcomes means you only pay for results. It does not. You pay for outcomes that were claimed, and a claim is not a cause. A shopper who already put a jacket in the cart, searched for a coupon code, clicked a code site, and bought, produced one order and one commission. The code site did not create the demand. It stood at the register and collected a toll.
\nI have watched teams celebrate a low affiliate cost per acquisition while the same customers were also credited to paid search and email. Every channel reported the win. The bank account reported something smaller.
\nNone of this means affiliate programs are bad. Good content publishers, reviewers, newsletter writers, and niche communities introduce your brand to people who would never have found it alone, and paying them per sale is a fair and efficient deal. The problem is that the same payment mechanism also rewards the cheapest tactics available: intercepting people who were already coming.
\nThis chapter treats an affiliate program as a ledger problem before it is a marketing problem. Here we stay on the money, the proof, and the guardrails.
\nA commission buys a claim on an order. Whether that claim is honest depends on three things: the order, the click, and the gap between them.
\nConsider the chain. A shopper clicks a partner link, the link sets a small browser file called a cookie (a tag the browser stores and sends back later), and the cookie carries a partner code. When the shopper checks out within the cookie window, the order is matched to that code. The program pays. Nothing in that chain asks whether the partner changed the shopper's mind.
\nThink of a basketball scorer's table. The official scorer records who made the basket, not who set the screen, ran the play, or drew the double team. Affiliate tracking is a scorer's table with one rule: whoever touched the ball last gets the points. Last touch rewards the partner closest to the checkout, and the partner closest to checkout is rarely the partner who created the intent.
\nA review site that compares six products is doing early work. A browser extension that pops a code box at checkout is doing late work. Both appear as one line item named \"affiliate\", and a flat rate pays them the same for very different value.
\nSo the first discipline is labeling. Tag every partner with a type at signup: content, review, newsletter, community, loyalty or cashback, coupon, comparison, extension, or sub-network. You will use that label in every later analysis.
\nCommission design is where a program becomes either a growth engine or a leak. There are four common structures, and each fits a different partner.
\nA flat rate pays the same share of order value to everyone. It is simple to explain and simple to audit, and it overpays late-funnel partners. A tiered rate pays more for partners who start the shopper's path and less for partners who finish it. A fixed bounty pays a set amount per qualified action, which suits lead generation or first-time buyers. A hybrid pays a small base share plus a bonus tied to a hard-to-fake outcome, such as a new customer who does not return the product.
\nHere is a rule that has served me well: the commission should rise with the partner's distance from the checkout. Content that introduces your brand to a cold reader earns more. A code box at checkout earns little, or nothing.
\nDefine a new customer yourself, in your own customer table, and hold commissions as pending until the return window closes so you pay on net orders only.
\nNow a worked example, illustrative only. Suppose a fictional outdoor retailer, Juniper Lane, sees 500 affiliate orders in a month at an average order value of 200 dollars, with a flat 10 percent commission. Reported revenue is 100,000 dollars and commissions are 10,000 dollars, so the dashboard shows 20 dollars of commission per order. Looks cheap.
\nNow suppose a holdout (a group deliberately kept from the program, covered below) suggests that nine in ten of the 300 coupon site orders would have happened anyway, eight in ten of the 100 partner brand search orders would have happened anyway, and four in ten of the 100 content orders would have happened anyway. Incremental orders: 30 plus 20 plus 60, or 110. The same 10,000 dollars now buys 110 caused orders, which is about 91 dollars each. If the retailer earns roughly 80 dollars of gross margin per order, the program loses money at the margin even though the dashboard glows.
\nRedesign it: pay content partners 12 percent, coupon partners three percent, and ban brand search bidding.
\nMost affiliate problems are not dramatic fraud. They are quiet leakage that follows the incentive. Learn the four patterns and you will catch most of the cost.
\nCoupon leakage happens when a discount code escapes the audience it was meant for. A code created for a creator's followers lands on a public code site, and now every shopper who searches for a code, finds one, and buys triggers the partner's commission and a discount. You paid twice for a sale you already had. Watch for orders where the code is public, the order follows a code search by seconds, and the session started on a coupon page.
\nCookie stuffing means setting the partner cookie without a real click, often through hidden page elements or automatic redirects. The signature is a gap: many partner cookies set, very few actual page views on your side, and orders attributed to partners whose click records show no engagement. Compare time on site, pages viewed, and landing page behavior between partner segments. Stuffed traffic looks oddly still.
\nAffiliate brand bidding means a partner buys search ads on your brand name, so the shopper who typed your name clicks the partner's ad, passes through a tracking link, and lands on your site. The partner earns a commission for traffic your brand already owned. Review the search results for your own name regularly, from different devices and locations, and check the partner's landing page host names.
\nBrowser extensions that inject links at checkout replace the existing partner cookie in the last moments. Look for orders where the click and the order are under a minute apart and the cart was already full.
\nSort orders by time between click and order. A partner whose median is a few seconds is almost certainly capturing, not creating.
\nPartners who earn a commission have a financial connection to you, and the shopper deserves to know. Under United States advertising guidance, material connections between an endorser and a brand should be disclosed clearly and close to the claim. A line buried in a footer, or a link labeled with a vague word, is the kind of thing that creates exposure.
\nYour duty does not end when you hand the partner a rulebook. A brand can be held responsible for what its partners say in its name, which means you need a monitoring habit, not just a policy. This chapter gives operational guidance only. Your counsel decides what your specific facts require.
\nHere is what I put in every partner agreement: 1. A plain-language disclosure requirement: a visible statement near the link that the partner may earn a commission. 2. A ban on health, income, or performance claims that you have not supplied and approved in writing. 3. A conservative default for audiences: no marketing aimed at children, and no partner content for sensitive categories without prior written approval. 4. A right to review, request edits, pause links, and withhold commission for violations.
\nAudit a sample monthly. Pick 10 partner pages at random, read them like a stranger would, and file dated screenshots in your ledger. A record of consistent review is worth far more than a promise.
\nNetworks track partners with their own pixels and redirects. That gives you convenience and a data set you do not own. A first-party design flips this: your domain issues the link, your server records the click, and your server records the order.
\nThe mechanics are simple enough to explain on one page. Each partner receives a link on your own domain with their code, such as a short path under your site. When a shopper clicks, your server logs a click record: partner code, a click identifier, time, landing page, and a hashed browser fingerprint you are permitted to keep. Your server then sets a first-party cookie with that click identifier. At checkout, your order system reads the cookie and writes the click identifier next to the order. When the payment succeeds, your server sends a conversion event (a short message from your back end, not the browser) to your analytics store.
\nThis buys you records that survive a change of network, and the ability to join click data to refunds, new-customer flags, and margin. Browser tracking limits hurt first-party cookies less than third-party ones, though they still apply, so test them.
\nMatomo, an open source analytics platform, can record campaign names, partner codes, and conversion goals from your own server, and its raw data stays in your database. You still need to respect user consent rules. Collect only what you need and honor opt-outs.
\nBe honest about the limits. Mature open source affiliate platforms are scarce, and no free tool matches a large network's partner marketplace, payment processing, and tax paperwork. For most small and mid-size programs, a handful of vetted partners tracked in your own system will outperform a crowded network.
\nYou do not need a platform to keep honest books. You need one table and one rule: nothing is paid until the ledger says it is payable.
\nBuild a table, in PostgreSQL or any relational database, with one row per partner order. Columns: order identifier, partner code, partner type, click identifier, click time, order time, gross value, discount code used, new-customer flag, refund status, commission rate, commission owed, status (pending, approved, reversed, paid), and a notes field. Add a second table for partners: legal name, tax form status, type, rate, agreement date, and approval status for any restricted claims.
\nRun four checks every cycle: 1. Click-to-order time by partner, to find the capture patterns described earlier. 2. Orders with public codes, to find coupon leakage. 3. Brand search terms in partner landing pages, to find brand bidding. 4. Refund and reversal rates by partner, to find inflated orders.
\nThen pay on approved rows only. Finance gets a number it can reconcile to the bank, and that is why it will trust your story.
\nMetabase, an open source reporting tool, can sit on top of that database and give your team a live view without anyone writing code.
\nThe ledger tells you what was claimed. A holdout tells you what was caused.
\nThe simplest method works at the partner level. Pick a partner you suspect is capturing. For a few weeks, pause their links for a random half of your traffic, or pause the partner entirely in one set of regions and compare. If orders in the paused group fall only slightly, the partner was collecting a toll. If they fall noticeably, the partner was moving shoppers.
\nA few rules keep this honest: 1. Randomize on something the partner cannot control, such as region, visitor identifier, or day. 2. Run long enough to cover your normal buying cycle. 3. Decide the decision rule before the test: for example, cut the rate if lift is below a threshold you wrote down in advance. 4. Test one change at a time.
\nA holdout is not always possible. Small programs may lack the volume, and some partners are contractually protected. In that case use a softer proxy: compare a partner's new-customer share, click-to-order time, and return rate with your program average, and pay accordingly. A proxy is a hint. A holdout is evidence.
\nRules that nobody reads are not rules. Onboarding is your one clean chance to set terms before habits form.
\nStart with a short application. Ask who the partner is, what their audience is, where they will place links, and how they will promote your brand. Look at their site, social account, or newsletter yourself. Reject partners with thin pages made entirely of offers, sites that imitate your brand, and anything that targets children.
\nThen send a one-page plain-language rulebook stating allowed channels, prohibited tactics (brand bidding, cookie stuffing, injected links, unapproved claims, code sharing outside the intended audience), your disclosure requirement with an example sentence, your commission schedule and payment timing, and the consequences of violations: warning, link pause, commission reversal, removal.
\nGive every partner a unique link and a unique code. If one code shows up on a public site, you know who let it out.
\nReview new partners closely for the first 30 days, then settle into a monthly sample.
\nReward the partners who do real work with early product information, approved images, and a clear contact.
\nAffiliate network: A third-party platform that lists brands and publishers, tracks clicks and orders, and handles commission payments for a fee. Attribution window: The number of days after a click during which an order can still be credited to the partner who sent it. Clawback: The reversal of a commission already paid, usually because an order was refunded, canceled, or found to violate program rules. Last-click credit: A rule that gives the full commission to the final partner link clicked before purchase, regardless of who introduced the shopper. Cookie stuffing: Setting a partner tracking cookie in a shopper's browser without that shopper actually clicking the partner's link.
\n1. Treating the network dashboard as the source of truth, which leaves you unable to reconcile partner payouts with your own order and refund records. 2. Paying one flat rate to every partner, which subsidizes late-funnel capture at the same price as real discovery work. 3. Allowing partners to bid on your brand name in search without a written rule, then discovering you are paying commissions on your own branded traffic. 4. Skipping disclosure review because the partner agreement already mentions it, when no one on your team actually reads the live pages. 5. Running a holdout once, getting a bad result, and either ignoring it or cutting the whole program instead of adjusting rates by partner type.
\n1. Label every partner by type at signup and report performance by type, not by program total. 2. Set commissions that rise with distance from checkout, and hold payment until the return window has closed. 3. Keep a first-party click and order ledger that you own, and reconcile it to finance every month. 4. Write a one-page rulebook with prohibited tactics and consequences, and require acceptance before any link goes live. 5. Test lift on your largest or most suspicious partners with a holdout at least twice a year, and write the decision rule before the test starts.
\n1. Search your own brand name monthly from a clean browser and several locations to spot partner ads before the invoice does. 2. Issue unique discount codes per partner so any leak points to a specific source. 3. Ask for a content calendar from your top partners, which both strengthens the relationship and surfaces claim problems early. 4. Keep screenshots of partner pages with dates, since pages change and your memory will not hold up under questions. 5. Compare each partner's click-to-order time with your program median, and investigate anyone whose orders arrive almost instantly.
\nMatomo (GPL license) records campaign and partner parameters, conversion goals, and visitor sessions on your own server, so your click data stays yours. PostgreSQL (PostgreSQL license) stores the partner ledger, click records, and refund data in tables you control and can audit. Metabase (AGPL license) connects to that database and shows finance and marketing the same live reports without custom code. There is no mature open source affiliate network that replicates a large publisher marketplace, so these three tools support a first-party program built on a small set of vetted partners.
\nAffiliate programs pay for claimed outcomes, and the gap between a claimed order and a caused order is where budgets quietly drain. A sound program starts with commission design that rises with distance from checkout, tracking on your own domain and server, and a ledger that finance can reconcile to the bank. It watches for coupon leakage, cookie stuffing, injected links, and partner brand bidding, and it treats disclosure as a monitored duty. Holdouts measure lift where volume allows, and proxies fill the gap where it does not. Clear onboarding rules, unique links and codes, and a human relationship with the best partners keep the program honest and profitable.
\n1. A dashboard total counts claimed orders, so ask finance-style questions about which orders would have happened anyway. 2. Partner type predicts value, so label partners at signup and set rates by their distance from checkout. 3. Coupon leakage, cookie stuffing, injected links, and brand bidding each leave a visible trace in click and order timing. 4. Your own click and order ledger, stored on your own database, lets you pay only on net, approved orders. 5. A written rulebook with unique codes and disclosure requirements makes violations traceable and enforceable.
\nThis week, export last month's partner orders, sort them by time between click and order, and pause commissions on any partner whose median is under one minute until a holdout proves they cause sales.
\n1. Federal Trade Commission. Guides Concerning the Use of Endorsements and Testimonials in Advertising, 16 CFR Part 255. 2. Federal Trade Commission. Disclosures 101 for Social Media Influencers. Plain-language guidance on disclosing material connections. 3. Federal Trade Commission. Dot Com Disclosures: How to Make Effective Disclosures in Digital Advertising. 4. Brodersen, K.H., Gallusser, F., Koehler, J., Remy, N., Scott, S.L. Inferring causal impact using Bayesian structural time-series models. Annals of Applied Statistics. Published methodology for estimating lift from a comparison group. 5. Matomo documentation. Campaign tracking, goals, and server-side tracking API.
"},{"id":"chapter-37-retail-media-and-marketplace-ads","title":"Chapter 37: Retail Media and Marketplace Ads","html":"By the end of this chapter, the reader will be able to: 1. Distinguish a sales figure a retailer reports from a sales lift a brand can verify. 2. Explain how sponsored listing auctions price visibility and why bids climb as more brands compete. 3. Compare the reporting limits of retailer networks, marketplaces and an owned site. 4. Calculate contribution margin after advertising, fees and promotions for a single product. 5. Design a holdout test that estimates incremental sales on a retail media campaign. 6. Evaluate a budget split across retailer, marketplace and own site using a written decision rule.
\nRetail media is the most trustworthy advertising you can buy. The ad runs next to the shelf, the sale happens on the same property, and the retailer sees both. No guessing, no stitched-together pixels, no modeled conversions. That is the pitch, and it is mostly true as far as it goes.
\nHere is the counterargument, and it is the stronger one. Seeing both ends of the path is not the same as proving the ad caused the sale. The retailer measures what happened after an ad appeared. It does not measure what would have happened anyway. A shopper who typed your exact product name into the search box was going to buy it. If your sponsored listing sat on top of the organic listing and caught the click, the report counts a sale. Your bank account counts a sale you already owned and paid extra to collect.
\nI have watched sharp buyers fall for this because the dashboard looks so clean. A return on ad spend (attributed sales divided by ad cost) of eight to one feels like a gift. Often it is a measure of how much of your existing demand you chose to rent back.
\nNone of this means you should stay away. Shoppers on these properties are close to a decision, and a well-run program can move real volume. It means you should treat the retailer's number the way a coach treats a stat sheet from the other team's scorekeeper. Useful. Probably accurate on its own terms. Not the final word on who won.
\nThis chapter gives you the working vocabulary, the auction mechanics, the reporting limits, a margin calculation you can run on one product, a test you can run with your own hands, and a rule for dividing money among retailer networks, marketplaces and your own site. The numbers are illustrative. Your categories, fees and conversion rates will differ, so build your own from your own data.
\nA closed loop means the seller of the ad space is also the seller of the product. When someone sees or clicks an ad and later buys on the same property, the retailer can link the two events to one account. Other channels lose that link at a browser boundary or an app boundary. Retailers do not.
\nThat is a real advantage, and it is a narrower one than the phrase suggests. The loop closes on the retailer's side only. You receive a summary: impressions, clicks, spend, attributed sales, sometimes new-to-brand share (the portion of attributed buyers who had not bought from your brand within a lookback window the retailer defines). You do not receive the shopper-level rows behind it. You cannot rerun the attribution with a different window. You cannot check whether the same buyer was counted under two campaigns.
\nThree questions sort what you can verify from what you must take on trust.
\nFirst, did the sale happen? Usually yes, and your own sales and settlement records will confirm it. Second, did the ad touch that shopper before the sale? Probably, though you are relying on the retailer's logging. Third, did the ad change the outcome? Nobody on the platform answers that for you. Attribution rules count a sale if an ad touch occurred inside a window, often a number of days after a click. That rule cannot separate a shopper persuaded by the ad from a shopper who was already walking to the register.
\nSo the verifiable portion is the sale and the cost. The unverifiable portion is the cause. Write both columns down before the first campaign launches, and agree internally which one drives budget decisions. If finance wants incremental profit and the dashboard supplies attributed revenue, someone has to translate, and the translation is your test design later in this chapter.
\nMost spend inside retail media goes to sponsored listings: ads that look like product results and appear in search or category pages. Understanding the auction behind them explains why costs rise.
\nA shopper searches. The platform finds the listings eligible for that query, scores each by bid and by predicted likelihood of a click or a purchase, and places the highest scores in the paid positions. You pay when someone clicks, typically an amount tied to the bid just below yours. Relevance matters because the retailer earns more from an ad that converts than from one that merely wins on price.
\nSeveral dynamics follow from this.
\nBranded queries are cheap and low in incrementality. Your own name or product title draws a high conversion rate, which makes the dashboard glow. But much of that traffic was yours before you paid. Defensive bidding on your own terms has a place, mainly when competitors bid on them, and its value is protection rather than growth.
\nCategory queries are expensive and contested. The shopper has not chosen a brand yet. Winning that click is closer to real persuasion, and every competitor knows it. Bids rise until the marginal sale barely covers the click cost, which is how auctions behave when sellers can see their own returns.
\nOrganic rank and paid rank feed each other. On many properties, sales driven by paid placement lift organic position over time because the algorithm reads sales velocity as relevance. A new product can use paid placement to build early velocity. A mature product with strong organic position may find that pausing ads costs little.
\nA simple illustration. Suppose a snack brand sells a variety pack and spends 30,000 dollars in a month on sponsored listings, with the dashboard reporting 180,000 dollars in attributed sales. Return on ad spend reads six to one. If half the spend sat on branded terms where most buyers would have purchased anyway, the picture changes quickly. The next sections show how to test it instead of arguing about it.
\nRetailers hold information you cannot buy elsewhere: purchase history across thousands of brands, loyalty card data, basket contents, price sensitivity, store versus online behavior. That is why their audience targeting can be precise. A retailer can reach households that bought a competing product last month without any third-party cookie.
\nThe same asymmetry shapes reporting. You see aggregates. The retailer decides the attribution window, the definition of new-to-brand, the treatment of in-store purchases tied to a loyalty account, and how views are credited. Each choice can change reported return by a wide margin, and each differs by network. A seven-day click window on one property and a 14-day view-and-click window on another produce numbers that cannot be placed side by side.
\nMarketplaces add one more wrinkle. A marketplace seller often competes against the marketplace's own private-label products and against other sellers of the identical item. Ad cost rises while price competition squeezes margin. Visibility inside the property is bought from the same party that sets the rules for your fees, your buy-box eligibility (the box that determines which seller gets the add-to-cart click) and your listing quality scores. That is not an accusation of bad faith. It is a structural fact, and the budget you set should account for it.
\nThree destinations compete for your dollars. Each earns its place for a different reason.
\nA retailer network reaches shoppers inside a store relationship you do not control, often alongside in-store purchase data. It suits brands whose growth depends on retail shelf presence, where sales through that retailer matter to the broader distribution relationship.
\nA marketplace reaches shoppers who start their search inside the marketplace, frequently with high purchase intent and low brand loyalty. It suits products where search is the main discovery path, and it carries heavier fee and price pressure.
\nYour own site gives you the customer record, the margin, and the repeat relationship. Spend there usually buys fewer customers per dollar at first. It buys something the other two destinations do not hand over: the ability to market to the same buyer again without paying a toll.
\nA workable rule: fund contribution margin, not revenue; assign each destination a role (defend, grow or build); fund defense at the minimum that protects branded terms; reserve 15 to 25 percent for the own site; move money only on test evidence. Fees, contracts and promotional terms are operational guidance here, and counsel and finance decide them.
\nIllustrative case, with a fictional brand. Ashgrove Outdoor sells trail socks and spends 40,000 dollars a month on commerce ads: 24,000 on marketplace listings, 12,000 on a retailer network, five thousand on its own site. Reported return looked strongest on the marketplace. After a holdout test and a margin calculation, the team found marketplace growth spend earned thin incremental profit once fees applied, while own-site email capture produced buyers who reordered. They moved to 18,000, 10,000 and 12,000 (own site at the larger figure). They did not decide the marketplace was bad. They decided defense and growth needed separate budgets with separate goals.
\nAdvertising does not live alone on a product's profit and loss statement. Price, fees, promotions and ad cost interact, and a campaign can look excellent while the product loses money.
\nBuild contribution margin for one product, line by line. Suppose a cleaning kit sells for 40 dollars. Cost of goods is 14 dollars. Marketplace referral fee is 15 percent of price, which is six dollars. Fulfillment and shipping cost eight dollars. That leaves 12 dollars before advertising.
\nNow add ad cost. If the average cost to win one sale through sponsored listings is seven dollars, contribution falls to five dollars. If a coupon takes another three dollars, it falls to two. A reported return on ad spend of 5.7 to one (40 divided by seven) looks wonderful, and the product is nearly break-even. Raise cost per sale to 12 dollars and each sale loses money.
\nBreak-even return on ad spend is the number to carry into every conversation. Divide price by contribution margin before advertising: here 40 divided by 12, about 3.3 to one. Any campaign below that rate loses money on its sales even if attributed revenue looks large. Any campaign above it earns something, but only to the extent the sales were incremental. If half the sales would have happened anyway, the effective return is half the reported figure, and the break-even bar doubles in practical terms.
\nPrice interacts with the auction too. Competitors undercutting your price can reduce your conversion rate on a listing, which raises your cost per sale even when your bids stay unchanged. Discount depth matters in the same way. A deep promotion lifts conversion and rank, then trains buyers to wait for the next one. Treat price changes and ad changes as one decision, test them separately where possible, and keep a log of what changed on which day so the results can be read later.
\nYou cannot audit the retailer's attribution. You can measure whether sales changed when advertising changed. Three designs fit most budgets.
\nDesign one is the pause test. Turn off a defined set of campaigns for two to four weeks while everything else stays steady, then compare total sales of the product (all sources, organic included) against the prior period and against a comparable product left running. A pause is crude, since seasons and stock shifts muddy it, but it answers the branded-term question quickly.
\nDesign two is the matched-product split. Pick two products in the same category with similar sales patterns. Advertise one and hold the other quiet for four to six weeks. Compare the lift in total sales for each. This works best when you control price and inventory for both.
\nWrite the decision rule before launch, measure total product sales from your own records, keep price and stock steady, run through at least one full purchase cycle, and repeat before any large budget move.
\nA worked example. Suppose a test group of 10 matched products shows average weekly total sales of 500 units with ads and 440 without. Lift is 60 units, or about 12 percent. If the retailer's report credited the ads with 300 units a week, only about a fifth of the attributed sales were incremental. The retailer's number is not false. It counts a different thing than you need to know.
\nRead every retailer report against your own records. Pull attributed units for a month, pull your own order or settlement records for the same product, and divide one by the other. If a campaign claims 70 percent of a product's sales and your sales did not move when spend moved, you have found the conversation worth having.
\nCheck for double counting. A sponsored product ad and a sponsored brand ad can both claim one purchase, so campaign totals added together may exceed actual sales. Watch search term reports too: terms with many clicks and no sales deserve exclusion, and terms on your own name deserve a separate budget line.
\nFamily-first note: where products are aimed at children, or sold in sensitive categories such as supplements or alcohol, set a conservative default, follow each platform's category rules, and have counsel review claims on the listing.
\nRetail media network: An advertising business run by a retailer that sells placements on its own site, app and sometimes off-property channels to brands that stock its products.
\nSponsored listing: A paid product placement inside search or category results that resembles organic results and is usually billed per click.
\nAttributed sales: Sales the retailer credits to an ad according to its own rules for window, click and view credit.
\nIncremental sales: Sales that happened because of the ad and would not have happened without it, estimated through a holdout or similar test.
\nBreak-even return on ad spend: The minimum ratio of ad-attributed revenue to ad cost at which a product's contribution margin covers its advertising.
\n1. Treating the retailer's reported return as proof of profit, when attribution counts sales the brand would have received without the ad. 2. Pouring budget into branded search terms because they convert well, and then reporting that conversion rate as campaign performance. 3. Comparing return on ad spend across networks that use different attribution windows and view-credit rules. 4. Setting ad budgets from revenue without calculating contribution margin after fees, shipping, coupons and ad cost. 5. Relying only on a retailer's dashboard for sales history, then losing the record when a definition or a reporting tool changes.
\n1. Calculate break-even return on ad spend for each product before launching any campaign, and publish it to everyone who reads the dashboard. 2. Split branded and category terms into separate campaigns with separate budgets and separate success measures. 3. Run at least one holdout or matched-product test per major category each year, and write the decision rule before the test begins. 4. Keep a monthly file of exports from every network in a fixed column layout, owned and stored by your own team. 5. Reserve a steady portion of commerce spend for the own site so the customer record keeps growing.
\n1. Ask the retailer for the exact attribution window and view-credit rule in writing, and note it beside every report you save. 2. Launch a new product with paid placement to build early sales velocity, and review after a few weeks whether organic rank now carries part of the load. 3. Check search term reports weekly in the first month of any campaign, since wasted clicks pile up fastest early. 4. When a competitor starts bidding on your name, test a modest defensive budget before matching them dollar for dollar. 5. Keep price changes and ad changes on a shared calendar so a result can be traced to its cause.
\nYou do not need paid analytics software to do the work in this chapter. pandas (BSD license) is a Python library for tables of data. Use it to load the retailer's exports, standardize column names, and calculate attributed share, break-even return and test lift in a few lines each. DuckDB (MIT license) is an embedded analytical database that runs on a laptop. Use it to store a year of monthly exports from every network and query them with standard SQL, without setting up a server. Together they give you your own record of what each network reported and what your own sales did, which is the foundation for every test described above.
\nRetail media and marketplace advertising put your product in front of shoppers who are near a decision, and the retailer can see the ad and the sale together. That visibility proves the sale happened and what it cost. It does not prove the ad caused it. Sponsored listings run on auctions where branded terms are cheap and low in incrementality, while category terms are costly and contested. Reporting limits, differing definitions and the retailer's control of the data mean you should keep your own records and compare them against every report. Contribution margin after fees, shipping, promotions and ad cost sets a break-even return that no dashboard supplies. Simple holdout tests estimate real lift. A written allocation rule across retailer, marketplace and own site keeps budgets tied to profit and not to the loudest number.
\n1. A closed loop proves the sale and the cost. It does not prove the ad caused the sale. 2. Branded search terms convert well because the buyer was already coming, so budget them as defense. 3. Break-even return on ad spend equals price divided by contribution margin before advertising, and every campaign must clear it. 4. A pause test, matched-product split or geographic holdout gives you an incrementality estimate from your own sales records. 5. Allocate across retailer, marketplace and own site by role (defend, grow, build), and move money only on test evidence.
\nBefore you raise any retail media budget, calculate break-even return on ad spend for the product and run one holdout test, then fund only the campaigns whose lift clears that bar in your own sales records.
\n1. pandas development team. pandas documentation, user guide on data manipulation and grouping. BSD licensed software. 2. DuckDB Foundation. DuckDB documentation, SQL reference and Python client. MIT licensed software. 3. Media Rating Council. Outcomes and Data Quality Standard, current version. 4. Interactive Advertising Bureau. Retail Media Measurement Guidelines, current version, published jointly with the Media Rating Council. 5. Federal Trade Commission. Advertising and Marketing guidance for businesses, including advertising disclosures for online and sponsored content.
"},{"id":"chapter-38-mobile-app-install-and-re-engagement","title":"Chapter 38: Mobile App Install and Re-engagement","html":"By the end of this chapter, the reader will be able to: 1. Define what an install event does and does not prove about a customer. 2. Calculate cost per retained user and cost per incremental install from round campaign figures. 3. Explain, in general terms, how privacy rules limit device-level attribution on phones. 4. Design a deferred deep link flow that carries a new user to the right first screen. 5. Evaluate an install source by cohort lifetime value instead of by cost per install. 6. Build a first-party event plan using open source tools and set a re-engagement test with a holdout group.
\nSuppose you spend 20,000 dollars on an install campaign and the dashboard reports 10,000 installs. That is two dollars an install, and your agency is pleased. Now do the arithmetic nobody puts on the slide. By day seven, 1,200 of those people have opened the app a second time. Of those, 150 have bought something, and the first order earns you 30 dollars in margin, so the campaign has returned 4,500 dollars on the first purchase. Your cost per retained user is about 17 dollars. Your cost per buyer is about 133 dollars. The number that looked like two dollars was never the price of a customer.
\nIt gets worse. Suppose four in 10 of those installs would have happened anyway, because the people were already searching for your app by name. The paid campaign created 6,000 installs, not 10,000, and the real cost of a created install was closer to three dollars and 33 cents.
\nMost brand managers I meet treat the install as the finish line of a campaign. It is the starting whistle. The install proves one thing: a file landed on a phone. It does not prove the person opened it, understood it, trusted it with a payment card, or wanted a notification from you. An installed app that is never opened again is a very expensive icon on a screen of 40 other icons.
\nThis chapter is about measuring the thing after the install, and about the quieter campaign most teams neglect: bringing back people who already know you. Phone attribution has also become harder to read, and I will be honest about what you can and cannot know. The reader here is a working buyer with a real budget, so the focus stays on decisions you can make with the data you actually own.
\nAn install is a count of a download completing, usually the first time the app opens on a device. Three different things hide inside that count, and they behave differently in your budget.
\nThe first is the paid install, caused by an ad. The second is the organic install, which comes from store search, word of mouth, or a recommendation. The third is the assisted install, where an ad mattered but not alone, because the person saw it, searched later, and downloaded. Platforms tend to claim the third group as their own, which is understandable and not always wrong. Your job is to decide how much of it you believe.
\nThink of a stadium ticket scan. The scan tells you a body came through the gate. It does not tell you whether that fan bought a drink, stayed for the fourth quarter, or will return next season. The gate count is the start of the data, not the answer.
\nFor that reason I treat the install as a cheap signal and the first meaningful action as the real conversion. Pick that action before launch. For a retailer it may be a first purchase or an account sign-up with a saved payment method. For a media app it may be three sessions in a week. For a service app it may be a booked appointment. Whatever you choose, it should correlate with people staying and happen early enough to adjust the campaign while money is still flowing.
\nCampaigns optimized for the raw install attract people who tap quickly, download freely, and leave. Campaigns optimized for a deeper event cost more per install and almost always cost less per customer. If your agency resists the deeper event because \"volume will drop,\" ask what volume of customers will drop, not volume of downloads.
\nPhone operating systems now put the user in control of whether an app may link activity across other companies' apps and websites. On Apple devices a person must say yes to a tracking prompt, and many say no. On Android the advertising identifier can be reset or removed, and platform privacy rules continue to change. I will not quote consent rates because they vary by app category, by how the prompt is explained, and by the audience. Assume that a meaningful share of your installs will arrive with no device-level link back to the ad that caused them.
\nWhat replaces the link? Mostly, aggregated and delayed reporting. The operating system or ad platform reports groups of conversions after a waiting period, often with limited detail about which campaign element did the work. Some signals are held back until enough users have acted, so small campaigns may see little or nothing. You can still learn from this data, but you cannot build a precise story of each user from it.
\nThree practical consequences follow.
\nFirst, expect your install counts across the ad platform, the store console, and your own analytics to disagree. That is normal. Decide in advance which source is the record for money decisions, and write it down. I recommend your own event data, because you control it and nobody else reinterprets it.
\nSecond, shorten the distance between ad and measurable action. A person who installs and registers in the same session gives you a first-party event with a clear timestamp. A person who sees an ad on Monday and installs on Friday leaves almost no trail.
\nThird, accept ranges. If a platform says it drove 900 installs and your holdout test suggests 600, the truth is likely between them. Running the business on a range is uncomfortable. Running it on a false point estimate is how budgets get burned.
\nThis is operational guidance, not legal advice. Your counsel decides how consent, disclosures, and store rules apply to your app and your audience.
\nHere is a leak that costs real money. Someone taps an ad for a specific pair of running shoes, goes to the store, installs your app, opens it, and lands on the generic home screen. The shoes are gone. They have to search again. Many will not.
\nA deferred deep link fixes this. An ordinary deep link opens a specific screen inside an app that is already installed. A deferred deep link remembers where the person was headed, survives the trip through the app store, and opens that screen on the first launch after install. Deferred means the instruction waits until the app exists on the phone.
\nYou do not need to buy anything to build this well. The pattern is simple: the ad link carries a campaign tag and a destination, your app reads that information on first launch, and your first-party event records that the destination was reached. Where the match cannot be made, the app should still do something kind, such as showing a short, honest welcome with two choices rather than a blank state.
\nTest the flow on a clean phone before you spend a dollar, on both major operating systems, with the app absent, then present. Test with a mobile data connection and with the wifi off. The most common failure I see is a link that works for the developer, who already has the app, and fails for every actual new customer.
\nFor a family-first standard, keep the first screen free of pressure. If your audience includes teenagers or parents managing a child's device, do not push a purchase prompt, a notification permission request, and a sign-up form in the same ten seconds. Ask for one thing at a time, and ask for permissions only when the person can see why you need them.
\nA cohort is a group of people who installed in the same period from the same source. Cohort lifetime value is what that group is worth over time, tracked from the day they arrived. It is the correct unit for comparing install sources, because it follows the people, not the clicks.
\nSuppose two sources each delivered 10,000 installs. Source A cost two dollars each and Source B cost four. A looks twice as efficient. Now follow the cohorts. By day 30, 800 people from Source A are still active and the cohort has produced 18,000 dollars in margin. From Source B, 2,600 are active and the cohort has produced 72,000 dollars. Source A cost 20,000 and returned 18,000. Source B cost 40,000 and returned 72,000. The expensive source is the profitable one by a wide margin.
\nBuild a simple cohort table with these columns: source, install week, installs, spend, day seven active users, day 30 active users, cumulative margin by day 30, and cumulative margin by day 90. Add a row for organic installs as the baseline. If your paid cohorts do not beat the organic cohort's behavior, you are paying to attract people you would have won anyway, or worse, people who are less engaged.
\nBe patient with the data. A cohort younger than 30 days is a forecast, not a result. A useful habit is to set a payback target (for example, recover spend within 60 days) and measure each source against it while resisting the urge to cut a source on day three.
\nRe-engagement means paying to reach people who already have your app and have gone quiet. It is the least glamorous campaign in the plan and often the most efficient, because the person already knows who you are and the store trip is already done.
\nDefine lapsed carefully. For a daily-use app, seven days of silence may be lapsed. For a seasonal app, 90 days may be normal. Pull the numbers from your own sessions, not from a template.
\nThen segment. Someone who installed and never opened it needs a different message than someone who bought three times and stopped. The first needs a reason to try. The second needs a reason to return, perhaps a restocked item, a new feature, or a reminder that points to their own history.
\nYour owned channels come first: push notifications, in-app messages, and email, all sent only to people who agreed to receive them. Paid re-engagement should cover the gap, targeting lapsed users with a deep link that opens a specific screen, not the home screen.
\nThe trap is credit. People who were about to return anyway will be reached by your ad and will reopen the app, and the platform will report a win. Only a holdout tells you the truth. Suppose you target 50,000 lapsed users and withhold ads from a randomly chosen 5000. If eight percent of the holdout returns on its own and 11 percent of the exposed group returns, your campaign moved three points, not 11. Those three points are what you pay for.
\nInstall fraud exists because installs are countable and often paid per unit. I will describe the patterns and leave out statistics, because the rates change and nobody can verify them from outside.
\nOne pattern is click injection or click flooding: a bad actor fires a flood of fake clicks so that when a real user installs organically, the fraudster's click is the last one on record and takes the credit. Another is install farming, where real devices or emulators install and open an app with no human purpose. A third is bots posing as engaged users who trigger the right event just once to satisfy a payout condition. A fourth is spoofed traffic, where the app name or publisher listed in a report is not where the ad actually ran.
\nYou can spot most of these without special software if you look at timing and behavior. Warning signs include an unusually short gap between click and install, a wave of installs from one source with almost no second-day activity, a sudden spike of events at identical time intervals, and a source with a high install count but retention far below your organic baseline.
\nProtect yourself with contracts and structure. Pay on a post-install event, not a raw install, which makes fraud more expensive to commit. Reserve the right to withhold payment on invalid traffic. Review source-level retention weekly. And remember that a cheap install source with terrible retention is a fraud signal even when nobody is cheating, because the result for you is the same.
\nAll of this depends on one asset: a clean stream of events from your own app. An event is a record that something happened, with a name, a time, and a few properties, such as \"purchase_completed,\" the order value, and the campaign tag that brought the person in.
\nKeep the event list short. Ten to 15 well-named events beat 150 sloppy ones. A good minimum includes app opened, account created, first key action, purchase completed, push permission granted, and push opened. Add the campaign tag and the deep link destination to the first-open event so every later action can be tied back to its source.
\nSend the same events to your own storage as well as to any ad platform, so that you can reconcile numbers when they disagree. Apply the consent choices of the person who generated the data, and collect no more than you need. If your app serves children or families, keep personal data to a minimum and have counsel review before launch.
\nFinally, give someone ownership of a one-page monthly review: installs by source, cost per first key action, cohort day 30 value, re-engagement lift against holdout, and a note on anything that looks like fraud.
\nDeferred deep link: A link that remembers the intended screen and opens it on the first launch after install, so the new user does not land on a generic home screen. Cohort: A group of users who installed in the same period or from the same source, tracked together so you can compare their behavior over time. Retention: The share of a cohort that returns to the app on a given later day, such as day seven or day 30. Holdout: A randomly chosen group that is deliberately not shown the ads, used to measure what would have happened without the campaign. Click injection: A fraud pattern where a fake click is fired just before a genuine install so the fraudster takes credit for it.
\n1. Buying on cost per install alone, which rewards sources that deliver cheap downloads from people who never return. 2. Skipping the clean-phone test of the deferred deep link, so every real new customer lands on the wrong screen while the developer's test passes. 3. Reporting re-engagement wins without a holdout, which credits ads for returns that would have happened anyway. 4. Treating the ad platform's install count as the record of truth when your own events say something different. 5. Judging a cohort on day three, cutting a source that would have paid back by day 60 or keeping one that never will.
\n1. Choose a first meaningful action before launch and optimize the campaign toward it rather than toward the raw install. 2. Track every install source as a cohort with spend, day seven and day 30 activity, and cumulative margin, and compare against the organic baseline. 3. Segment lapsed users by past behavior and send each segment a message tied to its own history. 4. Pay partners on a post-install event, with contract language that allows withholding for invalid traffic, and have counsel review the terms. 5. Keep a short, consistent event list in your own storage, with the campaign tag attached to the first open.
\n1. When two numbers disagree, ask which one you can reproduce from raw events, and use that one for money decisions. 2. Ask for the install-to-registration time on each source. Gaps of a few seconds across thousands of installs deserve a closer look. 3. Ask for notification permission after the person has seen value, not on first launch. 4. Run a small holdout on every re-engagement push, even one that feels obviously effective. Three percent of your audience is a cheap price for the truth. 5. Read the one-star and two-star store reviews monthly. They often reveal a broken first-run screen faster than any dashboard.
\nYou can build a credible first-party measurement setup without paying for any of it. Matomo SDKs (GPL) let your iOS and Android apps send screen views and custom events to a Matomo instance you host yourself. PostHog (MIT core) offers product analytics, funnels, retention tables, and cohort views that are well suited to the day seven and day 30 questions in this chapter. Jitsu (MIT) collects events from your app and routes them to your own data warehouse, so you can reconcile your numbers against whatever an ad platform reports. These tools do not replace a full demand-side platform or an ad network, and I will not pretend otherwise. They protect the one thing you need most: your own record of what happened.
\nAn install is a starting signal, not a customer. The honest price of a customer is found by following cohorts to a meaningful action, subtracting installs that would have happened anyway, and comparing sources on lifetime margin instead of cost per download. Privacy controls have reduced what ad platforms can see about individual users, so your own consented, first-party events are the record you should trust. Deferred deep links protect the first ten seconds. Re-engagement can be efficient, but only a holdout shows how much of the return your ads actually caused. Fraud patterns are easiest to catch by watching timing and retention, and by paying on post-install events.
\n1. The install event proves a download, so pick a first meaningful action and judge campaigns on that. 2. Subtract organic installs before you quote a cost per created install. 3. Compare install sources by cohort margin at day 30 and day 90, not by cost per install. 4. Re-engagement needs a holdout, or you are paying for returns that would have happened anyway. 5. Your own consented event data, tagged with the campaign at first open, is the record to settle disputes.
\nAn app campaign should launch with a named first meaningful action, a day 30 payback target, and a carved-out holdout group. Then judge every source on what its cohort earns, not on what its installs cost.
\n1. Apple Inc. App Tracking Transparency, developer documentation. Describes the user permission prompt for cross-app tracking on Apple devices. 2. Google. Android developer documentation on the advertising ID and privacy changes. Consult the current pages, since this area has changed more than once. 3. Matomo. Mobile SDK documentation for iOS and Android. GPL licensed software for self-hosted analytics. 4. PostHog. Product documentation for mobile libraries, retention, and cohort analysis. MIT licensed core. 5. Jitsu. Documentation for event collection and warehouse routing. MIT licensed software. 6. Interactive Advertising Bureau Tech Lab. Guidance on invalid traffic and ad fraud mitigation in mobile environments.
"},{"id":"chapter-39-sms-messaging-and-compliance","title":"Chapter 39: SMS, Messaging, and Compliance","html":"By the end of this chapter, the reader will be able to: 1. Describe what a defensible text message consent record contains and why a checkbox alone does not make one. 2. Design a consent capture flow that states the program, the frequency, the cost language and the opt-out path in plain words. 3. Apply quiet hours, frequency caps and opt-out handling rules to a live sending schedule. 4. Explain, at a general level, how carrier registration and content standards shape what a brand may send. 5. Evaluate a messaging program for value and fatigue using opt-out and complaint signals instead of send volume. 6. Build a consent ledger and sending workflow using open source software while recognizing which costs remain paid.
\nThe most common assumption in this channel goes like this: a customer gave us their phone number, so we may text them. It sounds reasonable. It fails almost every time it is tested.
\nA phone number handed over for a receipt, a delivery update or a loyalty lookup is permission for that purpose and nothing more. Marketing texts need their own consent, captured in its own moment, in language the person could read and understand. When a brand blurs the two, the problem rarely shows up on day one. It shows up eight months later, when a complaint arrives, a carrier throttles a sending number, or a demand letter lands on the desk of someone who has never heard the word \"TCPA.\" Then the question is not whether the texts worked. The question is whether you can prove the person said yes.
\nI have watched good marketers treat this channel like email with a faster clock. It is not. A text message arrives in the same place as a message from a spouse, a school or a doctor's office. People guard that space. The law guards it too, and the penalties are written per message, which means a mistake scales as fast as your send button.
\nNone of that makes texting a bad channel. Done properly it is one of the few direct lines to a customer that still gets read quickly. But the honest order of operations is compliance first, program design second, volume last. This chapter is operational guidance for brand managers and small business owners, not legal advice. Your counsel decides what the rules require for your business, your states and your message types. What I can do is show you how the work is organized so counsel has something solid to review.
\nConsent is an event. It happens at a specific time, through a specific interface, from a specific person, under a specific set of words. If you cannot reconstruct that event, you do not have consent. You have a belief.
\nThink about how a sports league handles a disputed call. They do not ask the referee to remember. They go to the replay. Your consent record is the replay. It needs to show the exact disclosure text that was on screen, the page or form where it appeared, the moment of agreement, the phone number, and how you know the person submitting the number controls it.
\nThat last point trips up many programs. Anyone can type a number into a web form. If the form says \"enter your phone number for offers\" and a prankster enters a stranger's number, the stranger now receives marketing messages they never requested. A confirmation step, where the new subscriber replies to a first message to confirm, closes that gap. This is often called double opt-in. It costs you a few sign-ups who never reply. It also removes the most common source of complaints, which is texting people who never asked.
\nSeparate consent by purpose. Someone who agrees to appointment reminders has not agreed to promotions. Someone who agrees to promotions from your store brand has not agreed to messages from your sister company.
\nA good disclosure is short, plain and placed where the person cannot miss it. Hidden terms in a footer do not count as a disclosure in any sense I would defend. The text needs to sit next to the button or field, in readable size, before the person submits.
\nAt minimum, the disclosure should say who is sending, what kind of messages to expect, roughly how often, that message and data rates may apply, and how to stop. It should link to your terms and privacy policy. Counsel may want more. Suppose a fictional boutique, Harbor Lane Outfitters, adds this line beside its sign-up box: \"Text JOIN to receive sale alerts and new arrival news from Harbor Lane Outfitters, up to four messages a month. Message and data rates may apply. Reply STOP to cancel, HELP for help. Terms and privacy policy linked below.\" It is plain, specific and it sets an expectation the program can keep.
\nDo not buy phone lists and do not import numbers from a partner and call it consent. Consent is not an asset you can transfer by contract. A seller's claim that \"everyone on this list opted in\" tells you nothing about what those people saw or agreed to, and the exposure lands on the sender, which is you.
\nFamily-first default: do not collect numbers from anyone you have reason to believe is a minor, and keep sensitive categories such as health, finances or children's products on a conservative footing with counsel involved from the start.
\nProof of consent lives in a database table, not in someone's memory and not in a marketing platform's export you hope to find later. I call this table the consent ledger. Every row is an event, and rows are never edited or deleted. If a person opts out, you add a row. If they opt back in, you add another.
\nA workable ledger records these fields: a unique person identifier, the phone number in a standard format, the purpose of the permission, the event type (opt-in, confirmation, opt-out, help request), the timestamp with time zone, the source (form name, keyword, kiosk), the disclosure text version, the page address or ad reference, and the technical details such as the submitting network address and device description. Keep a separate table of disclosure versions, so each row points to the exact words shown.
\nWhy append-only? Because a ledger you can edit is a ledger a skeptical reviewer will distrust. Append-only structure gives you a clean timeline: yes on this date, confirmed on this date, stop on this date. When a customer says \"I never signed up,\" you can show the form, the words, the confirmation reply and the time.
\nPostgreSQL is a strong fit for this work. It is free, mature and handles constraints well, so the database itself can refuse a send to anyone whose latest ledger event for a purpose is an opt-out. Put that rule in a view or query that your sending tool must consult. Do not rely on a marketer remembering to exclude a segment.
\nQuiet hours are the times when people should not be contacted. Federal telemarketing rules are commonly applied as eight in the morning to nine at night in the recipient's local time, and some states set narrower windows. The practical rule: figure out where each recipient is, not where your office is. Area codes are a weak guide because people move and keep their numbers. If you store a ZIP code or time zone from sign-up, use it. If you do not, default to a conservative window such as ten in the morning to eight at night in the earliest time zone you might reach.
\nOpt-out handling is the part regulators and carriers watch most closely. Rules of thumb:
\n1. Honor STOP, and common variants such as END, CANCEL, UNSUBSCRIBE and QUIT, in any letter case. 2. Send one confirmation message that the person is unsubscribed, with no marketing in it. 3. Stop all marketing to that number for that purpose promptly. Do not wait for a nightly batch if you can avoid it. 4. Treat plain-language requests (\"please stop texting me\") as opt-outs too. People do not read your instructions. 5. Answer HELP with your brand name and a real contact path.
\nFrequency caps belong in the system, not in a policy memo. Suppose Harbor Lane promises four messages a month. The sending tool should refuse a fifth, even if a manager is begging for one more sale blast. A promise made in the disclosure is a promise you must keep.
\nHere is the part many owners find irritating. You cannot simply buy a phone number and start texting thousands of customers. In the United States, mobile carriers require businesses that send application-to-person messages, meaning messages sent by software rather than a human thumb, to register the brand and the campaign. At a general level, you tell the carriers who you are, what you are sending, how people opt in and what a typical message looks like. The carriers use that to decide whether to deliver your traffic and at what throughput.
\nWhat does not change is the principle: the carriers are gatekeepers, and they judge you partly on your consent practices. If the opt-in flow you describe on the registration form differs from what your website really does, expect trouble. Registration answers should be true, and your ledger should match them.
\nContent standards matter as much. Carriers and regulators treat certain content with extra suspicion: loan and credit offers, gambling, cannabis, hemp products, tobacco, firearms, adult material and sweepstakes involve heightened scrutiny or outright prohibition on some routes. Messages that hide the sender, hide the link destination behind a generic shortener, or use shared numbers across unrelated brands also draw filtering. Use a short link on a domain you control, identify your brand in the first words of the message, and keep the number tied to one purpose.
\nNow the honest cost conversation. Software can be free. Delivery cannot. Every text passes through carrier networks that charge per message, plus registration and number fees, and no open source project removes that bill. You will pay a messaging provider or carrier connection for the transport. The open source layer controls your list, your rules, your records and your reporting. The paid utility moves the message. Budget for both, and be wary of any vendor that promises \"free texting at scale.\"
\nThe Telephone Consumer Protection Act is a federal law that restricts certain marketing calls and texts, especially those sent using automated systems. For marketing texts, the usual expectation is prior express written consent, which means a signed or electronically agreed permission that clearly authorizes marketing texts to a specific number. The law also provides for damages measured per message, and private parties can bring claims. That is why a single flawed sign-up flow, repeated across thousands of messages, turns into a serious matter so fast.
\nI will say this once more plainly: I am describing how the rules are usually discussed, not telling you what they require in your case. Court interpretations shift, agency rules get revised, and counsel is the one who tracks them.
\nThen come the state laws, sometimes called mini-TCPAs. Several states have their own telemarketing or text messaging statutes. They may set narrower calling hours, add registration or disclosure requirements, expand who can sue, or define covered technology differently than the federal law. A national brand that does not segment by state may find it simpler to follow the strictest state rule everywhere, and counsel decides. Most small businesses should simply adopt the stricter standard everywhere. It is cheaper than building 50 versions of a program.
\nCompliance here is not a launch task you finish. It is a monthly discipline, like checking tire pressure before every race.
\nA text program that exists only to blast discounts trains customers to ignore it or leave. Design for the moment when a message is actually welcome.
\nWelcome messages share a few qualities. They are timely (your table is ready, your order shipped), personal in a real sense (your size is back in stock), short, and they ask for little. Transactional and service messages tend to earn the most trust. Promotional messages work best when tied to something the customer asked for, such as a back-in-stock alert or a reminder for a service they booked.
\nSuppose Harbor Lane has 12,000 confirmed subscribers. A weak program sends the same sale text to everyone, every Friday. A better program splits into three purposes: order updates for anyone who bought, back-in-stock alerts for people who tapped \"notify me\" on a product page, and a monthly promotion for those who joined for offers. The second program sends fewer total messages and earns more attention on each, because the audience for each text asked for that type of text.
\nJudge the program on incremental value you can defend, and watch opt-out and complaint rates as health signals.
\nFatigue is the slow bleed. No single text causes it. The tenth text in a month does, when the customer's patience quietly runs out and they reply STOP, or worse, report your number as spam to their carrier.
\nThe signals are measurable. Rising opt-out rate per send. Falling click rate among long-term subscribers. Complaints, including carrier-reported spam flags. Growing delivery failures as carriers filter your traffic. Any of these is a warning that you are sending too much, to the wrong people, or with the wrong message.
\nA few controls help:
\n1. Set a hard monthly cap per person across all purposes, not per campaign. Three campaigns that each respect their own limit can still bury one customer. 2. Create a sunset rule. Subscribers who have not clicked or replied in a defined stretch get a \"still want these?\" message, then removal if they stay quiet. A shrinking list is not always a failure. A smaller list of people who want your texts beats a bigger list of people who are annoyed, because the annoyed ones damage your sender reputation for everyone else.
\nPrior express written consent: A documented agreement, signed or given electronically, in which a person clearly authorizes marketing texts to a specific number from a specific sender. Double opt-in: A two-step sign-up where the person submits a number and then confirms by replying to a first message, proving they control that phone. A2P (application-to-person): Messaging sent by software to people, as opposed to person-to-person texts, and the category that carriers require brands to register. Suppression list: A permanent list of numbers that must not receive messages for a given purpose, checked before every send. Quiet hours: The time window, based on the recipient's local time, outside of which marketing messages should not be sent.
\n1. Treating a phone number collected for receipts or support as permission for promotions, which fails because consent is tied to purpose. 2. Storing only a \"subscribed\" flag with no disclosure text, timestamp or source, which leaves nothing to show a reviewer when a complaint arrives. 3. Buying or borrowing phone lists and assuming the seller's word about consent protects you, when the exposure stays with the sender. 4. Scheduling sends by company headquarters time instead of recipient local time, so customers get messages at hours they consider intrusive. 5. Ignoring plain-language opt-outs and only honoring the exact word STOP, which is the fastest route to complaints and carrier filtering.
\n1. Capture consent separately for each purpose, with the disclosure beside the button and the exact wording version stored with every sign-up. 2. Confirm every new number with a reply step so you know the sender controls the phone and the sign-up was intentional. 3. Keep an append-only consent ledger and make your sending system query it before every message. 4. Enforce frequency caps and quiet hours in software, using one monthly limit per person across all campaigns. 5. Review the program with counsel on a schedule, and test opt-out behavior yourself with a real number each month.
\n1. Put your brand name in the first words of every message, because recipients decide in a glance whether it is spam. 2. Use a short link on a domain you control rather than a public shortener, which carriers often view with suspicion. 3. Keep one sending number tied to one brand and one purpose, since shared or mixed-use numbers inherit each other's reputation. 4. Send transactional and service texts to earn trust first, then ask for permission to send promotions. 5. Budget for carrier fees and registration as a fixed cost of the channel before you compare it with email or paid media.
\nMautic (GPL) is an open source marketing automation platform you can host yourself. Paired with SMS transports, meaning connectors that hand your messages to a carrier-facing provider, it lets you build segments, schedule sends and set frequency rules in software you control. PostgreSQL (PostgreSQL license) is the natural home for your consent ledger, with constraints that prevent bad records and views that your sending tool can consult. Be clear on the limit: the transport itself, meaning carrier delivery, is a paid utility that no open source project replaces. The software layer, the records and the rules are yours.
\nTexting is a high-intent channel because it reaches people where they read fastest, and it is high-risk for the same reason. Consent is a documented event tied to a purpose, proven by a ledger rather than a checkbox. Disclosures must be plain and placed beside the action. Quiet hours, frequency caps and opt-out handling belong in the sending system. Carriers act as gatekeepers through registration and content standards, and delivery is a paid utility even when the software is open source. Federal and state rules reward the strict standard. Value and fatigue management decide whether a program grows or burns out.
\n1. A phone number collected for one purpose is not permission for marketing, so capture consent separately and record it as an event. 2. An append-only ledger in PostgreSQL gives you the proof that a checkbox flag never will. 3. Quiet hours, frequency promises and opt-out rules should be enforced by the sending software, not by memory. 4. Carrier registration is a gate you must pass honestly, and delivery costs money no matter which software you choose. 5. Opt-out and complaint rates, not send volume, tell you whether the program is healthy.
\nThis week, pull one real subscriber from your list and try to reconstruct their opt-in: the exact words they saw, the time, the source and the confirmation. If you cannot, pause promotional sends to that segment until you can.
\n1. Telephone Consumer Protection Act, 47 U.S.C. 227, and the Federal Communications Commission rules implementing it, on consent, calling hours and opt-out handling. 2. Federal Communications Commission, consumer guidance on unwanted calls and text messages. 3. Federal Trade Commission, guidance on telemarketing and the Telemarketing Sales Rule. 4. CTIA, Messaging Principles, industry guidance for wireless messaging. 5. PostgreSQL Global Development Group, PostgreSQL documentation on constraints and views. 6. Mautic project documentation on SMS transports and segment-based campaigns.
"},{"id":"chapter-40-sponsorship-digital-activation","title":"Chapter 40: Sponsorship Digital Activation","html":"By the end of this chapter, the reader will be able to: 1. Explain why a rights fee buys permission and not performance, and why activation spend decides the outcome. 2. Apply the digital activation ladder (owned, earned, paid) to sequence a sponsorship budget. 3. Evaluate co-op and league inventory against a plan built on your own audience goals. 4. Design creator and athlete asset agreements that protect disclosure, approvals and reuse. 5. Construct a consented fan data capture flow that feeds your first-party records. 6. Assess sponsorship digital value using holdouts and plain counts instead of invented media-value equivalents.
\nOne slide from a vendor sold me a sponsorship. It said the logo exposure was worth four times the fee, and I believed it. Nobody ever showed me how the number was built. At the end of the season the team had a great year, my logo had been on the boards for every home game, and I could not point to a single customer who had walked in because of it. The slide had been a story, not a measurement.
\nThat is the assumption I want to challenge: that a sponsorship is an asset you buy and then own. You do not own it. You rent permission. The rights fee gives you the legal ability to use a team's name, marks and access. It does not give you attention, memory or sales. Those are earned after the signature, by whatever you spend on turning permission into activity.
\nThink of it like a season ticket to a stadium where you are not allowed to talk to anyone. The seat is yours. The crowd is not. Activation is the talking.
\nMost brands underfund that second step. They stretch to cover the rights fee, then treat the digital side as an afterthought: a few posts, a banner, a thank-you email. The property's sales team often encourages this, because the fee is their revenue and the activation is your problem.
\nThere is a second trap, and it is quieter. Digital makes sponsorship look measurable, so people grab the numbers that are easy to count (impressions, likes, follower growth) and call it proof. Easy to count is not the same as meaningful.
\nThis chapter treats sponsorship as a digital operating plan with a budget, a sequence and a test. It is operational guidance, not legal advice; your counsel should review any rights agreement before you sign it.
\nYou will hear a rule of thumb in sponsorship circles that you should spend some multiple of the rights fee on activation. Two to one, three to one, sometimes more. Treat any such ratio as a conversation starter and never as a plan, because a ratio ignores what you are trying to accomplish.
\nHere is the mistake inside the multiplier. It starts from the fee, so the fee sets the activation budget. That is backward. The fee is a sunk cost once you sign, and the right question is what each additional activation dollar buys.
\nSuppose a regional furniture retailer, a fictional brand I will call Larkmoor Home, signs a 100,000 dollar rights deal with a minor league arena. A multiplier rule says to add 200,000 dollars of activation. But Larkmoor's goal is showroom appointments. A better plan starts with the number of appointments the business can handle, works backward to the audience it needs to reach, and then prices the activity. That might be 60,000 dollars or 250,000. The fee does not tell you.
\nThere is also a floor. If you cannot afford to activate the rights well, the honest move is to buy fewer rights. A smaller package with proper digital support beats a large package with a logo and silence.
\nBefore signing, ask the property for three things in writing: which digital assets are included (posts, stories, email placements, video inventory), which are guaranteed in count and timing, and which you must pay for separately. Counsel decides the contract language. You decide whether the plan works.
\nI organize activation into three rungs, and I climb them in order: owned, earned, paid.
\nOwned means channels you control: your website, email list, social profiles, app and store pages. This is the cheapest rung, and it is where the sponsorship story has to live first. If a fan clicks through from a game-day post and lands on your ordinary homepage, the sponsorship has already leaked value. Build a dedicated landing page, with the property's marks used under your agreement, one clear offer and a simple way to opt in.
\nEarned means attention that others give you without a direct payment: the property reposting your content, a local reporter mentioning your community program, fans sharing a contest. Earned cannot be bought on demand, but you can design for it. Give people something worth sharing, like a behind-the-scenes clip, a player Q and A or a fan challenge with a real prize.
\nPaid means advertising you buy: social placements, video, connected TV (streaming television delivered over the internet) and audio. Paid is the rung everyone wants to start with, and it is the worst place to start. Paid media pointed at a weak landing page or a vague message simply buys expensive disappointment.
\nThe ladder is a sequencing rule. By the time you spend a paid dollar, the owned assets should be ready to catch traffic, and the earned content should already have shown which messages the audience cares about. Paid then amplifies what is working. A team's audience is also seasonal, so plan the paid rung around the calendar: preseason, opening week, rivalry games, playoffs and the offseason lull.
\nMany properties and leagues sell or bundle digital inventory: sponsored segments in a team newsletter, a branded section of the team app, in-stream video slots, or a place in a league-wide promotion with other sponsors. Some of it comes with the rights deal. Some is shared cost (co-op) money, where a manufacturer, a league or a group of sponsors pays part of the bill.
\nTreat this inventory like any other media offer. It is a place to buy attention, not a plan. Before you accept it, ask what audience sees it, how that audience was built, and what you receive afterward. A team newsletter with a loyal, opted-in readership can be excellent. A league-wide placement shared among a dozen sponsors can mean your message appears for a moment between three others.
\nCheck the data terms first. Does the property share any record of who clicked or signed up, or does it keep everything and give you a summary? If you receive nothing but a recap slide, you have rented an audience you can never reach again. Ask for tagged links so your own analytics can see the traffic, and for a clear statement of what the property will do with fan data collected on your behalf.
\nCo-op funds deserve care too. They often carry rules about creative approval, required language and proof of performance, and missing a requirement can cost you the reimbursement. Assign one person to track the paperwork. It sounds dull, and it protects real money.
\nWhen two sponsors in the same property fight for the same fans, the property's sponsor exclusivity terms matter. Read them before you plan a campaign that assumes you are alone.
\nAthletes and creators carry something a banner never will: a voice the audience already trusts. A veteran coach reading your name off a script reads like a billboard. A player telling a story in their own words reads like a friend.
\nBuild the asset agreement around four questions. First, what exactly will they create (posts, short video, appearances, a series)? Second, who approves the content and how fast? Third, how long may you reuse it, and in which paid placements? Fourth, how will the sponsorship be disclosed? In the United States, the Federal Trade Commission expects clear disclosure of material connections in endorsements, so the post must make the relationship obvious to the viewer. Your counsel can confirm the exact wording for your situation.
\nFor college athletes, name, image and likeness rules vary by state and institution, and the rules keep changing. Check them before you plan anything. Hold a conservative default for any athlete under 18: no promotion of products unsuited to minors, no data collection from them, and written parent or guardian involvement in whatever the law and the property require.
\nReusing creator content as paid media, often called whitelisting or partnership ads, can raise performance because the ad looks native. It also raises the stakes on usage rights. Get the paid usage term, the end date and the takedown process in writing.
\nOne last practical point: ask for raw files, not only posted versions. A clean clip can become six variations later at almost no extra cost.
\nThe most durable result of a sponsorship is a list of people who said yes to hearing from you. A banner impression disappears. An opt-in does not.
\nDesign the capture around a fair exchange. Offer a prize entry, early access to a ticket drop, a discount or exclusive content, and say plainly what the person will receive. Keep the form short. Name, email and one preference question is usually enough to start. Every extra field costs sign-ups.
\nConsent has to be real. That means an unchecked box, plain language about what you will send and how often, a link to your privacy policy, and a working unsubscribe in every message. Do not assume that a fan who gave their email to the team gave it to you. Unless the agreement and the sign-up screen say so, they did not. Laws on consent differ by state and country, and counsel should review your forms.
\nKeep children out of it. If the audience may include minors, as many youth sports audiences do, collect nothing without the legal consent route your counsel approves, or skip collection and run general awareness activity instead.
\nStore captured records in your own database with the source tagged (which game, which placement, which offer). Six months later, that tag lets you compare a rivalry-night contest against a season-opener giveaway and see which one brought people who actually bought. That comparison is worth more than any follower count.
\nSome agencies report sponsorship value by multiplying exposure time or impressions by an advertising rate and calling the result media value. Treat that figure with caution. Nobody paid that rate, no one verified that anyone noticed the logo, and the number says nothing about sales. I do not use it, and I would ask any vendor to show how theirs is built.
\nCount the things that happened. Visits to the sponsorship landing page. New opt-ins by source. Redemptions of a unique offer code. Showroom appointments mentioning the team. Those are small, honest numbers.
\nTo go further, use a holdout, a comparison group that does not see your activity. A geographic test works well for sponsorships: run paid support in some markets and hold back similar ones, then compare sales or sign-ups after the same period. Illustrative example: suppose you run support in six cities and hold back six matched cities. If the supported cities produce 30,000 sign-ups against 20,000 in the held-back cities, the lift is 10,000 sign-ups, not 30,000. That difference is what the activity added.
\nBrand surveys can help with awareness and association, as long as the sample is described and the questions are neutral. Be modest about what they prove.
\nTag every link and every offer so the data lands in your own analytics. A sponsorship you cannot trace is a sponsorship you cannot defend in next year's budget meeting.
\nLet me put the ladder into a plan. Suppose a fictional regional credit union, Prairie Shore Financial, signs a 60,000 dollar season package with a college athletics program and sets aside 90,000 dollars for activation. The goal is 800 new checking account applications across the season.
\nOwned work comes first: 12,000 dollars for the landing page, a creative kit and the opt-in system, plus staff time for email. Earned work gets 10,000 dollars for a player content series and a fan prize that the athletic department can share. Paid gets 55,000 dollars, held mostly back until the content shows which messages work, then released in bursts around home games. The last 13,000 dollars is a reserve for the playoffs, if they come, and for measurement.
\nNotice what the plan does not do. It does not spread money evenly across channels, and it does not set the budget from a multiplier. It also schedules a holdout test across the middle weeks of the season, so the credit union learns what the paid rung really added before it renews.
\nPlans like this fail when approvals slow everything. Agree up front with the property on who signs off on creative, and on what turnaround you expect during game weeks, because the news cycle in sports moves in hours.
\nThis is the second of the three places in this book where I describe my own companies, and I will keep it short and practical. If you want sponsorship activation run by operators who work inside sports media every week, Dakdan Worldwide and Sports Media Inc. are the in-house execution path I use myself: sports-anchored placement, athlete and creator coordination, and measurement that stays in your own accounts.
\nUse them only where they fit. Ask any partner, including me, the same questions you would ask a property: what is delivered, what you receive in data, and how the result is tested. If another route serves your goals better, take it. The ladder, the consent rules and the holdout work the same either way, and none of it depends on hiring anyone.
\nRights fee: The payment that gives a brand permission to use a property's name, marks and access under a written agreement. Activation: The marketing a brand does after signing a sponsorship to turn permission into attention, traffic and sales. Whitelisting: Running paid ads through a creator's or athlete's own social handle, with their permission, so the ad appears to come from them. Holdout: A group of people or markets deliberately kept from seeing a campaign so the results can be compared with the group that did. Opt-in: A clear, voluntary action by a person agreeing to receive messages from a brand.
\n1. Setting the activation budget as a multiple of the rights fee, instead of working backward from the result you need. 2. Reporting invented media-value figures or impression totals as proof of sales, when nobody can show how the number was built. 3. Sending sponsorship traffic to a general homepage instead of a dedicated page with one offer and an opt-in. 4. Assuming the property's fan list is yours, or collecting data from fans without a clear consent step. 5. Skipping the written terms for creator reuse, paid usage dates and disclosure, then discovering the limits after the campaign is live.
\n1. Build owned assets first, test messages with earned content, and then release paid spend in bursts around key dates. 2. Ask the property in writing what digital inventory is guaranteed, how large the audience is, and what data you receive. 3. Tag every link and offer by source so your own analytics can compare placements. 4. Run a geographic holdout on at least one stretch of the season and report the difference, not the total. 5. Hold a conservative default for minors: no data collection and no unsuitable products unless counsel approves the route.
\n1. Request raw creator files, not only posted versions, so one shoot can supply many placements. 2. Put one person in charge of co-op paperwork and proof of performance so reimbursements are not lost. 3. Read the exclusivity terms before planning a campaign that assumes you are the only sponsor in the category. 4. Schedule creative approvals for game weeks in advance, because sports content loses value within hours. 5. Keep a small reserve for playoffs or a breakout player, since the best moments cannot be scheduled.
\nTwo free, open source tools cover most of what a sponsorship activation needs. Matomo (GPL license) is a self-hosted web analytics platform that records visits, goals and campaign sources on pages you control, so you can see which placement drove which sign-up. Metabase (AGPL license) connects to your own database and turns captured records, offer redemptions and spend into simple dashboards that a non-technical team can read. Together they let you tag links, watch results by source, and compare supported and held-back markets without paying a subscription. Neither replaces the judgment of a person who understands the property, and neither tool is a full ad-buying platform.
\nA rights fee buys permission, and the outcome depends on what you spend and build afterward. Start from the result you need rather than a multiplier, and climb the ladder in order: owned assets first, earned content second, paid media last. Treat co-op and league inventory as media to be judged on audience and data terms, and write down creator and athlete rights, disclosure and reuse. Capture fan data with real consent and a conservative stance on minors. Measure with plain counts and holdouts instead of invented media value, and use a partner path only where it fits your goals.
\n1. A rights fee is rented permission, and activation decides whether it produces anything. 2. Budget activation from the result you need, not from a multiple of the fee. 3. Sequence the work as owned, earned and paid, so paid spend lands on assets that are ready. 4. Consented, source-tagged fan data is the lasting product of a sponsorship. 5. A holdout shows what the activity added, while invented media value shows only what someone wanted you to believe.
\nBuild the landing page and opt-in before you sign your next renewal, tag every link by source, and hold back a set of matched markets so you can show what the sponsorship added beyond what would have happened anyway.
\n1. Federal Trade Commission. Guides Concerning the Use of Endorsements and Testimonials in Advertising, 16 CFR Part 255. 2. Federal Trade Commission. Disclosures 101 for Social Media Influencers. Plain-language guidance on disclosing material connections. 3. Brodersen, K.H., Gallusser, F., Koehler, J., Remy, N., Scott, S.L. Inferring causal impact using Bayesian structural time-series models. Annals of Applied Statistics. Published methodology for estimating the effect of an intervention against a comparison. 4. Federal Trade Commission. Children's Online Privacy Protection Rule, 16 CFR Part 312. 5. Matomo and Metabase project documentation. Product documentation for self-hosted analytics and reporting.
"}]},{"title":"Part VII: Data, Measurement, and Attribution","items":[{"id":"chapter-41-first-party-data-strategy","title":"Chapter 41: First-Party Data Strategy","html":"By the end of this chapter, the reader will be able to: 1. Explain why rented audiences get repriced and why owned customer data does not. 2. Design a value exchange that gives a customer a clear reason to share information. 3. Construct a basic event schema and identity rule for a small customer data foundation. 4. Apply data minimization to decide what to collect, what to delay, and what to refuse. 5. Evaluate a governance and retention plan for gaps in consent, suppression, and deletion. 6. Assemble a customer data foundation from open tools and test whether it earns money.
\nHere is a question. If your largest ad platform doubled its prices next quarter, or quietly changed which of your customers it let you reach, how many people could you still contact tomorrow, at no cost, with a message you wrote?
\nMost brand managers answer by pointing at their audience sizes. Four hundred thousand followers. A lookalike pool in the millions. A retargeting list that feels enormous. Those numbers are real, and none of them are yours. The followers live inside a network that sets the reach. The lookalike pool is a mathematical guess the platform runs on its own data and sells back to you by the click. The retargeting list is a set of browser identifiers that browsers and regulators keep restricting. You are not holding an audience. You are holding a lease.
\nIn the Army we had a simple rule about supply: never let your whole mission depend on a line you do not control. Advertising is no different. Rented audiences get repriced, because the landlord can see exactly how much you depend on the building.
\nFirst-party data is the information your customers and prospects give you directly, through your own site, your own checkout, your own emails, and your own conversations. It is the only audience asset you fully own. It is also easy to get wrong. Plenty of businesses have a mountain of customer records they cannot use, cannot trust, or cannot legally touch. Size is not the point. Usable, permissioned, clean data is the point.
\nThis chapter gives the full argument and the build. We will cover why people share data, how to structure what you capture, how to recognize one person across channels, how to collect less on purpose, how to govern what you keep, and how to assemble the whole thing with free, open tools. Operational guidance only: your counsel decides what the law requires of you.
\nPaid media works like the broadcast deal. Every platform audience is built on the platform's collection, the platform's identity graph, and the platform's rules. When you target through it, you pay for access to people you never met. That access is repriced in three ways. The price rises as more advertisers compete for the same pool. The reach shrinks when privacy rules or browser changes cut the signals feeding it. And the definition shifts when the platform redesigns its targeting options and retires the one you relied on.
\nAn owned list does not face those pressures in the same way. If a customer gave you an email address and permission to use it, no platform can raise your cost to contact that person above the cost of sending the message. No algorithm change can delete the relationship. You can still use paid platforms, and you should, but you use them as a place to acquire customers and bring back the ones you already know, rather than as the vault where your audience is stored.
\nA simple test helps. Take your last quarter of sales and ask what share came from customers you could have contacted directly on any day you chose. For many small businesses, that share is smaller than they expect. Raising it is not a marketing tactic. It is a balance sheet decision about which assets you hold.
\nPeople do not give away information for free. They trade it, usually without thinking, whenever the trade feels fair. A strong first-party strategy starts by making the trade obvious and worth it.
\nA value exchange has two sides. The customer gives something small, such as an email address, a size, a preferred store, or a stated interest. You give something the customer actually wants, quickly. Think of a regional sporting goods retailer, illustrative only. A shopper picks a favorite team and sport on a short form. In return the shopper gets a game-day deal sent only for that team, plus early notice of restocks in the right size. The form asked two questions. It earned permission and two useful facts, and the shopper got something that felt personal.
\nAsk for one or two fields at the start, and more later, once the customer has received value. Say plainly what you will send and how often. Make the benefit immediate, such as a discount code, a download, or an order status text, rather than a vague future promise. Put the request at the moment of highest intent, like checkout, a booking confirmation, or the end of a helpful tool, instead of the moment a visitor lands.
\nProgressive profiling is the name for asking in small steps. The first visit captures an email address. The third email includes a one-question preference survey. The first purchase captures a birthday month if a birthday offer exists. Each ask earns the next one, and none of them feels like an interrogation.
\nAn event is a recorded action: someone viewed a product, started checkout, opened an email, redeemed an offer. A schema is the agreed list of event names and the details attached to each. It sounds dull. It is the difference between data you can analyze and a junk drawer.
\nWithout a schema, every team names things differently. One tool calls it \"purchase,\" another calls it \"order_complete,\" a third calls it \"Sale.\" Six months later nobody can build a simple funnel without a week of cleanup. A schema forces one name per action and a fixed set of properties.
\nFor a small team, start with a short list of events that match your sales path, roughly 10 to 15. For example: page viewed, product viewed, item added to cart, checkout started, order completed, form submitted, email subscribed, email unsubscribed, offer redeemed, support request opened. Each event carries a timestamp, a person identifier if known, a source, and only the properties you will truly use, such as product category, order value, and channel.
\nThree habits keep a schema healthy. Use consistent naming, and never rename an event without a recorded reason. Keep a plain shared document listing every event, every property, and what each means. Treat a new event like a new product feature: someone must say who will use it and what decision it supports before it ships.
\nIdentity resolution sounds like advanced engineering. At a basic level, it is simply deciding when two records belong to one person.
\nConsider a customer who browses on a phone, signs up for email on a laptop, and buys in a store using a loyalty phone number. Without identity rules you have three strangers. With them, you have one customer with a history. The aim is not to follow people around the internet. It is to keep your own records tidy.
\nStart with deterministic matching, which means joining records on something the customer gave you and confirmed. An email address, a loyalty number, or an account ID is a strong key. When a visitor logs in or clicks a link in an email, the anonymous browsing session can be connected to that known person, because the customer took an action that proves it.
\nChoose one durable internal customer ID, generated by you, and let every other identifier hang off it. An email address changes, a phone number is reassigned, but your own ID stays put. Store the mapping in one table so a merge or a deletion happens in one place.
\nBe cautious with probabilistic matching, which guesses that two records are the same person based on shared device or location traits. A wrong guess can mix two people's purchase histories or send one person another person's offer. At small and medium scale, deterministic matching covers most of the value with far less risk. If you cannot explain a match to the customer in one sentence, do not make it.
\nCollecting everything feels safe. It is the opposite. Every extra field is a cost: storage, security duties, deletion requests, and risk if the database leaks. Data you never collected cannot be breached, subpoenaed, or misused.
\nMinimization means collecting only what serves a stated purpose. A useful method is a purpose test for every field. Write the field name, the decision it informs, and who will use it. If the answer is \"we might want it someday,\" the field fails.
\nWhat is worth collecting for most brands: contact details the customer chose to share, purchase and order history, stated preferences, channel and campaign source, and engagement with your own emails and site. These directly support better offers, better service, and honest measurement.
\nWhat to delay or avoid. Do not ask for a birth date when a birth month serves the offer. Do not store full precise location when a region does the job. Do not collect anything about health, finances, or sexual life, and stay away from collecting data on children, because family-first is the standard I hold. If your business touches sensitive categories, the safest default is to collect nothing beyond what the transaction requires, and to put counsel in the room before changing that.
\nGovernance is the boring set of rules that keep a good asset from becoming a bad one. It answers who can see what, how long you keep it, what happens when a person says stop, and how you prove you did it.
\nStart with consent records. For each person, store what they agreed to, when, through which form, and which wording they saw. A consent banner that blocks collection until the visitor chooses is the visible half. The record behind it is the half that matters when someone asks questions.
\nNext, suppression. When a customer unsubscribes or opts out, that choice must reach every system that touches the person: the email tool, the audience files you upload to ad platforms, the analytics pipeline, and any personalization tool. A single list of suppressed IDs, checked before every send and every upload, is the simplest design. The common failure is one system honoring the opt-out while another keeps going. Chapter 42 covers privacy compliance as operating discipline, and the point here is only that suppression must be built into the data foundation itself.
\nThen retention. Decide how long each type of data stays useful and delete it on schedule. As an illustration, you might keep order records for the period your accountant requires, keep email engagement for a couple of years, and delete browsing events after a short window. A written retention schedule, with an automated job that enforces it, turns deletion from an emergency into a routine task.
\nFinally, access. Give people only what their job requires, log who exports customer files, and name one owner for the data.
\nYou do not need a six-figure customer data platform. A small team can assemble a sturdy foundation from free, open tools, though I will be plain about the trade: you pay in setup time and in someone who can maintain it.
\nThe pieces map to four jobs. Collection captures events on your site. Matomo, self-hosted, does this and keeps the data on servers you control. Consent is handled by Klaro, an open source consent manager that gates what loads based on the visitor's choice. Storage is a database you own, and PostgreSQL is a dependable choice for a customer table, an events table, and a consent table. Movement and shaping come from two tools: Meltano extracts data from your sources, such as your store, forms, and email tool, and loads it into the database, and dbt-core turns raw tables into clean, tested models, such as one row per customer with last purchase date and consent status.
\nA simple blueprint: raw events and records land in PostgreSQL. Meltano runs on a schedule to pull sources in. dbt-core builds a customer table, a segment table, and a suppression table. Your team reads from those tables to build email lists, ad audience files, and reports.
\nBe honest about the limits. These tools do not give you a hosted dashboard with one-click ad platform syncs, and a full activation platform at large scale has no open source equal. At small and mid scale, exporting a file from your own table and uploading it works fine.
\nAn owned audience is valuable in theory. You still have to show the money. The honest method is a holdout test, meaning you split a group of customers at random, message one half, hold the other half back, and compare results.
\nSuppose a brand has 20,000 customers who bought once but not in the last six months. Chosen at random, 10,000 receive a win-back message and the other 10,000 get nothing. If the messaged group buys at a rate of four percent and the held-back group buys at two percent, the lift is two percentage points. That is 200 orders the message plausibly caused, out of 400 orders the group placed. The other 200 would have happened anyway. The figures are illustrative, but the logic is real: without a holdout, you would credit the campaign with all 400.
\nTest the simplest segments first: buyers due to reorder, lapsed buyers, and engaged non-buyers, each defined by a plain rule in your own code rather than a platform's black box.
\nAlso use your data to improve paid media. Upload a suppression list of current customers so you stop paying to acquire people you already have, but only where a platform's terms and your customers' permissions allow it. Upload a high-value customer list to guide acquisition, with the same caution. Check platform requirements and have counsel review the permissions language first.
\nFirst-party data: Information a business collects directly from its own customers and visitors through channels it controls. Event schema: The agreed list of event names and attached properties that keeps recorded behavior consistent across tools and teams. Deterministic matching: Linking records to one person using an identifier the person provided and confirmed, such as a login or email address. Data minimization: The practice of collecting only the information needed for a stated purpose, and nothing more. Suppression list: A master list of people who must not be contacted or targeted, checked before every send and every audience upload.
\n1. Confusing audience size on a platform with owned audience. A large follower count or lookalike pool lives on the platform and can be repriced or restricted without notice. 2. Collecting every field on the first form. Long forms depress sign-ups and add records you will never use, along with the security and deletion burden attached to them. 3. Skipping the event schema. Teams that let every tool name events its own way spend weeks reconciling data before any analysis can begin. 4. Honoring an opt-out in one system and missing it in another. A customer who unsubscribes from email but stays in an ad audience file will notice, and so will regulators. 5. Crediting a campaign with every order from the messaged group. Without a holdout, ordinary repeat buying gets reported as campaign performance.
\n1. Offer a clear, immediate benefit at the moment of highest intent, and ask for one or two fields first. 2. Write the event schema in a shared document and require a stated purpose before any new event or property ships. 3. Generate your own durable customer ID and match on confirmed identifiers such as login and email. 4. Keep a written retention schedule with an automated deletion job, and one named owner for the data. 5. Run a random holdout on every owned-channel campaign large enough to read, and report incremental orders, not total orders.
\n1. Birth month, region, and favorite team often serve personalization as well as full birth dates or exact locations, with far less risk. 2. Add a short preference question to the third or fourth email instead of a long welcome survey. 3. Hash and test your suppression list before every upload to an ad platform, and keep a log of each upload. 4. Review three random customer records each month from signup to latest purchase. Real examples reveal schema and matching errors that dashboards hide. 5. Put the consent record, the wording version, and the timestamp in the same row. When a question comes, you will answer it in minutes.
\nA small team can assemble the foundation from five free tools. Matomo (GPL) collects site analytics on servers you control. Klaro (BSD) manages the consent banner and gates what loads. PostgreSQL (PostgreSQL license) stores the customer, event, consent, and suppression tables. Meltano (MIT) pulls data from your sources into the database on a schedule. dbt-core (Apache 2.0) turns raw tables into tested, documented customer and segment models. You supply the hosting and the time to maintain them. At large scale, no open tool matches a full-featured commercial activation platform, and I would rather say that than pretend otherwise.
\nPlatform audiences are leased, and landlords reprice. First-party data is the one audience asset a brand fully owns, but it only has value when it is permissioned, clean, and usable. Earn it through a fair value exchange, record it through a consistent event schema, connect it through simple deterministic identity, and limit it through minimization. Govern it with consent records, one suppression list, a retention schedule, and clear ownership. Build the foundation with free, open tools, and prove its worth with holdout tests that count only the orders the activity caused.
\n1. Reach on a platform is rented, and rented reach gets repriced, restricted, or redefined by the owner. 2. A value exchange that gives the customer something useful immediately produces better data than a request for general updates. 3. A short, documented event schema and one internal customer ID make data analyzable instead of messy. 4. Minimization is a risk control: fields you never collect cannot be leaked or mishandled. 5. A holdout test is the only reliable way to show what your owned audience adds beyond orders that would have happened anyway.
\nThis week, write down every field your forms collect, assign each a stated purpose, delete the ones that fail, and build one suppression list that every send and every ad upload must check first.
\n1. Global Privacy Control specification, published through the W3C Privacy Community Group. Technical standard for a browser signal expressing a consumer privacy preference. 2. California Consumer Privacy Act and California Privacy Rights Act regulation text. Definitions of sale, sharing, consent, and consumer opt-out rights. 3. Brodersen, K.H., Gallusser, F., Koehler, J., Remy, N., Scott, S.L. Inferring causal impact using Bayesian structural time-series models. Annals of Applied Statistics. Published methodology for estimating incremental effects. 4. Project documentation for PostgreSQL, Meltano, dbt-core, Matomo, and Klaro. Official guides for self-hosted installation, licensing, and configuration.
"},{"id":"chapter-42-consent-privacy-law-and-compliance-operations","title":"Chapter 42: Consent, Privacy Law, and Compliance Operations","html":"By the end of this chapter, the reader will be able to: 1. Explain which privacy, messaging, children's and disclosure rules attach to a given advertising activity. 2. Design a consent record that can answer a regulator, a partner or a customer months after the click. 3. Configure a banner and tag layer so that ad and analytics tags stay silent until consent exists. 4. Build a request workflow that carries an opt-out or deletion through every system that holds the person. 5. Evaluate a vendor's data processing terms and flag the gaps before signature. 6. Construct a compliance calendar that assigns owners, cadences and evidence for each recurring duty.
\nLet me argue the other side first, because I have heard it in every budget meeting. Compliance is overhead. Regulators chase the giants, not a regional brand spending a few thousand dollars a month. The lawyer will tell us what to do if something happens. A banner from a plugin handles the rest. And every extra consent screen costs us measurable audience, which means every hour spent on privacy is an hour taken from performance.
\nSome of that is true. Banners do cost audience. Counsel does decide the hard calls, and nothing in this chapter replaces counsel. Everything here is operational guidance, not legal advice.
\nHere is where the argument breaks. The expensive failures are rarely the ones a lawyer could have prevented with a memo. They are the ones nobody owned. A customer clicked unsubscribe in March and got a retargeting ad in June because the opt-out lived in one tool and the audience upload lived in another. A banner said \"reject\" while a tag fired anyway. None of these needed a legal theory. They needed a calendar, an owner and a record.
\nThat is the position of this chapter. Compliance is operating discipline, the same kind of discipline that keeps a pit crew from leaving a wheel loose. You do not win the race in the pit, but you lose it there if the routine is sloppy. A brand that runs compliance as routine spends less on emergencies, moves faster with partners who ask for proof, and keeps the first-party data that later chapters depend on. The rest of this chapter builds that routine piece by piece.
\nMost brand managers try to learn privacy law statute by statute. That is a lawyer's way of sorting, and it produces a long list nobody can apply on a Tuesday afternoon. Sort by activity instead. Ask what you are doing, then ask which rules touch it.
\nCollecting data through a website or app touches the European regulation called GDPR if you reach people in Europe, the California statutes known as CCPA and CPRA if you reach California residents, and a growing set of other state privacy laws. They differ in thresholds and wording, but they share a shape: tell people what you collect, honor their rights to see, correct and delete it, and give them a way to stop certain uses. Several treat the sale or sharing of data for cross-context advertising as something a person can refuse. California regulations expect a business to treat a Global Privacy Control signal, a browser setting that tells every site \"do not sell or share my data,\" as a valid refusal. Counsel confirms how each rule applies to your size and footprint.
\nSending email touches CAN-SPAM. Texting or calling touches the Telephone Consumer Protection Act, usually called TCPA. Reaching children touches the Children's Online Privacy Protection Act, COPPA. Paying a creator to promote you touches the Federal Trade Commission's endorsement and disclosure rules. Publishing any page or ad touches accessibility expectations, because people who use screen readers or keyboards are customers too.
\nThis sorting gives every duty an owner. The email manager owns CAN-SPAM. The influencer lead owns disclosure. The analytics lead owns the banner and tag behavior. A statute with no owner is a statute nobody follows.
\nA banner click is not a record. It is a moment. A record is what you can retrieve later to show what a specific visitor saw and chose.
\nConsent that holds up has a few traits. It is given before collection starts, not after. It is specific, meaning a visitor can accept analytics and refuse advertising. It is unbundled from other agreements, so nobody consents to marketing by agreeing to terms of service. Pre-checked boxes do not count as consent under the stricter regimes, and a design that makes refusing harder than accepting invites trouble.
\nNow the record. For every consent event, store a random visitor identifier (not a name or email unless the person is already known), the timestamp, the exact version of the banner text shown, the categories accepted or refused, the page where the choice was made, and the signal source: a click, a Global Privacy Control signal, or a withdrawal. Version the banner text, so you know which visitors saw which sentence.
\nA plain PostgreSQL table handles this well. One append-only table of events, never edited, with a second view showing each visitor's current state. Append-only matters because a ledger that can be quietly changed is not evidence. Give write access to the consent service alone.
\nSuppose a regional outfitter, which I will call Ashgrove Outdoors (an illustrative composite), receives a partner inquiry asking whether a particular email subscriber ever agreed to marketing. With a ledger, the answer takes a short query: yes, on this date, on this form, with this wording, from this source. Without one, the answer is an apology and a long week.
\nThe most common failure I see is a banner that looks right and does nothing. The visitor clicks reject, and the advertising pixel has already fired, because the tag loaded before the banner finished. Litigation in some states has targeted pixels and session replay tools that capture activity before consent, and counsel will advise on your exposure. The engineering fix is simple to state: nothing fires first.
\nKlaro is an open source consent manager under a BSD license. It works by holding back scripts until a visitor approves the category they belong to. You mark each third-party script as belonging to a purpose, such as analytics or advertising, and Klaro keeps those scripts from executing until consent arrives. That design is the right one, because it blocks at the source rather than asking a tag to behave.
\nThree practices make the difference between a banner and a control.
\nFirst, test with the network panel open. Load the page in a fresh browser profile, refuse everything, and watch which requests leave the page. Any request to an ad network, social platform or analytics host after refusal is a defect.
\nSecond, many programs honor Global Privacy Control. If the browser sends the signal, they treat it as a refusal for advertising categories without another click, and log the event with its source. Counsel confirms where it is required.
\nThird, configure Matomo, the GPL-licensed analytics platform, with its privacy settings: anonymize IP addresses, respect browser Do Not Track settings, and set a retention window for raw logs. Collecting less is the cheapest compliance control there is.
\nThe banner itself must work with a keyboard and a screen reader. An inaccessible consent screen is a compliance gap.
\nRights requests fail in the handoffs. The law gives people the right to ask what you hold, to correct it, to delete it and, in many places, to stop its sale or sharing. Deadlines exist, typically measured in weeks. GDPR typically sets about one month, California typically sets 45 days with limited extensions, and counsel confirms the clock for each law that applies to you.
\nThe workflow has five steps, and each needs a named owner.
\n1. Intake: one public form and one email address, both logged with a timestamp. 2. Identity check: proportionate to the request. Verify more for deletion than for an opt-out, and never demand more personal data than you need to confirm the person. 3. Fan-out: send the instruction to every system that holds the person. 4. Confirmation: each system reports back, and a human checks the list. 5. Closure: the requester receives a reply, and the ledger records the outcome.
\nFan-out is where brands break. Suppose a request touches seven systems and the team handles four. That leaves three places where the person can still be reached, and the first one to reach them creates the complaint.
\nBuild a data map before the first request arrives. A spreadsheet works: every system, what personal data it holds, why, who can access it, which vendors touch it, and how long it keeps it. The map is the table of contents for every request you will ever receive.
\nStore each opt-out as a suppression record that outlives every other deletion, because you must remember not to contact someone after forgetting the rest.
\nMessaging rules are narrower than privacy statutes and more mechanical, which makes them easy to operationalize and expensive to ignore.
\nUnder CAN-SPAM, commercial email needs a truthful sender name and subject line, a valid physical postal address, a clear way to unsubscribe, and a promise to honor unsubscribes within 10 business days. Honor them immediately, and make sure suppression reaches every sending tool.
\nTCPA is stricter. Automated marketing texts and certain calls generally require prior express written consent, which means a clear agreement that names your brand and states the purpose. A checkbox buried in a contest form is a weak foundation. Keep the consent language, the timestamp, the phone number and the form version in the ledger. Honor stop words such as STOP at once, and treat quiet hours as a rule rather than a suggestion.
\nHere is a decision rule I give teams. If you cannot show, within one minute, where and when a person agreed to hear from you by this channel, do not send.
\nThe sweepstakes entry form is the danger zone for sports brands. Fans enter for a jersey and the fine print enrolls them in marketing. Let the entry stand alone and offer marketing as a separate, unticked choice.
\nSet a conservative default whenever young people might see or touch your advertising. COPPA applies to online services directed to children under 13 and to services with actual knowledge that a user is under 13. It generally requires verifiable parental consent before collecting personal information from them. Advertising built on behavioral tracking is exactly what the rule targets.
\nThe operational answer is to avoid the problem. Do not target audiences you know skew young with data collection, exclude child-directed placements from retargeting pools, and do not build lookalike audiences from sources that may contain minors. If a campaign must reach families, aim it at parents through contextual placements and keep tracking out of the experience. Ask counsel before launching anything where children are the intended audience.
\nCreators bring a different duty. The FTC's endorsement rules expect a clear and conspicuous disclosure when a creator has a material connection to your brand, such as payment or free product. Vague tags buried in a pile of hashtags, or hidden behind a \"more\" link, are the classic failure. The rules reach the brand as well as the creator, which is why disclosure belongs in your contract and your approval process, not in the creator's good intentions.
\nBuild a one-page disclosure sheet for every creator brief. State the exact wording expected, where it must appear (spoken in the video as well as written in the caption, for video content), and what happens if it is missing. Review the first post of every new creator before it goes live.
\nEvery ad platform, analytics tool, email provider and agency you use processes personal data for you or about you. The contract decides who is responsible for what, and it is usually the least-read document in the stack.
\nWhen a vendor processes data on your behalf, privacy laws generally expect a written agreement covering what the vendor may do with the data, security obligations, help with rights requests, deletion at the end of the relationship, and notice of sub-processors, meaning the other companies the vendor uses. Read the vendor's standard addendum. Counsel decides whether the terms are acceptable.
\nAsk these five. Can the vendor use your data to improve its own products? Does it share data with other advertisers? Where is data stored and who else touches it? How quickly will it tell you about a security incident? And what happens to your data when you leave?
\nSome vendors act as independent controllers of the data rather than your processor, and the legal responsibilities shift accordingly. Ask which role the vendor claims and record the answer.
\nKeep a vendor register: name, purpose, data types, agreement date, renewal date, owner and review date. When a vendor emails a notice of changed terms, the owner reads it and logs the decision.
\nEverything above decays without a schedule. Banners drift when a developer ships a new template. New pixels appear when a coworker pastes a snippet.
\nA calendar converts good intentions into routine. Here is a simple rhythm, to be adapted with counsel.
\n| Cadence | Task | Evidence kept | |---|---|---| | Weekly | Review new rights requests and unsubscribes; confirm all systems closed | Request log | | Monthly | Scan the site for new third-party scripts; refuse-all network test | Test notes and screenshots | | Quarterly | Review vendor register, creator disclosures, and sampled consent records | Signed review sheet | | Twice a year | Update the data map; review banner wording and accessibility | Dated data map version | | Annually | Refresh policies with counsel; retrain every person who touches data | Training roster |
\nAssign each row a person, not a department. A task owned by \"marketing\" is owned by nobody.
\nAdd a change gate. Any new tool, form, pixel or audience source needs a short privacy check before launch: what data, what purpose, which vendor, what consent. Five questions, answered in writing, take 10 minutes and prevent the kind of surprise that takes 10 weeks to unwind.
\nMeasure the routine too. Track the share of requests closed on time, the share of systems covered by suppression, the share of data sources on the map and the number of unplanned scripts found each month. These are control measures, not vanity numbers. If the unplanned script count climbs, your change gate is leaking.
\nRun a drill once a year: submit a test deletion request and watch it travel.
\nConsent ledger: An append-only record of every consent event, including time, wording version, categories and source. Global Privacy Control: A browser signal that tells sites a person refuses the sale or sharing of their data for advertising. Data processing addendum: A contract add-on that sets what a vendor may do with personal data on your behalf. Suppression record: A stored instruction not to contact or target a person, kept even after other data is deleted. Data map: An inventory of the systems holding personal data, with purpose, owner, vendors and retention for each.
\n1. Treating the banner as the whole program, so tags load before the visitor chooses and the choice never reaches the other systems. 2. Letting an opt-out live in one tool, so retargeting pools and uploaded audiences keep reaching people who asked to stop. 3. Bundling marketing consent into contest entries, terms of service or checkout, which leaves no clean record of agreement. 4. Skipping disclosure instructions in creator contracts and discovering the gap after the post is live. 5. Signing vendor terms without reading the data use clause, then learning that your event data feeds someone else's product.
\n1. Sort rules by activity and give every activity a named owner with a written duty. 2. Store consent as an append-only ledger with banner wording versions, and test the query that answers a partner's question. 3. Block scripts until consent arrives, honor browser privacy signals automatically, and verify with a refuse-all network test every month. 4. Build the data map first, then run every rights request as a five-step workflow with confirmation from each system. 5. Gate every new tool, form and pixel with a short written privacy check before launch.
\n1. Ask counsel for a one-page summary of which rules apply to your footprint, then pin it where the team works. 2. Record screenshots with each monthly test, because a dated image settles disputes faster than memory. 3. Use contextual placements, which target the page rather than the person, when your audience may include children or when consent rates are low. 4. Put an unsubscribe and a stop word in every outbound message, and test both with a real address and phone. 5. Rehearse an incident once a year: who calls counsel, who tells partners, who pauses campaigns, and who writes the customer notice.
\nThree open source tools cover the core of this chapter without a subscription. Klaro (BSD license) is a consent manager that blocks third-party scripts until a visitor approves their category. A PostgreSQL consent ledger, using the open source PostgreSQL database, stores append-only consent events, suppression records and request logs under access controls you set. Matomo privacy settings (GPL license) let you run analytics with anonymized IP addresses, respected Do Not Track settings and limited retention, which reduces the data you must protect. All three run on your own infrastructure, so the records stay yours.
\nCompliance works when it is treated as routine rather than rescue. Sorting rules by activity gives each duty an owner. A consent ledger turns a click into evidence, and a banner built to block scripts turns evidence into control. Rights requests, opt-outs and unsubscribes succeed only when they travel through a mapped set of systems with confirmation. Email, text, children's and creator rules each demand a specific habit, and vendor contracts decide who carries risk when data moves. A calendar, a change gate and a yearly drill keep the routine alive. Counsel decides the legal questions, and operations make those decisions stick.
\n1. Organize compliance by the activities your team performs, and give each one a named owner. 2. A consent record needs a visitor identifier, timestamp, wording version, categories and source to count as evidence. 3. Test the banner by refusing everything and watching the network, because appearance proves nothing. 4. Opt-outs only work when every system holding the person receives the instruction and confirms it. 5. Vendor terms, creator disclosures and messaging permissions need written records and scheduled reviews.
\nThis week, refuse every cookie on your own homepage with the network panel open, list each request that still leaves the page, and assign each one to a named owner with a fix date.
\n1. Federal Trade Commission. Guides Concerning the Use of Endorsements and Testimonials in Advertising, and the accompanying disclosure guidance. 2. Global Privacy Control specification, published through the W3C Privacy Community Group, current version. 3. General Data Protection Regulation, official text. Lawful basis, consent conditions and data subject rights. 4. California Consumer Privacy Act and California Privacy Rights Act statute and regulation text. Notice, deletion, opt-out of sale or sharing. 5. Federal Trade Commission. CAN-SPAM Act: A Compliance Guide for Business, and COPPA business guidance. 6. Federal Communications Commission. Telephone Consumer Protection Act consumer and business guidance.
"},{"id":"chapter-43-tagging-server-side-collection-and-data-layers","title":"Chapter 43: Tagging, Server-Side Collection, and Data Layers","html":"By the end of this chapter, the reader will be able to: 1. Describe how a data layer, a tag manager, and a destination platform divide the work of measurement. 2. Design an event naming convention and schema that survives a site redesign. 3. Apply change control to a tag container so that no untested edit reaches production. 4. Evaluate what server-side collection fixes and what it leaves untouched, including consent. 5. Execute a repeatable quality assurance routine on a new or changed tag. 6. Build a monitoring routine that catches tag breakage before a bidding algorithm learns from bad data.
\nHere is an illustrative scene. Ashgrove Outfitters, a fictional regional retailer, redesigns its checkout. The developer, doing a perfectly sensible thing, renames the purchase event to something cleaner and more descriptive. Nobody tells the media buyer. The new page looks better, loads faster, and converts about the same. Launch day goes well.
\nEleven days later the paid search lead notices that spend is flat but reported sales have fallen by half. The finance team says revenue is steady. The platform's automated bidding, fed almost no conversions, has been pulling back on every campaign. It did exactly what it was told to do with the data it was given. The data was wrong, and nobody was watching the pipe.
\nMost brand managers assume tagging is plumbing: a technical chore that belongs to the web team and only matters when someone asks why a dashboard looks odd. That assumption is expensive. Tagging is the point where your business decides what the advertising platforms are allowed to know. Every optimization, every audience, every report you will ever show a finance partner is downstream of a handful of small code snippets and the naming choices behind them.
\nA second assumption deserves equal suspicion: that moving collection to a server solves the problem. It solves some problems. It creates others. It does not, under any circumstances, replace the permission a visitor gave or declined to give.
\nThis chapter treats tagging the way a head coach treats the playbook. The plays only work if everyone calls them by the same name, the changes get reviewed before game day, and somebody is watching the field when something breaks. You do not need to write the code. You need to know what to demand, what to test, and what to watch.
\nA data layer is a structured object on each page that holds facts about what is happening: which page this is, which product was viewed, what the order total was. The website writes the facts into it. The tag manager, a tool that decides which snippets run and when, reads from it and passes the values on to analytics and advertising platforms.
\nWhy bother with the middle step? Because without it, tags scrape the page. They read a price from a visible label, or an order number from a web address, and the moment a designer rearranges the page, the scraped value changes or vanishes. A data layer separates what the business knows from how the page happens to look. Redesign the page all you like. The contract stays put.
\nThink of it as the official scorer at a ball game. The scoreboard graphics change every season. The scorer's book follows the same rules every night.
\nFor a brand manager, the practical consequence is ownership. Someone has to own the contract, and it should not be the person who happens to be most annoyed this week. In a small business that owner may be you. In a larger one it is a named analytics lead who signs off on every change. Either way the contract is a written document, stored somewhere everyone can find it, with an example payload for every event.
\nKeep three categories of fact in the layer: page context, event facts, and visitor facts, the last only in forms you have the right to hold and only after permission rules apply. If a fact will not drive a decision or a report, leave it out. Every extra field is another thing that can leak or break.
\nNames look trivial and are not. If one team writes \"add_to_cart\" and another writes \"AddToCart\" and a third writes \"cart-add,\" your reports will count three different behaviors, and every join between systems will fail quietly.
\nPick a convention and write it on one page. A sound default: lowercase, underscores, an object followed by a past tense action (product_viewed, cart_updated, order_completed). No marketing slang, no names tied to a campaign. A name should still make sense two redesigns from now.
\nA schema goes one step further. For each event it states the required properties, their types, and their allowed values. Consider an order_completed event. It might require a transaction identifier (text, unique per order), an order total (a number, not text with a currency symbol attached), a currency code, and a list of items, each with an identifier, quantity, and unit price. Required means required: if the field is missing, the event is incomplete, and incomplete events should be flagged, not silently accepted.
\nTwo properties earn special attention.
\nThe transaction identifier is your deduplication key. When the same purchase is reported twice, as happens when a visitor reloads a confirmation page, the identifier lets any downstream system recognize the repeat and count it once. Without it you inflate conversions and teach the bidding algorithm to overpay.
\nCurrency and units need explicit declaration. A total of 1,500 could be fifteen dollars in cents or fifteen hundred dollars. Decide once, state it in the schema, and test for it.
\nVersion your schema, and never quietly alter the meaning of an existing field.
\nA tag container is a live production system. Treat it like one.
\nIn many accounts the container grows by accretion: a pixel for a campaign that ended long ago, a custom script nobody can explain. Each leftover tag slows the page, widens the privacy exposure, and adds another way for a conversion to be counted wrongly.
\nGovernance is the set of habits that prevents this. Keep it light enough that people will actually follow it.
\nAccess: Give edit rights to few people and publish rights to fewer. Agencies get named accounts, never a shared login.
\nEnvironments: Maintain separate development, staging, and production versions of the container. Tests happen in the first two. The third receives only what was approved.
\nReview: Every change gets a short written request stating what is changing, why, which events it touches, and who tested it. A second person reads it before publishing. A trusted colleague spending ten minutes is enough.
\nVersion notes: Every published version carries a note in plain language. When something breaks in March, you want to read exactly what changed in February.
\nQuarterly audit: For each tag, ask who requested it and what decision it supports. If nobody can answer, pause it, and remove it if nothing complains for a month.
\nVendor requests deserve suspicion. Ask what data a new snippet receives and where it goes. This is operational guidance only: your counsel and your contracts decide what is permitted.
\nStandard tagging runs in the visitor's browser. Each platform's code loads there, reads what it can, and sends data directly to its own servers. Server-side collection adds a step. The browser sends one stream of events to an endpoint you control. That endpoint, running on infrastructure you operate or rent, decides what gets forwarded to which platform and in what form.
\nWhat it genuinely helps with:
\nWhat it does not solve:
\nA useful way to hold this: server-side collection is a better loading dock, not a different warehouse. It improves how goods move through. It does not change what you are permitted to store.
\nGet the browser-side contract clean first. Move to a server endpoint when you have the people to keep it running.
\nSuppose a fictional home goods brand, Linden and Pike, sees its reported online orders drift away from its order system. Illustrative figures: the order system records 200 orders in a week, while the ad platform reports 150 and web analytics reports 230.
\nThe gap in two directions suggests two different faults, which is why you should not average them.
\nStart with the undercount. Visitors who decline tracking or use blocking tools will not appear, and that explains part of it. It does not explain all of it. A test purchase in staging reveals that the confirmation page fires the conversion tag before the data layer has finished loading, so on slower connections the order total arrives empty and the platform discards the event. The fix is ordering: the tag waits for a data layer event, not for page load.
\nNow the overcount. Analytics reports 230 against 200 real orders, a sign of duplicates. Reloading the confirmation page in staging fires the event again. The transaction identifier is present, but analytics was never configured to use it for deduplication. The fix is a rule in the destination plus a one-time guard so the page marks an order as reported.
\nAfter both fixes, a week later the numbers read 200 orders in the system, about 185 in the platform, and 198 in analytics. They still do not match perfectly, and they should not be expected to. Consent declines, blockers, and timing differences leave a gap that is known, bounded, and stable. A stable gap you can explain is far more valuable than a perfect match you cannot trust.
\nQuality assurance for tags is not glamorous, and it is the cheapest insurance in this book. A repeatable routine takes about an hour for a typical change.
\n1. Read the request. Confirm what should change and which events it touches. 2. Use the tag manager's preview mode in staging, never on live traffic first. 3. Walk the actual user path from add to cart through a test purchase, because most breakage appears between steps. 4. Inspect the data layer at each step. Confirm event names, required properties, types, and the transaction identifier. Browser developer tools show this directly. 5. Inspect the outgoing network requests. Confirm each destination receives the expected values and that nothing extra is being sent. 6. Test the negative cases. Decline consent and confirm nothing fires that should not. Reload the confirmation page and confirm no duplicate. Remove a required field and confirm the event is flagged. 7. Test on a phone and a slow connection, where timing faults hide. 8. Record what was tested, by whom, and the result, then publish with a version note.
\nRun the same path again after any site release, template change, checkout update, or consent banner adjustment. Most tag breakage is caused by a change somewhere else.
\nA tag does not announce that it has stopped working. The page still loads, the site still sells, and the only symptom is a number that quietly gets smaller or stranger. Monitoring turns that silence into an alert.
\nWatch five signals.
\nVolume: Count events by type each day. A sudden drop, or a sudden spike, deserves a look. Compare with the same weekday in prior weeks, since traffic has a weekly shape.
\nCompleteness: For each event, measure how often required properties are present. A purchase event arriving without a total is a failure even though the event technically exists.
\nDuplicates: Count repeated transaction identifiers. A rising rate points to a page or script change.
\nErrors: Track rejected or failed requests to each destination where the platform exposes them.
\nReconciliation: Each week, compare order counts and revenue from your own order system against what each platform reports. Track the gap as a ratio over time. The size of the gap matters less than a sudden change in it.
\nSend alerts to a named person, not to a shared inbox nobody reads. Alerts that fire constantly get ignored, which is worse than having none.
\nKeep a short incident log: when it started, when it was found, the cause, and how long data was affected.
\nTell the people who buy media. When a conversion signal was broken for three days, annotate that window so nobody makes a budget decision off a measurement fault.
\nYou can run a capable collection and monitoring setup without paying for software. The tools are a tag manager, an event collector, and a database, all self-hosted. The trade is your time: installation, updates, security patches, and backups are your responsibility. Be honest about that cost before choosing it.
\nWhere does this stop? Expect to configure some destinations by hand, since commercial templates will be missing. For large operations needing audience activation across many platforms, open source options do not match commercial products, and I will say so plainly. For most brand managers with a defined set of events, they are sufficient.
\nData layer: A structured object on a page that holds the facts about a visit and its events, read by the tag manager rather than scraped from the visible page. Tag manager: A tool that stores and controls which tracking snippets run, under which conditions, with version history and preview. Server-side collection: A setup in which the browser sends events to an endpoint you control, which then decides what to forward to each platform. Deduplication key: A unique identifier, such as an order number, that lets systems recognize and discard a repeated report of the same event. Schema: A written definition of each event's required properties, their types, and their allowed values.
\n1. Letting the website team rename events or change checkout flow without telling whoever owns measurement, which silently breaks conversion reporting. 2. Believing a server-side endpoint changes consent obligations, when the visitor's permission state must still be honored on every forwarded event. 3. Testing a tag on one page instead of the full path from product view to completed purchase, which misses the timing faults between steps. 4. Leaving old tags in the container indefinitely, which slows the page and widens the privacy exposure for no decision value. 5. Expecting every platform to report the same count, then discarding good data when it does not, instead of tracking a stable, explained gap.
\n1. Write the data layer contract as a one-page document with an example payload for every event, and store it where marketing and engineering can both reach it. 2. Require a transaction identifier on every purchase event and configure each destination to use it for deduplication. 3. Keep development, staging, and production containers separate, with a second reviewer and a plain-language version note on every publish. 4. Reconcile platform conversion counts against your own order system every week and chart the gap as a ratio over time. 5. Pass the consent state with each event and make the collection endpoint enforce it, rather than relying on the browser alone.
\n1. Ask every agency and vendor exactly which fields their tag reads, and refuse snippets that cannot answer clearly. 2. Prefix test orders so they can be filtered out, and delete them from any audience the platform builds from conversions. 3. Schedule a tag path re-test after every site release, because the most common cause of broken tags is an unrelated change. 4. Add a measurement annotation to your reporting calendar for every outage, so a broken week is never compared as though it were clean. 5. Collect only fields tied to a named decision, since each unused field is risk without return.
\nMatomo Tag Manager (GPL) manages tags, triggers, and published versions on infrastructure you control, with preview and version history built in. Jitsu (MIT) collects events and forwards them to destinations you configure, so it can serve as the first-party endpoint described above. PostgreSQL (PostgreSQL license) stores the raw event stream, which lets you run completeness, duplicate, and reconciliation queries against your own data. Together these cover collection, control, and storage with no license fees. Self-hosting means you handle security, updates, and backups, and none of these tools settles consent questions for you.
\nTagging decides what advertising platforms are allowed to know, so it deserves the discipline of any production system. A written data layer contract keeps measurement stable through redesigns. A plain naming convention and versioned schema keep reports joinable. Change control, with separate environments and a second reviewer, prevents untested edits from reaching live traffic. Server-side collection gives control, lighter pages, and durability, but it does not replace consent, repair bad data, or run itself. A routine of full-path testing, steady monitoring, and weekly reconciliation turns tag breakage from a silent loss into a short, logged incident.
\n1. A data layer is a written contract between the website and the tag manager, and it needs a named owner. 2. Event names and schemas should be plain, versioned, and written down before anyone builds anything. 3. Every container change needs a second reviewer, a separate test environment, and a version note. 4. Server-side collection improves control and page weight, yet it never replaces the visitor's consent decision. 5. Reconcile against your own order system weekly, and aim for a stable explained gap, not perfect agreement.
\nA site release is only done when one person's name sits on the data layer contract and a full purchase path test is scheduled for the day after launch. Then start a weekly reconciliation of platform conversions against your own order records.
\n1. World Wide Web Consortium. Web Content Accessibility Guidelines 2.1. Standard relevant to consent banners and tag-driven page elements. 2. Interactive Advertising Bureau Tech Lab. Transparency and Consent Framework technical specifications. Published signal format for passing consent state. 3. Matomo documentation. Matomo Tag Manager user guide. Container, trigger, and versioning reference. 4. Jitsu project documentation. Event collection and destinations reference. Open source (MIT) collector. 5. PostgreSQL Global Development Group. PostgreSQL documentation. Reference for event storage and query design.
"},{"id":"chapter-44-attribution-models-and-their-failure-modes","title":"Chapter 44: Attribution Models and Their Failure Modes","html":"By the end of this chapter, the reader will be able to: 1. Name the direction of bias in first click, last click, linear, time decay, position based and data-driven models. 2. Explain why a platform-reported conversion count and a finance ledger rarely agree. 3. Compare multi-touch attribution, marketing mix modeling and controlled experiments by what each can and cannot see. 4. Build a side-by-side attribution comparison from one set of conversion paths. 5. Evaluate a channel's reported credit against an experiment result and decide how to adjust budget. 6. Design a short uncertainty briefing that a finance partner will accept.
\nPicture the monthly review. Your dashboard says paid social drove 1,250 sales, search drove 1550, and email drove 900. Add the three up and you get 3700. The finance partner across the table has the bank deposits, and they cover 2,100 sales. Nobody in the room is lying. The dashboard is simply counting the same sale more than once.
\nHere is the assumption I want to challenge: that attribution is a measurement problem you solve by finding the right model. It is not. It is a bookkeeping rule, a decision about who gets credit when several channels touched the same buyer. Choosing a rule does not discover the truth. It only decides which channels will look good.
\nThat is why the better question is never \"which model is correct?\" The better question is \"in which direction is each model wrong, and by roughly how much?\" A scale that always reads heavy is still useful if you know it reads heavy. A scale that reads randomly is useless, and a scale whose owner sells you the groceries is suspect. Every attribution model falls somewhere on that spectrum, and none of them reads true.
\nThink of a football broadcast crediting a touchdown. The announcer credits the receiver who caught it. The film room credits the lineman whose block opened the lane, the coordinator who drew the play, and the quarterback who read the coverage. Each account is a defensible story, and each one hands the trophy to a different person. Attribution models work the same way.
\nChapter 1 explained why digital results are trackable but not truly measurable. This chapter extends that point into practice. We will catalog the known biases, show how three different instruments cross-check one another, build a working comparison, and then handle the hardest audience of all: the person who signs off on the budget and reads the ledger.
\nStart with the six common rules and their tilt.
\nLast click gives all credit to the final touch before purchase. Its bias runs toward channels that sit close to the sale: branded search, retargeting, email to existing customers. It rewards whoever stands at the checkout door and ignores whoever sent the shopper there.
\nFirst click does the reverse. It gives all credit to the opening touch, so it favors prospecting channels such as video, social and display. It flatters whatever introduced the buyer, even if that buyer would have found you anyway or never intended to buy.
\nLinear splits credit equally. It sounds fair, but its bias is toward channels that show up often, because frequency earns share. A cheap retargeting ad shown six times collects six slices.
\nTime decay weights recent touches more heavily. Its tilt resembles last click, softened. Long consideration cycles, such as a home renovation or a business software purchase, get shortchanged on their early stages.
\nPosition based (often called U-shaped) typically gives 40 percent to the first touch, 40 percent to the last and splits the remaining 20 percent across the middle. The weights are a guess made by whoever built the tool. Middle touches such as comparison content, reviews and nurture email are structurally undervalued.
\nData-driven models use statistical methods to estimate each touch's contribution from observed paths. This sounds like an escape from guesswork. It is only a partial one, for reasons covered next.
\nNone of these rules measures cause. All six only redistribute credit among touches that already happened to a person who already converted. A buyer who would have purchased with no advertising at all still generates credit. That shared flaw is the deeper bias, and it points the same way for every model: toward overstating the total contribution of advertising and making measured channels look better than they are.
\nWhen a platform offers a data-driven model, ask who built it, who runs it and who benefits from its output. The answer is the same company that sells you the media.
\nA platform's model can only see what the platform sees. It observes impressions and clicks inside its own system, and it has limited sight of a competitor's ad, a podcast read, a store visit or a conversation between friends. So a platform's data-driven report tends to be sound at distributing credit among its own touches and silent about everything outside them. Credit that should be shared with outside channels gets absorbed inside.
\nThere is a second problem. Many platforms count a conversion when a person saw an ad and later converted, even without a click. Suppose a shopper sees your video ad on a Tuesday, never clicks, and buys on Friday after a search. The video platform may claim the sale. So may the search platform. So may your email tool. Each report is honest under its own rules, and the sum of the reports exceeds reality. That is the 3,700 versus 2,100 problem from the opening.
\nData-driven models are also hungry. They need a large volume of conversions with varied paths before the estimates settle. A small account running a few hundred sales a month is effectively feeding the model too little data, and the output can swing from week to week for reasons unrelated to performance.
\nNone of this means ignoring platform reports. They are excellent for in-platform optimization: comparing one ad set against another inside the same system under the same rules. They are poor referees between systems. Use them as a ruler for tuning, not as a verdict on budget.
\nEven a perfectly fair model inherits the limits of its inputs. Three input problems appear in nearly every account.
\nThe lookback window is the period before conversion in which a touch can earn credit. If your buyers typically take 45 days from first visit to purchase and your window is 30 days, every buyer path loses its first third. The model is not wrong about what it saw. It was simply handed a truncated story. Calibrate the window by pulling actual time-to-purchase from your orders, then set the window near the point where most buyers have converted.
\nDevices break paths. A person sees your ad on a phone during lunch and buys on a laptop that night. Without a shared login or a first-party identifier, the two visits look like two strangers. Phone-heavy upper-funnel channels lose credit and the laptop session's last touch collects it. This is another reason last click tilts toward direct visits and branded search.
\nOffline touches vanish entirely. A trade show conversation, a sales call, a billboard or a recommendation from a coach never enters the path. A business with a long sales cycle can have most of its decisive moments invisible to every model on this list.
\nPrivacy rules and browser restrictions add gaps on top of these. Operationally, treat any digital path as an incomplete sample. If you capture consented first-party identifiers where the law and your own policy allow, the paths improve, but gaps remain. Your counsel decides what is permitted; this is operational guidance only.
\nNo single instrument sees everything. Three together get closer.
\nMulti-touch attribution (the path-based approach above) works at the individual level. It is fast, granular and cheap to refresh. Its weakness is that it only sees tracked touches and cannot separate causation from correlation.
\nMarketing mix modeling is a statistical approach that relates total sales over time to total spend by channel, plus outside factors such as season, price and promotions. It needs no individual tracking, so privacy changes barely touch it, and it can include offline media. Its weaknesses are coarse resolution, slow refresh and a need for enough spending variation. If you spend the same amount every week in every channel, the model cannot separate the channels from one another.
\nControlled experiments include geographic holdouts, audience holdouts and spend on-off tests. You withhold advertising from a comparable group and compare outcomes. Experiments are the closest thing to a cause-and-effect answer. Their weaknesses are cost, duration and narrow scope: each test answers one question for one channel at one moment.
\nThink of a baseball scouting department. Video is fast and detailed but only shows what the camera caught. The stat sheet sees a full season but averages away context. The tryout puts a player in a controlled situation and shows what he can do. No scout relies on one alone, and each one checks the other two.
\nThe practical rule: use path-based reports weekly for tuning, use a mix model quarterly for budget direction, and use experiments to settle the largest disputes between the other two. When all three agree, act confidently. When they disagree, the disagreement tells you where to test next.
\nYou do not need new software to start. You need one table of conversion paths and the patience to replay it under several rules.
\nSuppose a fictional outdoor retailer, Marrow Creek Gear (illustrative), logs 10 orders in a sample, each with an ordered list of touches. Four paths run video, then social, then branded search. Three run email, then branded search. Two run display, then retargeting. One runs branded search alone.
\nReplay these under last click and branded search receives eight of 10 orders, retargeting receives two, and video, social, display and email receive none. Replay them under first click and video receives four, email three, display two, branded search one. Under linear, credit spreads: video and social earn about 1.3 orders each from the first group, and branded search collects a mix across groups.
\nNow read the pattern. Branded search ranks high under last click and low under first click. That is the signature of a channel that captures demand someone else created. Video ranks low under last click and high under first click, the signature of a channel that introduces buyers. Neither reading proves cause. But the gap between the two columns is a map of where to look.
\nTo build this for real, export paths from your analytics tool into a database such as PostgreSQL, then replay them with a few short queries in DuckDB. Each rule is a small piece of logic: pick the first touch, pick the last, split evenly, or apply weights. Store the output in one table with a column per model and one row per channel. Then a dashboard tool can display the columns side by side.
\nOnce the columns sit side by side, rank channels under each. Three patterns recur.
\nA channel that ranks high everywhere is probably real, or at least stable under every rule. A channel that ranks high only under last click is a candidate demand capturer. A channel that ranks high only under first click needs a skeptical eye, because introducing a person is not the same as winning them.
\nThe gap is where experiments belong. Suppose Marrow Creek Gear's table shows branded search taking 80 percent of credit under last click. Run a holdout: pause branded search ads in a comparable set of regions for four weeks and leave the rest running. If sales in the paused regions barely drop, because shoppers click the organic listing instead, then most of that credit was demand that would have arrived anyway. If sales drop sharply, the channel was doing real work, perhaps defending against competitors who bid on your name.
\nBoth results are useful. The point is that the experiment answers something no model can answer from paths alone. For a small business with few regions, a simple on-off test on one channel for several weeks, with seasonality in mind, still beats guessing. Keep the test short and the question narrow.
\nRecord every experiment result next to the model columns. Over a few quarters you build a calibration table: how much each model overstated or understated each channel when compared to a test. That table becomes your internal correction factor, and it is worth more than any vendor report.
\nA finance partner does not want a philosophy lecture. They want a number they can defend and a plain statement of how firm it is.
\nStart by reconciling. Show platform-reported conversions, your analytics count and the ledger side by side, and explain that the first two overlap while the third is the anchor. Finance will trust you faster when you volunteer the gap than when they discover it.
\nThen present a range instead of a single figure. For example: \"Paid social contributed between 300 and 600 incremental sales last quarter. The low end comes from our holdout test. The high end comes from our path-based report. We plan to narrow it with a second test.\" A range with a stated source for each end sounds more credible than a precise figure with no source.
\nUse three labels: measured by experiment, estimated by model, and assumed. Mark every number accordingly. Finance people live with estimates and assumptions every day, from forecasts to depreciation schedules. They only object when the label is missing.
\nFinally, tie requests to decisions. Do not ask for budget because a model says a channel works. Ask for a defined test, a cost, a duration and a rule for what you will do with each possible result. That framing shifts the conversation from defending a dashboard to buying information. It also protects you: when an experiment kills a favorite channel, the decision was agreed in advance.
\nIncrementality: The sales that happened because of advertising and would not have happened without it, as distinct from sales that were merely credited to it.
\nLookback window: The period before a conversion during which a touch is eligible to receive credit.
\nHoldout: A comparable group of people or regions deliberately kept from seeing advertising, used as the baseline in an experiment.
\nView-through conversion: A conversion credited to an ad that was seen but not clicked.
\nMarketing mix modeling: A statistical method that relates total sales over time to spend by channel and outside factors, without tracking individuals.
\n1. Treating the sum of platform reports as total contribution, when overlapping credit makes the sum exceed actual sales. 2. Crediting branded search with growth, when it often collects demand that video, social, audio or word of mouth created. 3. Setting a lookback window shorter than the real sales cycle, which erases the first stages of the buyer's path. 4. Switching models whenever a result looks disappointing, which turns attribution into a tool for justifying decisions already made. 5. Running a holdout without checking that the test and control groups are comparable, which produces a clean-looking but meaningless result.
\n1. Replay the same conversion paths under at least four rules and keep every column visible in one table. 2. Pull time-to-purchase from your own orders and set the lookback window from that evidence, then review it every quarter. 3. Reconcile every reporting period to the finance ledger before presenting channel numbers. 4. Run at least one holdout per year on your largest channel, and record how far each model missed. 5. Label every number in a report as measured, estimated or assumed.
\n1. Ask any vendor offering a data-driven model what data it cannot see, and who audits its output. 2. When a channel looks great only under last click, check what share of its conversions come from people who already searched your brand name. 3. Keep spending variation in your plan on purpose. A mix model cannot learn from a budget that never changes. 4. Pre-register the decision rule before a test starts: write down what you will do if the result is high, low or flat. 5. Treat view-through credit as a lead to investigate, not as proof, and reduce its weight when comparing channels across platforms.
\nYou can build the entire comparison with free software. Matomo (GPL) collects first-party site analytics that you host yourself, including conversion paths. PostgreSQL stores the exported path data in tables you control. DuckDB (MIT) runs fast analytical queries on those tables or on plain files, which makes replaying paths under different rules a short script. Metabase (AGPL) turns the output into a shared dashboard where model columns sit side by side. No open source option here replaces a full commercial marketing mix platform for large budgets, so for that job plan on a custom statistical workflow built by someone with the skills, or skip it and rely on experiments.
\nAttribution is a rule for sharing credit, not an instrument for finding cause. Last click favors the bottom of the funnel, first click favors the top, linear favors frequency, time decay favors recency, position based favors the weights its builder guessed, and data-driven models favor whatever the platform can see. Windows, devices and offline touches add further gaps. Path-based reports, mix models and controlled experiments cover each other's blind spots, and the disagreements among them show where to test next. A side-by-side comparison built from your own paths, reconciled to the ledger and presented as a labeled range, gives finance something it can defend.
\n1. Every model is wrong in a direction you can name, so name it before you present a number. 2. A platform's own reports are good for tuning inside that platform and poor for judging it against others. 3. Path-based reports, mix models and experiments each see something the others miss, so use all three at different speeds. 4. The gap between first click and last click columns marks the channels worth testing. 5. Finance accepts uncertainty when it is labeled, ranged and tied to a decision rule.
\nYour budget review starts with one quarter of conversion paths replayed under four rules, the total reconciled to the ledger, and the channel with the widest gap between columns picked for a four-week holdout test.
\n1. Brodersen, K.H., Gallusser, F., Koehler, J., Remy, N., and Scott, S.L. Inferring causal impact using Bayesian structural time-series models. Annals of Applied Statistics. Published methodology for estimating the effect of an intervention without a randomized control. 2. Media Rating Council. Outcomes and Data Quality Standard. Industry accreditation guidance on measurement and outcome reporting. 3. Interactive Advertising Bureau and Media Rating Council. Guidelines for viewability and impression measurement. Industry definitions for what counts as a served and seen ad. 4. Matomo, PostgreSQL, DuckDB and Metabase project documentation. Official documentation for the open source tools named in this chapter.
"},{"id":"chapter-45-incrementality-and-geo-experiments","title":"Chapter 45: Incrementality and Geo Experiments","html":"By the end of this chapter, the reader will be able to: 1. Define incrementality and explain why attributed conversions and incremental conversions are different quantities. 2. Compare audience holdouts, geo holdouts, matched markets, synthetic control and platform lift studies, and choose one for a given budget. 3. Estimate in plain language whether a planned test has enough power to detect the effect the business needs. 4. Write a decision rule before launch and apply it to results without bending it. 5. Analyze an illustrative geo test and present the lift with an interval to a finance partner. 6. Evaluate a test plan for contamination, short windows and peeking.
\nExperiments are the only honest answer to the question \"did this advertising cause those sales?\" Nothing else comes close. Dashboards report who saw an ad and later bought. Models guess at what would have happened otherwise. A properly run holdout simply removes the advertising from a slice of the world and watches what happens.
\nRunning one well is harder than it sounds, and the objections are real. Most brand managers cannot run a clean experiment. The budget is small, the markets are uneven, the quarter is short, and the executive team wants an answer by Friday. A badly run experiment is worse than no experiment, because it carries the authority of the word \"test\" while delivering noise. I have watched a team cut a profitable campaign because a two week go-dark test happened to overlap a regional holiday. The test said the ads did nothing. The calendar said otherwise.
\nSo I will not tell you that every brand needs a lab. I will tell you something narrower. Any campaign that consumes a meaningful share of your budget deserves one honest test per year, designed with a decision attached to it. Not a test to learn something interesting. A test whose result moves money.
\nIncrementality is the difference between what happened with your advertising and what would have happened without it. The second half of that sentence can never be observed directly. Nobody gets to live the same week twice. Every experiment is therefore an attempt to build a believable stand-in for the world where the ads did not run.
\nConsider a retailer running brand search ads. Many of the people who click were already typing the retailer's name. They would have found the site through the free listing a line below. The platform counts those purchases as conversions, and it is not lying. The shoppers did click and did buy. What it cannot tell you is how many would have bought anyway. That number is the baseline, and the baseline is what experiments exist to estimate.
\nThe useful way to think about this is in three layers. Total outcomes are everything that happened: all orders, all visits. Attributed outcomes are the slice a platform or model assigns to your ads. Incremental outcomes are the slice that would have vanished if the ads had been off. Incremental is always less than or equal to total, and it can be far lower than attributed, especially for channels that reach people who were already close to buying.
\nThis is why the ratio matters more than any single number. Divide incremental conversions by attributed conversions and you get an incrementality factor. A factor near one means the platform's count is roughly honest. A factor of one half means half the credited conversions would have happened anyway. A factor above one is possible too, when ads create sales that tracking cannot see, such as a store visit after a video view. Whatever the factor turns out to be, it is a property of that campaign, that audience and that period.
\nThere are five experiment designs worth knowing. Each trades cost, precision and practicality differently.
\nAudience holdout. The platform or your own system randomly withholds ads from a set share of eligible people, say 10 percent, while the rest see the normal campaign. Randomization at the person level is the strongest design in principle, because the two groups differ only by chance. Its weakness is where it can run. It needs a system that controls delivery to individuals, and it needs a clean way to link outcomes back to those individuals. For online conversions inside one platform, this is often the easiest option. For sales that happen in stores, it is hard.
\nGeo holdout. You withhold advertising from some geographic areas and keep it running in others, then compare outcomes by area. Because the unit is a place and not a person, it works for any outcome you can count by place: store sales, call volume, orders by delivery zip code. It also avoids the privacy and tracking problems of person-level data. The cost is fewer units. You may have 40 markets instead of four million people, so noise is a serious problem.
\nMatched markets. A refinement of the geo holdout. Before launch, you pair markets that behave similarly, then assign one in each pair to treatment and one to control, usually by coin flip. Pairing shrinks noise because each pair shares seasonality, weather and local events.
\nSynthetic control. Instead of finding one twin for each treated market, you build a weighted blend of many untreated markets whose history, added together, tracks the treated market closely before the test. After launch, the blend is your forecast of what the treated market would have done. CausalImpact, the open source package named in the callout, works in this spirit using a statistical model of the pre-test relationship. Synthetic control suits a small number of treated markets, or one, where pairing is impractical.
\nPlatform lift studies. The advertising platform runs the holdout itself and reports the lift. They are convenient and often free to the advertiser. They are also a case of the seller grading its own homework. Treat them as one input, check what the platform defines as a conversion, and, where budget allows, confirm with a test you control.
\nPower answers one question before you spend a dollar: if the advertising really works at the level the business needs, will this test be able to see it?
\nImagine weighing yourself on a scale that jumps by two pounds every time you step on it. If you lose half a pound, the scale cannot tell you. If you lose 10 pounds, it can. Your test has a scale too, and its jumpiness comes from how much the gap between treatment and control moves around from week to week with no advertising at all. You can measure that jumpiness from history. Take the weeks before the test, compute the gap between your would-be treatment and control groups, and look at how much that gap wobbles.
\nThree things shrink the wobble relative to the effect: more units, a longer test, and better matching. Three things you want from the result: a lift large enough to matter financially, enough confidence to act, and a low chance of missing a real effect. As a rule of thumb, the smallest lift a test can reliably detect is about two and a half times the standard error of the estimate. Suppose the weekly gap wobbles by five percent of the baseline. Averaged over a six week test, the standard error shrinks to about two percent (five divided by the square root of six). The smallest lift the test can reliably see is then roughly five percent.
\nNow ask the uncomfortable question. Is five percent lift enough to justify the campaign? If the campaign only pays back at a three percent lift, this test cannot answer your question. You have three choices: run longer, add markets, or accept that the experiment cannot settle the decision and do not run it. That last choice is rarely popular, but it beats spending money on an answer that will be noise.
\nBefore the test launches, write down what you will do under each outcome. Put it in a document, date it, and send it to the person who controls the budget. This is pre-registration, and it is the single cheapest protection against fooling yourself.
\nA good decision rule has five parts: 1. The primary outcome (orders, sales, store visits) and the one window it will be measured over. 2. The breakeven: the highest cost per incremental outcome the business can afford. 3. The confidence level for the interval, chosen now and not after. 4. What happens when the interval sits entirely on the profitable side, entirely on the unprofitable side, or straddles breakeven. 5. Who decides, and by what date.
\nThe straddle case is where most teams get stuck, so decide it ahead of time. A sensible default is a partial move: increase spend in measured steps and retest, instead of betting the full budget on a point estimate.
\nEverything below is illustrative, with round numbers chosen to show the logic. Real tests will be messier.
\nA fictional running shoe brand, Ostrander Footwear, wants to know whether its paid social video campaign drives orders or merely harvests shoppers who were coming anyway. Its platform dashboard attributes 2,700 orders to the campaign over the test period.
\nSetup. The brand has 24 delivery markets. Using 16 weeks of order history, the analyst pairs markets by how closely their weekly orders move together, producing 12 pairs. Within each pair a coin flip assigns one market to run the campaign and the other to go without. Before launch, the analyst checks that the treatment group and control group had nearly the same total orders in the pre-period: about 3,000 per week each. The weekly gap between the groups wobbles by about five percent.
\nPre-registered rule. Breakeven is 45 dollars of ad spend per incremental order, because the contribution margin on an average order covers that much. The interval will be at the 90 percent level. If the entire interval for cost per incremental order sits below 45 dollars, scale. If the lower end of the lift interval is above zero but the cost interval straddles 45 dollars, raise spend by 25 percent and retest. If the lift interval includes zero, hold spend flat, and do not extend the test. The test runs six weeks, with a two week cooldown to catch late orders, and the result will be read once, at the end.
\nResult. The treatment markets spent 54000 dollars on the campaign. Predicted orders in those markets, based on the control markets and the pre-period relationship, were 18000 over six weeks (3,000 per week). Observed orders were 19260. That gap is 1,260 incremental orders, a lift of seven percent. With a standard error of about two percent, the 90 percent interval on the lift runs from roughly 3.6 percent to 10.4 percent, or 650 to 1.9 thousand incremental orders.
\n| Measure | Point estimate | Low end | High end | |---|---|---|---| | Incremental orders | 1,260 | 650 | 1.9 thousand | | Cost per incremental order (dollars) | 43 | 83 | 29 |
\nThe platform's own count implies about 20 dollars per order (54000 divided by 2,700). The experiment says 43 dollars at the midpoint. The incrementality factor is 1,260 divided by 2700, about 0.47. Roughly half of what the dashboard claimed was real.
\nDecision. Lift is clearly above zero, so the campaign works. The point estimate of cost sits under breakeven. But the low end of the interval, 83 dollars, is well above breakeven, so the cost interval straddles it. The pre-registered rule says: raise spend by 25 percent and retest. Not scale hard, not cut. Notice that nobody had to argue. The rule had already done the arguing.
\nNever show a lift as a single number. A point estimate of seven percent sounds precise. An interval of 3.6 to 10.4 percent tells the truth: the real effect is probably in that range, and we cannot say more.
\nWhen you present to a finance partner or a skeptical executive, use this order. First the decision the test was built to inform. Second the result in the units they care about, dollars per incremental order. Third the interval, in plain words: \"We are 90 percent confident the true cost is between 29 and 83 dollars.\" Fourth the decision rule you committed to and what it triggers. Fifth the limits: one market type, one season, one creative set.
\nBe careful with the phrase \"statistically significant.\" It only says the interval excludes zero. It says nothing about whether the effect is big enough to matter. A lift of one percent can be significant with enough data and still lose money.
\nExperiments fail in a small number of predictable ways, almost all of them preventable.
\nContamination. Treatment leaks into control, or the reverse. Shoppers in a control market see ads served to a neighboring treated market. A regional television buy overlaps with your control area. A national promotion lands in only some groups. In geo tests, choose markets with little commuting between them, and check that no other media buy hits one side only.
\nShort windows. A test that ends before the buying cycle does measures only the fastest buyers. If customers take three weeks to decide on a purchase, a two week test will understate the effect. The reverse also occurs: a test running through a promotion may overstate it. Align the window to the buying cycle and add a cooldown period after spending stops.
\nPeeking. Checking the result every day and stopping when it looks good. Noise swings above and below the truth, and a stopping rule based on the current number will eventually catch a lucky swing. Set the end date in the pre-registration and read once. If you must monitor, watch for operational problems such as delivery failures, not for the lift.
\nPoor matching. Control markets that never tracked the treatment markets in the pre-period will give a forecast that drifts. Always plot the pre-period fit. If the lines do not move together before launch, they will not be a reliable guide after it.
\nIncrementality: the portion of an outcome that would not have happened without the advertising. Holdout: a group of people or places deliberately kept from seeing the advertising so they can serve as the comparison. Counterfactual: the estimate of what would have happened in the treated group had the advertising not run. Minimum detectable effect: the smallest lift a test can reliably tell apart from random noise. Pre-registration: writing and dating the primary outcome, window and decision rule before launch so results cannot be reinterpreted afterward.
\n1. Running a test with no decision attached, so that whatever the result says, nothing changes and the spend is wasted. 2. Launching a short test in a noisy market set and treating a wide interval as proof that the advertising does not work. 3. Pairing markets by size or region name instead of by how closely their pre-period outcomes actually moved together. 4. Checking results daily and ending the test the moment the number looks favorable. 5. Treating one campaign's incrementality factor as a permanent constant and applying it to other channels, audiences or seasons.
\n1. Estimate power from your own pre-period data before launch, and cancel the test if it cannot detect the lift the business needs. 2. Write the decision rule, including the straddle case, and share it with the budget owner before spending starts. 3. Randomize assignment of markets within matched pairs so that your own preferences cannot bias the groups. 4. Align the test window to the real buying cycle and include a cooldown period for delayed conversions. 5. Report every lift with its interval, in dollars per incremental outcome, and state the limits of what the test covered.
\n1. Start with the largest line item that you cannot currently defend, because the first test pays for itself fastest where doubt is greatest. 2. Keep a calendar of local events, promotions and other media buys for each market, since the best contamination check is a plain list. 3. Retest on a schedule, because audiences, creative and competitors change and old results grow stale. 4. When a platform lift study and your own geo test disagree, trust the one whose design you control, and ask the platform to explain the gap. 5. Hand finance the spreadsheet with the pre-period fit, the interval and the rule, so they can reproduce the logic themselves.
\nYou can run every analysis in this chapter with free, open source software. CausalImpact (Apache 2.0) builds a forecast of the counterfactual from control series and reports the lift with an interval, which suits synthetic control designs. statsmodels (BSD) provides regression, time series and confidence interval routines for matched market analysis and power calculations. scipy (BSD) supplies the statistical tests, distributions and optimization tools underneath. A spreadsheet can handle the simplest paired comparison, but these three cover the rest. For a full production-scale experimentation platform, no open source option reaches parity, and for most brands a script and a clear document are enough.
\nIncrementality asks what the advertising caused, not what it touched, and the only reliable way to answer is to remove the advertising from a comparable slice of the audience or map and watch what happens. Audience holdouts, geo holdouts, matched markets, synthetic control and platform lift studies each fit different situations, and the right choice depends on what you can randomize and where outcomes are counted. Power calculations tell you before launch whether the test can detect the lift the business needs, and a pre-registered decision rule turns the result into an action without argument. The worked example showed a campaign that did work but cost roughly double what its dashboard claimed, and a rule that called for a measured step up and a retest. Contamination, short windows, peeking, weak matching and mid-test changes are the errors that most often ruin a design.
\n1. Incremental outcomes are the ones that would have vanished without the ads, and they are often well below the count a platform reports. 2. Pick the design that matches what you can randomize: people, markets in pairs, or a few markets with a synthetic comparison. 3. Check power before launch, because a test that cannot see the needed lift is a waste of budget. 4. Write the decision rule first, including what to do when the interval straddles breakeven. 5. Present every result with an interval, in dollars per incremental outcome, and read it once at the end.
\nPick your single largest unproven channel this quarter, write a one page decision rule with a breakeven cost per incremental order, and run one matched market test against it before you approve next quarter's budget.
\n1. Brodersen, K.H., Gallusser, F., Koehler, J., Remy, N., Scott, S.L. Inferring causal impact using Bayesian structural time-series models. Annals of Applied Statistics. Methodology behind the CausalImpact software (Apache 2.0). 2. Vaver, J., Koehler, J. Measuring Ad Effectiveness Using Geo Experiments. Google research technical report. Describes geo experiment design and analysis. 3. Abadie, A., Diamond, A., Hainmueller, J. Synthetic Control Methods for Comparative Case Studies. Journal of the American Statistical Association. Foundational paper on synthetic control. 4. Kohavi, R., Tang, D., Xu, Y. Trustworthy Online Controlled Experiments. Cambridge University Press. Practical guidance on power, peeking and experiment pitfalls. 5. statsmodels and scipy project documentation. Reference for confidence interval, power and statistical test routines (both BSD licensed).
"},{"id":"chapter-46-marketing-mix-modeling-with-open-tools","title":"Chapter 46: Marketing Mix Modeling with Open Tools","html":"By the end of this chapter, the reader will be able to: 1. Explain what a marketing mix model can and cannot answer, and state the questions it is the wrong tool for. 2. Assemble a weekly modeling dataset and set a refresh cadence that fits a real team. 3. Describe adstock (lingering effect) and saturation (diminishing returns) in plain language and spot when a model has them wrong. 4. Apply experiment results as priors to pull a model toward observed incrementality. 5. Diagnose collinearity, thin history, and overfit before a model reaches a budget meeting. 6. Evaluate Meridian, Robyn and PyMC-Marketing against a team's skills, and convert model output into a staged budget move.
\nStart with a small piece of arithmetic. Suppose you spend 100 dollars on a video campaign in a single week, and suppose each week keeps 60 percent of the previous week's effect. Week one delivers 100 units of influence. Week two delivers 60. Week three, 36. Week four, about 22. Keep going and the pile adds up to 250 units, from a single 100 dollar push. Now imagine a report that only looks at sales inside the same week. It sees 100 of the 250 and calls video a 40 percent performer relative to its true effect. Cut the budget based on that report and you have starved a channel that was working.
\nThat is the quiet danger of measurement built on short windows, and it is the reason marketing mix modeling (MMM, a statistical method that relates total weekly sales to total weekly spend by channel) keeps returning to the table. It works on aggregate numbers, so it does not need cookies, device identifiers or user-level tracking.
\nHere is the assumption I want to challenge. Many teams believe a free open source modeling tool gives them a free answer. It does not. The software costs nothing. The answer costs months of data work, a person who understands the statistics, and the discipline to test the model against reality. A model fitted in an afternoon on messy spend files will produce charts with decimal points and confident labels, and every one of those numbers can be wrong.
\nA model output is not a verdict. It is closer to a well-informed scout report: it tells you where to look and how much to bet, and it earns trust only after you check it against a game you actually watched.
\nAn MMM answers questions of the form: across many weeks, when spend in a channel went up or down, what happened to the outcome, after accounting for season, price, promotions and other forces? It can estimate how much of your sales would likely have arrived with no paid media at all (the baseline), how quickly returns flatten as spend grows, and where the next dollar probably does the most good.
\nIt cannot tell you which ad, audience or keyword worked. It cannot settle questions at a fine grain, such as whether one creative beat another, because there is not enough variation in weekly totals to say. It cannot prove cause by itself either. A model sees that two things moved together. It does not see why. If you raised search spend every December and sales also rose every December, the model has to guess how to split the credit, and a guess is all it is.
\nSo treat the MMM as the top layer of a measurement stack. Experiments give you ground truth on a narrow question. Platform reports give you fast, biased signals. The model gives you a slow, wide view across everything. When the three disagree, that disagreement is the most useful finding you will get, because it shows you where to run the next test.
\nRefresh monthly or quarterly. Weekly refreshes invite you to chase noise.
\nThink of the dataset as a table with one row per week. Columns hold the outcome (sales, leads, orders, whichever your business treats as the real result), spend by channel, and the other forces that move the outcome.
\nThe outcome should be the one finance already trusts. If revenue in your model does not match revenue in the accounting system, expect the first meeting to be about the data and not the findings.
\nSpend should be split finely enough to be useful. A single column called \"digital\" hides the very differences you are trying to learn. Search, paid social, video, display and creator placements deserve separate columns, and offline channels need the same treatment. Where a channel has two quite different modes, such as brand search and non-brand search, separate them. Brand search often just collects people who were already coming, and lumping it in with discovery search flatters the whole channel.
\nThen come the control variables. These include price changes, promotions, product launches, distribution changes, holidays, weather where it matters, competitor activity if you can find a proxy, and any big event such as a stock-out. Leave out an important control and the model will hand its credit to whichever channel happened to move at the same time.
\nHow much history? As a working rule, aim for at least two years of weekly data, so the model sees each season more than once, and three is better. That is 104 to 156 rows. The number sounds large until you compare it with the number of things you want to estimate. Each channel needs several parameters, and rows are the scarce resource.
\nBudget six or more weeks for what I call data archaeology: tracking down who owns each file, reconciling definitions, fixing gaps. Finance owns revenue, the brand team owns television, the agency owns digital, and nobody has ever lined them up. This work is dull and it is most of the job.
\nTwo ideas carry most of the weight in any MMM, and both are simple once you strip the vocabulary.
\nAdstock is the lingering effect of advertising. The arithmetic in the introduction was adstock: each week keeps a share of the one before. The share is called the decay rate. Channels differ. A search click is nearly instant, so its decay is small. A television flight or a sponsorship can echo for weeks. If your model assumes everything works instantly, it will undervalue slow channels and overvalue fast ones, and the error always leans toward the channels that are easiest to see.
\nSaturation is the idea that each additional dollar buys less than the last. Think of a stadium. The first thousand tickets sell to your most devoted fans. By the time you are filling the upper deck on a Tuesday night, every seat costs more effort to fill. Ad spend works the same way. A channel has a curve that rises steeply, bends, and eventually flattens. The tools describe this curve with a shape parameter and a half-saturation point, the spend level at which you have earned half the maximum effect.
\nWhy does this matter for a brand manager? Because average return and marginal return are different numbers, and budgets should move on the marginal one. Suppose, for illustration, a channel at 40,000 dollars a week averages a return of three to one, but the model shows its curve has nearly flattened, so the next 10,000 dollars returns closer to one to one. A smaller channel averaging only two to one might still be on the steep part of its curve and return well over that on added spend. The channel with the better average is the wrong place to add money.
\nIf your analyst cannot show where each channel sits on its curve, the model is not ready for a budget decision.
\nA prior is what the model believes before it sees your data. In a Bayesian model (one that combines prior beliefs with evidence and reports a range of plausible answers instead of a single number), priors are explicit. In others they hide inside default settings. Either way, they shape the result, especially when data is thin.
\nDefaults are not neutral. A tool that assumes every channel has a similar return will pull weak signals toward the average. A tool that allows any return at all, with no limits, may give you a channel with a negative effect from a noisy stretch of data. Neither is wrong by nature. Both need to be a conscious choice.
\nThe best source for better priors is an experiment. Suppose a geographic holdout test (a test where some regions see the ads and matched regions do not) found that paid social drove roughly a certain lift in orders. You can feed that result to the model as a prior on social's return, with the test's uncertainty attached. The model then has to reconcile its weekly patterns with something that was actually measured. This is called calibration, and it is the single most effective guard against a confident wrong answer.
\nUse the experiment's range, not its midpoint, so the model knows how sure you are. Match the time period and conditions as closely as you can. If the calibrated and raw models disagree sharply, show leadership both and say which you trust.
\nThree problems cause most failed models.
\nCollinearity means two inputs move together so closely that the model cannot separate them. The classic case is a brand that raises every channel in the fourth quarter and cuts every channel in January. The model sees one pattern and several suspects. The result is unstable credit: the model may assign most of it to search in one run and to video in the next, with nearly identical fit. Wide ranges on channel returns are the warning sign. The fix is not statistical. It is operational. You need spend to vary independently, which means planned pulses, staggered flights, or regional differences that you design on purpose.
\nToo few cycles is the second problem. With one year of data, the model cannot tell a seasonal bump from an advertising effect, because it has seen the season once. If you are young, either collect more history or lean harder on experiment priors and say plainly that the model is early.
\nOverfit is the third. A model with many knobs and few rows can match history almost perfectly and predict nothing. It has memorized the past. The check is simple: hide the most recent eight to 13 weeks, fit on the rest, and see how well the model predicts the hidden stretch. Good fit on training data is not evidence. Good prediction on weeks it never saw is.
\nAlso sanity check the baseline. If the model says paid media drives almost all of your sales, or almost none, ask whether that fits what you know.
\nThree open tools cover most needs, and they differ more in philosophy than in power.
\nRobyn is an R package that automates much of the search for good settings and presents a set of candidate models. It is approachable and quick to get running. The risk is that automation makes it easy to pick the model that looks best on a chart and skip the thinking. Treat the candidates as a menu to question, not a ranking to obey.
\nMeridian is a Python library built around Bayesian modeling and designed for geographic-level data, where you have weekly rows for many regions rather than one national series. More rows per week means better separation of channels and tighter ranges. It also has direct support for adding experiment results as priors. If you can supply data by region, it is the strongest option for the question of where returns are well identified.
\nPyMC-Marketing is a Python library built on a general-purpose Bayesian toolkit. It is the most flexible, and the most demanding. You can shape the model to match your business, including custom controls and unusual channel behavior, but you need someone who can read and critique a probabilistic model.
\nTwo cautions apply to all three. First, two of these tools come from companies that sell advertising, and their defaults may reflect that. The code is open, which is a real advantage, because you can read what the assumptions are. Read them. Know who owns the data layer and ask whether that creates a conflict in the modeling layer. Second, the software does not do the hard parts. No tool in this family reaches the level of a full managed measurement service in convenience, and that is fine, because convenience is not the goal. A model you understand and can defend beats a faster one you cannot.
\nMatch the tool to the team: Robyn for an R-fluent analyst, Meridian for regional data and Python skills, PyMC-Marketing for a statistician who wants control of every assumption.
\nA model is only useful if it changes a decision, and it should change it carefully.
\nBegin with the question the business is actually asking. Is it how to split a fixed budget, how much to add, or whether to cut something? Each calls for a different read of the response curves. For a fixed budget, you compare marginal returns across channels and shift money from the flattest curve toward the steepest. For added budget, you look for the channel with the most headroom. For a cut, you find the channel whose loss would hurt least.
\nThen size the move to the confidence. Suppose, for illustration, the model suggests shifting 15 percent of the display budget into video, with a wide range on video's return. A cautious operator moves a third of that, about five percent, for one flight and watches. A model that is both wide and untested earns a small step. A model that is narrow and agrees with an experiment earns a bigger one.
\nThird, make the move a test. If you shift budget into a channel and nothing else changes, you have just made a natural experiment. Record the date, the size, and what you expected. Next quarter's model gets a fresh signal on that channel, and your priors get better. This is how a model improves over time: its recommendations create the variation that sharpens its next estimate.
\nFourth, report ranges. Tell leadership that video's return is likely between a low and a high figure, and that the plan holds across that range or it does not. Executives can handle uncertainty. What they cannot forgive is a precise number that turns out false.
\nFinally, keep every recommendation inside the spend range the data has actually seen. Outside it, a model is guessing.
\nAdstock: A way of modeling the lingering effect of advertising, where each week carries over a share of the previous week's impact. Saturation: The point at which added spend in a channel produces smaller and smaller gains, shown as a curve that flattens. Prior: The belief a Bayesian model holds before it sees your data, which can be set from experiments or outside knowledge. Posterior: The range of plausible answers a Bayesian model reports after combining its prior with your data. Calibration: Adjusting a model so its estimates agree with the results of real experiments.
\n1. Fitting a model on one year of data and treating its seasonal patterns as advertising effects. 2. Lumping all digital spend into one column, which erases the differences the model was built to find. 3. Accepting default priors without reading them, then presenting the output as neutral. 4. Moving budget on average return instead of marginal return, which pushes money into channels that are already saturated. 5. Quoting a single point estimate to leadership and leaving out the range, so a wide model sounds certain.
\n1. Spend the first six weeks on data reconciliation with finance before any modeling begins. 2. Hold out the most recent eight to 13 weeks and judge the model on how well it predicts them. 3. Calibrate with at least one geographic or holdout experiment and record the test's range, not only its midpoint. 4. Refresh monthly or quarterly, and record every budget move so the model learns from your decisions. 5. Show response curves with current spend marked on each, so every conversation is about the margin.
\n1. Plan spend pulses and staggered flights on purpose, because a model cannot separate channels that always move together. 2. Split brand search from non-brand search, since brand search often harvests demand created elsewhere. 3. Read the default settings of any tool built by an advertising company, and ask who gains from them. 4. Run an occasional model with a key control removed and see which channel absorbs the credit. 5. Size a budget shift to the model's confidence: small for wide ranges, larger when experiments agree.
\nThree libraries cover the work, and all three are free to use and run on your own machine. Meridian (Apache 2.0) is a Bayesian library that works best with region-level data and accepts experiment results as priors. Robyn (MIT) is an R package that automates the search for model settings and returns a set of candidates to review. PyMC-Marketing (Apache 2.0) is a flexible Python library for building a custom Bayesian model. No open tool matches the convenience of a managed measurement service, so budget for analyst time and training, which is where the real cost sits.
\nMarketing mix modeling relates weekly outcomes to weekly spend, so it works without user-level tracking, but it answers broad allocation questions and not questions about individual ads. It needs clean weekly data, at least two years of history, finely split channels and honest controls. Adstock and saturation explain why short windows mislead and why marginal return, not average return, should guide budgets. Experiments supply the priors that keep a model tethered to reality, while collinearity, thin history and overfit are the failures to check first. Meridian, Robyn and PyMC-Marketing each fit a different team, and each needs a person who reads the assumptions. Budget moves should be sized to confidence, reported as ranges, and treated as tests that improve the next model.
\n1. A marketing mix model sees aggregate weekly patterns, so use it for allocation across channels and never for judging a single ad. 2. Adstock means spend keeps working after the week it runs, and ignoring it undervalues slow channels. 3. Move money on marginal return, using response curves with current spend marked. 4. Calibrate with experiments and validate on weeks the model never saw before trusting any output. 5. Size each budget shift to the model's confidence and log it, so the move doubles as the next test.
\nNever move more than a third of any recommended shift until an experiment backs it, and start by asking your analyst to show the response curve for your largest channel with current spend marked on it.
\n1. Jin, Y., Wang, Y., Sun, Y., Chan, D., Koehler, J. Bayesian Methods for Media Mix Modeling with Carryover and Shape Effects. Google Research publication. 2. Chan, D., Perry, M. Challenges and Opportunities in Media Mix Modeling. Google Research publication. 3. Meridian project documentation. Open source Bayesian marketing mix modeling library, Apache 2.0. 4. Robyn project documentation. Open source marketing mix modeling package, MIT license. 5. PyMC-Marketing project documentation. Open source Bayesian marketing mix modeling library, Apache 2.0.
"},{"id":"chapter-47-dashboards-and-a-single-source-of-truth","title":"Chapter 47: Dashboards and a Single Source of Truth","html":"By the end of this chapter, the reader will be able to: 1. Explain why a dashboard and a source of truth are different things and why the first cannot replace the second. 2. Write a metric dictionary entry that fixes a definition, a formula, a grain, an owner and a version. 3. Apply a reconciliation routine that ties platform-reported spend and revenue to the general ledger and records the gap. 4. Design one dashboard per recurring decision, with a named decision owner and a defined refresh rhythm. 5. Evaluate the cost of vendor lock-in in measurement and judge how an owned stack changes agency and platform negotiations. 6. Plan a small self-hosted measurement stack using the open source tools in this chapter.
\nMost teams believe a dashboard is the single source of truth. It is the most common assumption in marketing measurement, and it fails the first time two people open two dashboards in the same meeting.
\nHere is how it plays out. The paid media manager shows a conversion count from the ad platform. The analytics lead shows a different count from the website tool. The finance partner shows a revenue figure that matches neither. The meeting becomes an argument about whose screen to believe, and the budget decision waits another week.
\nThe screen was never the problem. A dashboard is a window. It shows whatever sits behind the glass, and if three windows look into three different rooms, you get three different views no matter how well the charts are drawn. A source of truth is the room itself: one agreed place where the raw records land, one set of written rules that turns them into numbers, and one reconciliation back to money that the business actually received and spent.
\nThere is a second assumption sitting beneath the first, and it is expensive. Many brands assume that reporting is a service they rent, bundled with the platform or the agency, and that the rented view is good enough. It is good enough until you want to leave, or negotiate, or audit. Then you find that the history, the definitions and the logic all live in someone else's system. This chapter is about owning the room, not just decorating the window.
\nTwo dashboards disagree for reasons that are boring and fixable. Learn the short list and you can diagnose most arguments in minutes.
\nThe first reason is a definition difference. One report counts a conversion when a form is submitted. Another counts it when the lead is accepted by sales. Both are reasonable, and both are called \"conversions.\" The second reason is a time difference. A platform may credit a sale to the day of the ad click, while the ledger books it on the day the order shipped. Over a month those two calendars produce different totals even with perfect data. The third is a filter difference: one view excludes internal test orders, refunds or a particular region, and the other does not.
\nThe fourth reason is the quietest. Some figures are not facts at all but modeled estimates, such as conversions a platform infers for people it cannot observe. Chapter 44 covers why every attribution model leans in a known direction, so here I will say only that a modeled figure and a counted figure should never share a column without a label.
\nThe fifth is plain breakage. A tag stops firing after a site release, an export fails silently, a connector times out, and a chart quietly flatlines or, worse, keeps showing yesterday's number.
\nA trustworthy measurement stack has three layers, and it helps to keep them physically and mentally separate.
\nThe first is the raw layer. This is where data lands exactly as it arrived: platform exports, order records, lead records, call logs, email engagement files. You do not fix anything here. You keep the original so that any later question (\"what did the platform actually send us on that day?\") has an answer. Raw data is evidence, and evidence is not edited.
\nThe second is the modeled layer. Here the raw records are cleaned, joined and shaped using written rules. Duplicates are removed. Campaign names are standardized. Spend from five platforms is stacked into one table with the same columns. Orders are linked to the leads that produced them where the data allows. Every rule in this layer is code in a file, not a manual spreadsheet step that lives in one analyst's head.
\nThe third is the metric layer, which is a plain name for the place where each business measure is defined once. Revenue, qualified lead, customer acquisition cost, blended return on ad spend. A dashboard tile does not calculate its own version. It asks the metric layer.
\nThe metric dictionary is the most valuable document in this chapter, and it costs nothing but attention. It is a shared file, one entry per metric, that settles every definition argument before it starts.
\nEach entry needs six fields. The name, in plain words. The business question it answers. The exact formula, including what is excluded. The grain, meaning the level at which it is calculated, such as per day, per campaign or per order. The owner, a single named person who can approve a change. And the version, with a date-free change note describing what changed and why.
\nConsider an entry for \"qualified lead.\" Suppose a business sells installed home equipment and runs paid search, social and a call center. The dictionary says: a qualified lead is a contact with a verified phone or email, in a serviceable area, who has either booked an estimate or spoken with a representative for at least 90 seconds, excluding employees, existing customers inside a 12 month window and records flagged as test. Grain is per lead per day. Owner is the sales operations manager. Version is v2, with a note that v1 counted any form submission and inflated the number whenever a promotion drew casual clicks.
\nThat small entry does real work. When the paid search agency reports 600 leads and sales reports 420, the dictionary tells you the agency used v1 and sales used v2. The argument ends in a minute, and the conversation moves to the decision.
\nPlatforms report what they believe happened. The general ledger records what the business paid and received. When the two disagree, the ledger wins, because the ledger is where taxes are filed and loans are judged.
\nReconciliation is the routine of tying advertising numbers back to ledger numbers every month and writing down the gap. It is not glamorous. It is also the single practice that most separates teams who are trusted by finance from teams who are tolerated.
\nWork it in two directions. For spend, compare what each platform says you spent against what was actually invoiced or charged to the card. For revenue, compare what the modeled layer says advertising drove against what the books recognized in the same period, and note which portion the modeled layer attributes to paid media versus other sources.
\nHere is an illustrative worked example. Suppose a regional retailer's ad platforms report a combined 83,000 dollars of spend for the month. The ledger shows 80,000 dollars paid. The gap is about three thousand dollars, nearly four percent. Investigation finds that credits for a billing correction and a refunded invalid-traffic charge landed in the ledger but never appeared in the platform export. The fix is not to edit the dashboard. The fix is a documented reconciling item in the modeled layer: platform spend, less credits, equals ledger spend. Next month the gap is explained before anyone asks.
\nRevenue is harder. Suppose the platforms jointly claim 310,000 dollars of attributed sales, while the books show 240,000 dollars of total online sales for the month. The claimed figure exceeds the whole. That is not a bug in the arithmetic. It is the natural result of each platform taking credit for the same order. Your reconciliation must show total ledger revenue as the ceiling, platform claims as a labeled overlap, and your own deduplicated attribution as the working number. A finance partner who sees this table will relax, because you have named the problem before they found it.
\nOperational guidance only here: how revenue is recognized, how credits are booked and how refunds are treated are accounting decisions. Your accountant decides them. Your job is to make the advertising numbers easy to tie to whatever they decide.
\nMost reporting bloat comes from building dashboards around data sources instead of decisions. \"The social dashboard\" tells you nothing about what anyone should do on Tuesday morning. A better test is to ask, for every proposed dashboard, which decision it supports, who makes that decision and how often.
\nThe answers usually produce a short list. A weekly budget allocation view supports the media buyer and the brand manager deciding where next week's money goes. It needs spend, cost per qualified lead and a conversion trend by channel, with the reconciliation status visible in the corner. A creative review view supports the creative lead deciding which ads to retire and which to extend. It needs ad-level results with a clear sample size, so nobody retires a winner after a weekend of thin data.
\nA monthly executive view supports the leadership decision about overall spend. It needs blended cost per acquisition, total ledger-tied revenue and a plain note on anything that moved the definitions. And an operational health view supports whoever owns the plumbing. It needs freshness, event volume against its normal range, pipeline run status and failed tests.
\nEach of these has a named owner, a refresh rhythm and a written purpose at the top of the page. If a dashboard has no decision attached, retire it.
\nGovernance sounds like paperwork, but in measurement it is mostly one idea: changes to a number's meaning must be deliberate, visible and reversible.
\nTreat definitions the way a software team treats code. Keep the dictionary and the modeling rules in a version control system, which is free software that records every change, who made it and why. A change to a definition goes through a proposal and a review by the metric owner. When it is approved, the version number moves, and the change note explains what shifts.
\nThis matters because definitions drift. Someone adjusts an exclusion to fix a one-time oddity, nobody records it, and six months later the trend line has a step in it that no one can explain. The business cuts a channel because it appears to decline, when the decline was a definition change. Versioning turns that mystery into a one-line lookup.
\nWhen a definition must change, do two things. Restate the recent history under the new definition so the trend stays comparable, and keep the old series available under its old version name for as long as anyone relies on it. A step change with no restated history is a trap for every later reader.
\nHere is the idea this book returns to when it matters. Open tools remove vendor lock-in, and a brand that owns its measurement stack owns its negotiating position.
\nConsider what lock-in looks like in practice. An agency builds your reporting in its proprietary workspace. The campaign history, the naming conventions and the attribution logic live inside that workspace. When the contract comes up for renewal, the agency knows something you cannot afford to test: leaving means losing two years of comparable history and rebuilding the logic from memory. That knowledge gives the agency the upper hand in the negotiation.
\nNow flip it. Suppose your own server holds the raw exports, the modeled tables and the dictionary. The agency reports into your stack, using your definitions, and you can recompute their claimed results yourself. A renewal conversation changes character. You can ask for the fee to be tied to a metric you defined and can verify. You can run a competing agency's trial against the same history. You can walk away without losing a day of comparability. An agency that knows this performs differently, because its numbers will be checked against a record it does not control.
\nBe honest about the limits. Owning measurement does not give you the platforms' internal data, and it does not replace their reporting for in-platform optimization. A full demand-side buying system at scale has no open source equivalent at parity, so you will still buy media through marketplaces and exchanges. The claim is narrower and still valuable: your measurement, your history and your definitions should not be hostage to anyone you pay.
\nYou can build this in a sequence that respects a small team's time, and each step produces value before the next begins.
\nStart with the dictionary and a spreadsheet-sized reconciliation. Before touching any software, write the 10 to 15 metrics you actually argue about, and tie one month of spend to the ledger by hand. You will learn more from that exercise than from any tool, and you will discover which definitions are contested.
\nNext, choose a database. PostgreSQL is a free, mature database that handles concurrent users and runs happily on a single modest server. DuckDB is a free analytical database that runs inside a single file on a laptop, which makes it a good place to prototype the modeling rules before moving them to a shared server. Many small teams start with DuckDB for development and PostgreSQL for the shared copy.
\nThen add the modeling tool. dbt-core lets you write each cleaning and joining rule as a small text file of structured query language, test it, document it and keep it in version control. This is where the modeled layer and the tests live, and it is the practical home of your versioned definitions.
\nFinally, put a dashboard tool on top. Metabase is approachable for business users who want to ask questions without writing code, and Apache Superset offers more control for analysts who want finer charting. If your organization plans to modify or redistribute Metabase rather than simply run it internally, its AGPL license has conditions, so ask counsel before you do.
\nSingle source of truth: One agreed location and set of rules from which every report draws its numbers, so reports agree because they read the same records.
\nMetric dictionary: A versioned document that defines each business measure by name, formula, grain, owner and exclusions.
\nGrain: The level of detail at which a figure is calculated, such as per order, per campaign or per day.
\nReconciliation: The routine of tying reported advertising figures back to the general ledger and documenting any remaining gap.
\nData freshness: The time elapsed since a dataset was last successfully updated, used as an early warning that a pipeline has stalled.
\n1. Building dashboards before agreeing on definitions, which produces beautiful charts that disagree with each other and with finance. 2. Letting each dashboard tile calculate its own version of a metric, so a single edit in one place silently changes one report and not the others. 3. Trusting platform-reported revenue as a total, when several platforms routinely claim the same order and the sum can exceed what the books recorded. 4. Changing a definition without restating history or recording the change, which creates trend breaks that get mistaken for real performance shifts. 5. Keeping reporting logic inside an agency or platform workspace, which leaves the brand unable to verify results or leave without losing its history.
\n1. Write the metric dictionary first, keep it short, and give every entry a single named owner and a version. 2. Reconcile spend and revenue to the general ledger every month, set a written tolerance, and record every explained gap. 3. Attach every dashboard to one decision, one owner and one refresh rhythm, and retire any dashboard that cannot name all three. 4. Keep raw data untouched, put all cleaning rules in version-controlled code with automated tests, and show test status on a health view. 5. Require agencies and partners to deliver raw exports to storage you control, in a format you can recompute from.
\n1. Print the last successful refresh time on every dashboard page; it prevents more false alarms than any other single feature. 2. When two numbers disagree, check definition, then calendar, then filters, then modeled versus counted, then breakage, in that order. 3. Show the sample size beside every rate in creative and audience views so thin data stops driving decisions. 4. Ask your finance partner to approve the revenue and spend definitions once; their sign-off is worth more than any chart polish. 5. Before any renewal, recompute the vendor's claimed results from your own records at least a quarter ahead, so you arrive with evidence instead of impressions.
\nEvery tool here is free and self-hostable. PostgreSQL provides a dependable shared database, and DuckDB provides a fast analytical database that runs from a single file, which suits prototyping. dbt-core (Apache 2.0) holds your modeling rules and automated tests as version-controlled text. Metabase (AGPL) offers approachable dashboards for business users, while Apache Superset (Apache 2.0) offers finer control for analysts. Choose one dashboard tool, not both. No open option replaces a full-scale demand-side buying platform, so media is still bought through marketplaces, but measurement, history and definitions can sit entirely under your control. Ask counsel about license conditions before modifying or redistributing any of them.
\nA dashboard is a window, and the source of truth is the room behind it. Dashboards disagree because of definitions, calendars, filters, modeled estimates and breakage, and all of those are fixable with a layered stack: raw records kept as evidence, modeled tables built by tested and versioned code, and a metric dictionary that defines each measure once. Monthly reconciliation to the general ledger turns advertising numbers into figures finance will trust, and one dashboard per decision keeps reporting short and useful. Governance means definition changes are reviewed, versioned and restated. Owning this stack with open tools also changes negotiation, because the brand holds its own history and can verify what agencies and platforms claim.
\n1. A dashboard displays numbers; a source of truth defines and protects them, so build the second before polishing the first. 2. A metric dictionary with named owners and versions ends most reporting arguments before they begin. 3. The ledger is the ceiling: reconcile spend and revenue monthly, and label any platform claim that exceeds what the books show. 4. Every dashboard should serve one decision, one owner and one rhythm, and be retired when it cannot. 5. Owning the measurement stack turns renewals and platform pitches into conversations based on your own evidence.
\nThree things should be on paper before any renewal or platform pitch: the 10 metrics you argue about, last month's spend tied to the ledger, and a vendor commitment to deliver raw exports to storage you own.
\n1. PostgreSQL Global Development Group. PostgreSQL documentation, current release. Reference for the free relational database used as the shared store. 2. DuckDB Foundation. DuckDB documentation, current release. Reference for the embedded analytical database. 3. dbt Labs. dbt-core documentation: models, tests and documentation. Reference for version-controlled transformation code under the Apache 2.0 license. 4. Metabase project. Metabase open source documentation. Reference for self-hosted dashboards under the AGPL license. 5. Apache Software Foundation. Apache Superset documentation. Reference for self-hosted analytics under the Apache 2.0 license.
"},{"id":"chapter-48-forecasting-and-budget-pacing","title":"Chapter 48: Forecasting and Budget Pacing","html":"By the end of this chapter, the reader will be able to: 1. Explain why a forecast for paid media should be stated as a range with a stated confidence, not as a single figure. 2. Describe how a response curve shows diminishing returns and apply it to a budget increase or cut. 3. Calculate cumulative pacing, required daily rate, and projected end-of-flight spend using the formulas in this chapter. 4. Diagnose the common causes of underspend and overspend and match each to a corrective action. 5. Design a reforecast cadence and score forecast quality with error and bias measures. 6. Build a one-page range summary that a finance partner can approve, track and hold you to.
\nA campaign of mine once finished the month at exactly 100 percent of budget, and I printed the report and felt good about it. Then the sales team told me the best stretch of demand had come in the second week, and by the second week my daily budgets had already been throttled to keep the line smooth. I had hit the number and missed the market. The spend was right. The pacing was wrong. And nobody had asked me the one question that mattered, which was whether the dollars landed where the demand was.
\nMost brand managers carry the same assumption I did: that a good plan is a flat plan. Divide the monthly budget by the days in the month, set the daily caps, and watch the line climb at a steady angle. It looks disciplined. It is often the opposite. Demand does not arrive evenly, platforms do not spend evenly, and the last week of any flight is where a flat plan either strands money or burns it.
\nThe second assumption is quieter and does more damage. It is that a forecast is a promise. You tell finance that 60,000 dollars will produce 1,200 leads, and from that moment the number is a debt. Every deviation becomes a conversation about whether you failed. So people forecast defensively, pad the cost, hide the uncertainty and then explain misses for the rest of the quarter.
\nA forecast is not a promise. It is a disciplined statement of what you expect, how wrong you might be, and what you will do when reality moves. Pacing is the control side of that statement: the daily mechanics that keep spend and outcomes inside the range you described. This chapter treats the two as one system. You will learn to forecast with response curves and scenario ranges, pace with formulas you can run in a spreadsheet, manage flights so the end of the period does not surprise you, and explain all of it to the person who owns the budget without flinching.
\nEvery forecast has two parts, a most likely outcome and a spread around it. Beginners publish the first and hide the second. Experienced buyers do the reverse of what feels safe: they lead with the spread.
\nStart with what you actually have. A baseline forecast takes recent performance, adjusts it for the seasonal pattern in your category, and projects forward. For a business with at least a year of weekly history, you can estimate a seasonal index, which is simply how much a given week runs above or below the average week. A week with an index of 1.30 tends to deliver about 30 percent more volume than an average week. A week at 0.80 delivers about a fifth less. Multiply your baseline weekly expectation by each week's index and you have a shape instead of a flat line.
\nThen ask how much that shape could be off. Look at your own past forecasts, or if you have none, at the week to week swings in your results. If weekly conversions in a stable stretch bounced around by roughly 15 percent either way, a forecast that claims precision to the nearest conversion is fiction. The honest statement is \"most likely 850 conversions, with a plausible range of 720 to 980.\"
\nSpending twice as much rarely produces twice the result. A response curve is a picture, or an equation, of how outcomes rise as spend rises. It starts steep, because your first dollars reach the most willing audience and the best inventory. Then it flattens, because later dollars reach people who are less interested, pay more per impression, or see your ad for the sixth time.
\nHere is an illustrative case. Suppose a regional outfitter called Ashgrove sees roughly 800 sign-ups from 20,000 dollars of monthly search and social spend. At 30,000 dollars the account delivers about 1000. At 40,000 dollars it delivers about 1150. The average cost per sign-up at 20,000 is 25 dollars. But the first extra 10,000 dollars bought 200 sign-ups, an incremental cost of 50 dollars each. The next 10,000 bought 150, an incremental cost of about 67 dollars each.
\nThat second number, the marginal cost, is the one that should drive the decision. Finance sees an average of 25 dollars and expects every added dollar to behave the same way. You know the marginal dollar costs more than double. Showing that gap on a simple chart is often the single most useful thing you can bring to a budget meeting.
\nYou do not need a sophisticated model to sketch a curve. Pull weekly spend and weekly outcomes for one channel across stretches where spend moved up or down, plot them, and draw the smooth shape the points suggest. A fitted curve in Python, with a log or power relationship, works too. Be careful about one trap: a curve built from weeks when spend and demand rose together will make the channel look better than it is, because seasonality is doing some of the lifting. Where you can, compare periods that match for season, or let the seasonal index remove that effect first.
\nA scenario is a forecast with a specific story attached. Three is usually enough: a low case, a base case and a high case. The skill is making each one tied to named assumptions instead of arbitrary plus and minus adjustments.
\nFor the Ashgrove example, a base case might assume the spring season index holds, no new competitor bids on branded terms, and the creative in market keeps its current response. That produces roughly 850 sign-ups at 30,000 dollars. A low case assumes a competitor enters and the cost of the same traffic rises, along with creative fatigue setting in by week three. That might land near 720. A high case assumes a strong seasonal peak and a fresh creative set that lifts response, pushing toward 980.
\nWrite the assumptions beside the numbers. Then you can tell finance what to watch: if competitor bid pressure appears in the first two weeks, the plan is sliding toward the low case, and here is the action we take.
\nPacing answers a narrow question: given what I have spent and what days remain, am I on track? Every formula below can be built in a spreadsheet with basic arithmetic. I write them in words because the logic matters more than the cell references.
\nPlanned cumulative spend to date. Add up the planned daily spend for every day from the start of the flight through today. If you used a seasonal shape, the plan is not a straight line, so you sum the shaped daily amounts rather than multiplying a flat figure.
\nPacing ratio. Divide actual cumulative spend by planned cumulative spend. A ratio of 1.00 is on plan. Above 1.00 means you are ahead of plan; below means behind.
\nRemaining budget. Total flight budget minus actual cumulative spend.
\nRequired daily rate. Remaining budget divided by the number of days left. This tells you what you must spend per day from here to finish on plan.
\nRun rate projection. Take the average daily spend of the last seven days, multiply by the days remaining, and add actual cumulative spend. This is where you finish if nothing changes.
\nProjected variance. Run rate projection minus total flight budget. Positive means projected overspend. Negative means projected underspend.
\nNow an illustrative example. Ashgrove has a 45,000 dollar budget for a 30 day flight, planned flat for simplicity at 1,500 dollars a day. On day 12, planned cumulative spend is 18,000 dollars. Actual cumulative spend is 20,700 dollars. The pacing ratio is 1.15. Remaining budget is 24,300 dollars with 18 days left, so the required daily rate is 1,350 dollars. But the last seven days averaged 1,725 dollars a day. Run rate projection is 20,700 plus 18 times 1725, which is 51,750 dollars. Projected variance is 6,750 dollars over.
\nLook at what the numbers tell you. You are 15 percent ahead of plan, but the more urgent reading is that your current run rate would overshoot by 6,750 dollars. The fix is not to slash everything on day 12. It is to find why the daily rate crept up and bring it toward 1,350 over a few days, while checking whether the extra spend bought outcomes that justify keeping some of it.
\nSpend that misses the plan always has a cause. Finding the cause is faster than staring at the symptom.
\nCommon causes of underspend. Targeting that is too narrow for the budget, so the platform cannot find enough eligible impressions. Bids set below what the auction requires. Creative or landing pages stuck in review. Audiences that shrank after a privacy or list change. Schedules or geographic limits that exclude the hours when your customers are online. Daily budgets set so low that campaigns pause early in the day.
\nCommon causes of overspend. Platforms that are allowed to exceed a daily budget on high traffic days, which most major platforms permit within a monthly limit, so check each platform's own rules. Automated bidding that raises bids when it sees strong conversion signals. Several campaigns competing for the same audience and bidding against each other. Budgets that were increased mid-flight without the plan being updated. Delayed billing or reporting that makes yesterday's spend look smaller than it was.
\nA flight is a defined period with a start date, an end date and a budget. Managing it well is partly arithmetic and partly judgment about shape.
\nFront-loading means spending more early. It makes sense for a product launch, an event with a fixed date, or a sale that rewards early awareness. Back-loading means spending more late, which fits when conversion intent builds toward a deadline such as a registration cutoff. Flat pacing fits steady demand. Most flights deserve a deliberate shape, drawn from your seasonal index and your calendar, not the default of equal days.
\nLeave a reserve. I like holding back five to 10 percent of a flight budget unallocated until the midpoint. That reserve lets you fund a channel that is outperforming without begging for a budget change. It also gives you a cushion when a platform overspends on a spike.
\nA forecast made once and never touched is a snapshot of ignorance. Set a rhythm.
\nDaily: check pacing, delivery and budget caps. This is a five minute scan, not a meeting.
\nWeekly: compare actual outcomes to the forecast range, review channel level pacing, and decide whether any trigger from your scenarios has been hit.
\nMonthly or at flight midpoint: formally reforecast. Update the baseline with new data, refresh the response curve, and restate the range for the remaining period.
\nMeasure accuracy with two simple tools. The first is mean absolute percentage error, which is the average size of your misses expressed as a share of the actual result, ignoring direction. If your forecasts are typically off by 12 percent, you can say so and set ranges that reflect it. The second is bias, which is the average of your signed errors. If you have over-forecast five periods in a row, your bias is telling you the model is optimistic, even if each individual miss looked like bad luck. Mean absolute percentage error tells you how big your misses are. Bias tells you which way they lean.
\nWatch for the habit of changing the forecast to match results. A reforecast should use new information about the future, not erase the record of past misses. Keep the original forecast alongside every revision so you can score yourself against it honestly. Sports analysts do this with pregame lines: the closing line is smarter than the opening line, and the gap between them is how you learn.
\nYou can do all of this in LibreOffice Calc with a table of dates, forecast ranges and actuals, plus a column for each error measure. DuckDB can query your exported platform reports with plain SQL when the data grows beyond what a sheet handles comfortably.
\nFinance teams are not hostile to uncertainty. They are hostile to surprise. Your job is to turn a range into something they can plan around.
\nLead with the decision, not the model. \"I am requesting 30,000 dollars for the spring flight. The most likely outcome is about 850 sign-ups. I would not be surprised by anything from 720 to 980.\" That is clear in three sentences.
\nGive the range a name and a meaning. Say what the low case requires, what the high case requires, and how likely each feels. Do not claim a statistical confidence level you did not calculate. \"Our past forecasts landed inside this band most of the time\" is honest if it is true.
\nThis is operational guidance only. Your finance and legal advisers decide how budgets are booked and approved.
\nPacing ratio: Actual cumulative spend divided by planned cumulative spend to date, where a value near 1.00 means on plan. Response curve: A line showing how outcomes change as spend changes, usually rising quickly at first and flattening as spend grows. Marginal cost: The extra cost of the next unit of outcome gained by adding spend, which is usually higher than the average cost. Seasonal index: A multiplier showing how far a given week or month typically runs above or below the average period. Flight: A defined period with a start date, an end date and a budget for a campaign or set of campaigns.
\n1. Treating a flat daily budget as the safe default, which ignores the seasonal shape of demand and often starves the best days. 2. Publishing a single-number forecast with no range, which turns every normal fluctuation into an apparent failure. 3. Judging a budget increase by average cost per result instead of marginal cost, which overstates what the extra money will buy. 4. Waiting until the last week to correct a pacing gap, then pushing large budget jumps that raise costs and weaken delivery. 5. Rewriting the original forecast after the fact so it looks accurate, which destroys the record you need to improve.
\n1. Build the plan from a seasonal shape and a named set of assumptions, and write those assumptions beside each scenario. 2. Calculate the pacing ratio, required daily rate and run rate projection daily during live flights, by channel and in total. 3. Hold a modest reserve of unallocated budget until the flight midpoint so you can fund what works without a new approval. 4. Keep every original forecast on file next to its revisions, and score both with error and bias measures each quarter. 5. Give finance a spend commitment, a reporting rhythm and a three-scenario range, not a promised result.
\n1. When a platform shows spend that looks too high for a single day, check the monthly limit before cutting, since many platforms balance daily overages across the month. 2. Seven day averages smooth out weekday and weekend swings, so use them for run rate projection instead of yesterday's figure. 3. If two campaigns target the same audience, merge them or set exclusions before blaming the platform for overspend. 4. Raise a lagging daily rate gradually over several days so delivery systems have time to find quality inventory. 5. Mark the date and the reason for every mid-flight budget change in the same sheet as the plan, so the pacing math always reflects the real budget.
\nstatsmodels (BSD license) is a Python library for time series work, including exponential smoothing and seasonal models that return prediction intervals, which gives you the range around a forecast. DuckDB (MIT license) runs SQL directly on exported platform reports and CSV files, so daily pacing tables and error measures can be rebuilt in seconds. LibreOffice Calc (MPL license) handles the pacing sheet, the scenario table and the finance summary with ordinary formulas and charts. Together they cover forecasting, data handling and presentation at no cost. For large marketing mix modeling, no tool in this short list matches a full dedicated platform, so keep scenario work simple and transparent.
\nA forecast for paid media is a stated range built from a seasonal baseline, a response curve and a small set of scenarios tied to named assumptions, and pacing is the daily control that keeps spend and outcomes inside that range. The pacing formulas in this chapter, which compare actual and planned cumulative spend, compute the required daily rate and project the end of the flight, can be run in a spreadsheet and applied by channel. Underspend and overspend both have findable causes, and flights work best with a deliberate shape, a small reserve and early correction. A regular reforecast cadence, scored with error and bias, keeps the method honest, and a one-page range summary gives finance a commitment it can plan around.
\n1. A forecast is a range with stated assumptions, and leading with the spread builds more trust than hiding it. 2. Response curves show that the marginal dollar costs more than the average dollar, so judge budget changes by marginal cost. 3. Pacing ratio, required daily rate and run rate projection let you see a miss while there is still time to fix it. 4. Underspend and overspend deserve equal attention, and checking approvals, bids, audience size and overlaps finds most causes quickly. 5. Keep original forecasts, reforecast on a fixed rhythm, and score yourself with error and bias.
\nThis week, build one sheet that shows planned versus actual cumulative spend, the required daily rate and a seven day run rate projection for every live flight, then bring finance a three-scenario range instead of a single number.
\n1. Hyndman, R.J. and Athanasopoulos, G. Forecasting: Principles and Practice. Open online textbook covering seasonal methods, exponential smoothing and forecast accuracy measures. 2. Seabold, S. and Perktold, J. statsmodels: Econometric and statistical modeling with Python. Project documentation for time series and exponential smoothing models. 3. DuckDB project documentation. Reference for SQL analysis of CSV and Parquet files. 4. LibreOffice Calc Guide. The Document Foundation user documentation for formulas, charts and scenarios. 5. Advertising platform help documentation for budget delivery rules. Consult each platform's current rules on daily and monthly budget behavior before relying on any pacing assumption.
"}]},{"title":"Part VIII: Creative, Testing, and Operations","items":[{"id":"chapter-49-creative-strategy-for-performance-and-brand","title":"Chapter 49: Creative Strategy for Performance and Brand","html":"By the end of this chapter, the reader will be able to: 1. Distinguish a creative concept from its executions and explain why the difference controls testing results. 2. Rank creative variables by expected impact and sequence tests from the biggest swing to the smallest. 3. Differentiate brand-building creative from response creative and assign each to the right job and measurement. 4. Construct a message architecture that keeps one core promise consistent across formats and audiences. 5. Evaluate an ad against accessibility requirements before it ships. 6. Design a creative learning system with a test log, a decision rule, and a refresh cadence.
\nHere is a question. Last quarter, which moved your results more: the audience you chose, the bid strategy you set, or the ad itself?
\nMost brand managers answer audience or bid. They are wrong, and the reason matters. Targeting options keep shrinking as platforms automate more of the work, and every competitor sees the same auctions and the same tools. Bidding is increasingly handled by the platform's own algorithm. What you control, and what your competitors cannot copy by flipping a setting, is what the person on the other end of the impression actually sees and hears.
\nThat is my argument for this chapter: creative is the largest controllable variable in digital advertising, and most teams under-invest in testing it. They spend weeks debating audience splits and then upload three near-identical ads, shot on the same afternoon, with the same promise worded three slightly different ways. Then they declare a winner after four days.
\nI will not tell you that creative is magic. It is not. A brilliant ad pointed at the wrong offer fails, and a dull ad with a strong offer sometimes wins. Creative also fatigues, which means yesterday's winner is a depreciating asset. What I will show you is how to treat creative as something you can structure, test, and learn from, the same way you would treat a pricing decision or a supplier contract. The goal is a system that produces better ads every month, not one lucky hit.
\nA concept is the core idea: the single promise or tension the ad is built around. An execution is how that idea is dressed: the headline wording, the photo, the length, the voiceover, the button text. Teams confuse the two constantly.
\nSuppose a regional gutter-cleaning company wants more booked estimates. One concept says \"You will not climb that ladder again.\" Another says \"Clogged gutters ruin foundations, and repairs cost far more than a cleaning.\" A third says \"Booked by Thursday, done by Saturday.\" Those are three different ideas about what the customer cares about: avoiding risk, avoiding cost, and speed. Now suppose the team shoots one video per concept, then makes five cuts of each. They have 15 ads but only three concepts.
\nThis is where most testing goes sideways. If you test five cuts of the ladder idea against five cuts of the same ladder idea, you learn which cut is best. Maybe the gain is small. If you test the ladder concept against the foundation concept, you may discover that one of them is worth double the other. Concept differences are large. Execution differences are usually small by comparison.
\nA practical rule: before you commission anything, write each concept as one sentence that a stranger could repeat back. If two ads produce the same sentence, they are one concept. If you cannot write the sentence, the ad has no concept, only a look.
\nThe under-investment is rarely a decision. It is a pile of small habits.
\nProduction is slow and visible, while media changes are instant and invisible. A budget shift takes two minutes. A new video takes two weeks and an invoice. So teams tune what is cheap to change and neglect what matters more.
\nCreative also feels subjective, and arguments about taste end with the most senior opinion winning. A test ends the argument, but only if the team agreed in advance to accept the result. Dashboards do not help: they are organized by campaign and placement, and creative sits one level down, if it is reported at all.
\nThen there is the fear of waste. An ad that loses a test costs a small slice of budget and buys you a fact. An ad that runs untested for three months costs far more, because you never learn what you were missing. Think of a baseball team that never scouts. It has the same stadium and payroll as everyone else, and it hopes.
\nNot every variable deserves the same attention. I rank them by how large a swing they typically produce, and I test in that order.
\n1. Offer and concept. What are you promising, and why should anyone care? This is the largest swing. Test it first. 2. Audience-message fit. The same concept can land differently with new customers than with past buyers. Test whether the angle changes by group. 3. Format. Short video, static image, carousel, long-form video. Different formats serve different placements and different attention spans. 4. Hook. The first moments of a video or the first line of copy. A strong hook earns the rest of the viewing. 5. Visual treatment. Colors of the scene, framing, who appears, whether a person speaks to camera. 6. Copy details and call to action. Headline wording, button text, supporting line.
\nButton text is where many teams spend most of their testing energy, because it is easy to change. It is also where gains are smallest.
\nSequence matters for another reason: each level assumes the one above is settled. Once a concept wins, test formats within that concept. Once a format wins, test hooks within that format. You narrow the search the way a coach narrows a roster: first the position, then the player, then the playbook details.
\nChange one thing per test when you can. If the hook and the offer both change, a lift cannot be assigned to either one. The exception is the first round, when you are testing wholly different concepts. There, the ads differ in many ways by design, and the test answers a broad question: which idea has pull?
\nResponse creative asks for an action now: click, call, buy, book. Brand-building creative tries to be remembered and preferred later, when the person is actually ready to buy. They have different jobs, so judging them by the same metric guarantees one of them looks bad.
\nResponse ads tend to win on cost per action in the short run, because they speak to people already close to deciding. Brand ads look weak on click-based reports because their audience is not ready to click. Read only last-click numbers and you will cut them, then wonder why response ads cost more. You removed what was filling the pool of ready buyers.
\nWhat makes brand creative work is usually distinctiveness and memory, not persuasion. Does the viewer know whose ad it is without seeing the logo? Is there a consistent sound, phrase, character, or visual habit that you repeat? A sports comparison helps: fans know a team's chant or entrance song before they see the uniform. That recognition is built through repetition across many exposures, which is why changing the brand look every quarter is expensive.
\nWhat makes response creative work is clarity and a reason to act now. State the offer, the proof, and the next step in plain language. Remove anything that makes the viewer work to understand it.
\nThe mistake is treating this as a rigid split. Good response ads still carry the brand's recognizable elements, and good brand ads still give the viewer an easy next step. I treat it as a dial. Early in the buying process, turn toward memory. Close to purchase, turn toward action. The right setting depends on how long your buyers take to decide. A pizza shop needs the dial near action. A kitchen remodeler needs it near memory for months before action matters.
\nResponse ads can be judged by cost per qualified action. Brand ads need different evidence: recall surveys where you can run them, changes in branded search volume, and holdout comparisons between matched regions.
\nMessage architecture is a plain idea: decide, in writing, what your brand says first, second, and third, and keep it consistent everywhere.
\nI use three layers. The first is the core promise: one sentence about what the customer gets that they cannot easily get elsewhere. The second is a small set of proof points, usually three: the facts, guarantees, or experiences that make the promise believable. The third is the set of audience angles: how the promise is framed for different customers.
\nTake a fictional trailer-rental company, Prairie Haul. Its core promise: \"A trailer that is ready when you arrive, or the rental is free.\" Proof points: a posted inspection checklist, a 24-hour phone line answered by a person, and a no-surprise price at booking. Audience angles: for contractors, uptime on a work day; for families, a stress-free move; for hunters, trailers cleaned and ready before dawn.
\nEvery ad, on every platform, pulls from that structure. A short vertical video for families might open on the angle and end on the promise. A search ad for contractors might use the proof point about phone support. The wording changes. The promise does not.
\nWithout this, different teams write whatever sounds good, and customers meet five slightly different companies. With it, testing gets cleaner, because you know which layer you are testing.
\nAccessibility is not a final checklist item. It is a design constraint that improves ads for everyone and widens the audience that can use them. Many people watch video with the sound off, and many have hearing loss or low vision. An ad they cannot read or hear does not reach them, and you paid for the impression anyway.
\nBuild these into the brief, not the revision round:
\nFor ads that may reach children or families, set a conservative default: no pressure language aimed at young viewers, plain claims, and review by someone who did not make the ad. Where an ad uses a creator or endorsement, disclosure requirements apply. Treat this as operational guidance only and let counsel decide on specific wording.
\nI run accessibility checks before the test, so every variation meets the same floor and the test measures ideas, not compliance gaps.
\nHere is an illustrative case, with round numbers and a fictional brand. Fennel Crate sells subscription meal kits and spends 24,000 dollars a month on social video. For months it ran six ads, all variations on \"Dinner in 20 minutes.\" Results were flat.
\nThe team changed its approach. In round one, it tested three concepts against each other, one video each, with equal budgets: speed (\"Dinner in 20 minutes\"), waste (\"Stop throwing out groceries you never cooked\"), and confidence (\"Cook something you are proud of, even if you never cook\"). The audience, budget, and time window were the same for all three.
\nSuppose the waste concept produced meaningfully more trial signups per thousand impressions than the other two, and the speed concept, the long-time default, came last. Nobody on the team predicted that. The old default was a habit, not a finding.
\nRound two held the waste concept fixed and tested format: a 15-second cut, a 30-second cut, and a static image. Round three tested hooks within the winning format. Each round asked one question, held everything else steady, and started from the previous answer. The figures here are made up for teaching. The sequence is the lesson: biggest variable first, narrow as you go, record what you learn.
\nA test that is not recorded teaches nothing the next team member can use. A learning system turns single tests into accumulated knowledge.
\nThe core piece is a test log. For each test, record the question, the hypothesis, the variations, the audience, the budget, the dates, the result, and the decision. Add one more field: what you would do differently. A spreadsheet works. So does a shared document. What matters is that anyone can search it.
\nPair the log with a decision rule written before the test begins. For example: \"We will run each variation until it has reached at least a set number of conversions or two full weeks, whichever is later, and we will declare a winner only if the gap holds across both weeks.\" The exact number depends on your volume and counsel from your analyst. The point is to fix the rule in advance, because people stop tests early when they see numbers they like.
\nPlan for fatigue: repeated exposure weakens response and raises cost, so schedule refreshes rather than wait for decline. A common approach is to keep the winning concept but swap hooks and visuals on a regular rhythm, while running new concepts at a small share of budget so that the next winner is already in progress.
\nBudget for the system. I suggest treating a fixed share of media spend, often around 10 to 20 percent for accounts with room to experiment, as the learning budget. That money is allowed to lose. It is protected from the pressure to hit short-term targets, because its product is information.
\nConcept: The central idea or promise an ad is built on, independent of its wording, format, or visuals.
\nHook: The opening seconds of a video or the first line of copy, designed to earn attention before the viewer scrolls past.
\nCreative fatigue: The decline in response when the same audience sees the same ad too often.
\nControl: The current best-performing ad, which new variations must beat to replace it.
\nHoldout: A group of people or a region deliberately not shown the ads, used as a baseline to estimate what the ads added.
\n1. Testing five versions of one concept and calling it a creative test, when only the concept can produce a large difference. 2. Stopping a test after a few days because one ad looks ahead, instead of following a decision rule written before launch. 3. Judging brand-building ads by last-click results, then cutting them and watching response ads become more expensive. 4. Letting accessibility arrive as a last-minute edit, so the variations in a test differ in readability and the result is muddied. 5. Failing to record results, so the same losing idea gets proposed again a year later by someone who never saw the log.
\n1. Write every concept as one sentence a stranger could repeat, and refuse to brief an ad that cannot be summarized that way. 2. Test in order of expected swing: offer and concept, then audience fit, format, hook, visual treatment, and copy details. 3. Keep a written message architecture with one core promise, about three proof points, and a short set of audience angles. 4. Set aside a protected learning budget that is allowed to lose, and keep the share steady through good and bad months. 5. Check captions, contrast, and motion before any test launches so every variation meets the same accessibility floor.
\n1. Mine reviews, support calls, and sales notes for the customer's own words about the problem, and use those phrases in concepts and hooks. 2. Name creative files with concept, format, and version in a fixed pattern so reports can be grouped and compared later. 3. Run new concepts at a small share of budget alongside the current control, so a successor is always in progress before fatigue sets in. 4. Review the test log monthly with the whole team, including the losers, and ask which assumptions the results broke. 5. Rerun major winners in a new season or audience before building a whole quarter around them.
\nPenpot (MPL 2.0) is a design and prototyping tool where teams build and share ad layouts in a browser, and it works well for message architecture reviews and static ad templates. Krita (GPL) handles illustration and painted or photo-based image work. Inkscape (GPL) produces clean vector logos, icons, and scalable graphics that resize across formats without losing quality. GrowthBook (MIT core) runs experiments on pages and apps you control, with traffic splitting and statistical reporting for landing page and creative tests. All four are free to use and can be self-hosted. None replaces the discipline of a written test log.
\nCreative is the largest controllable variable in digital advertising, yet most teams spend far more effort on audience and bidding than on testing the ad itself. Concepts, not executions, produce the biggest differences, so tests should begin with offer and concept and narrow through audience fit, format, hook, visual treatment, and copy. Response creative and brand-building creative do different jobs and must be measured differently, or the brand work will be cut for failing a test it was never meant to pass. A written message architecture keeps the promise consistent, accessibility checks protect reach and test validity, and a learning system with a log, a decision rule, a refresh schedule, and a protected budget turns individual tests into knowledge that outlasts any one campaign.
\n1. Creative controls more of your results than targeting or bidding, and it is the variable competitors cannot copy by changing a setting. 2. Concepts differ far more than executions, so test big ideas first and polish details last. 3. Brand-building and response creative need different jobs and different measures, and last-click reports alone will mislead you about brand work. 4. A written message architecture and an accessibility floor make tests cleaner and customers' experience consistent. 5. A test log, a pre-set decision rule, scheduled refreshes, and a protected learning budget make creative improvement repeatable.
\nA launch is ready when each ad's concept is written as one sentence, a protected share of budget is set aside for testing concepts against each other, and every result is logged where the whole team can read it.
\n1. World Wide Web Consortium. Web Content Accessibility Guidelines, version 2.1. Success criteria for contrast, text alternatives, and flashing content. 2. Federal Trade Commission. Guides Concerning the Use of Endorsements and Testimonials in Advertising, 16 CFR Part 255. Disclosure guidance for creator and endorsement content. 3. Sharp, Byron. How Brands Grow. Oxford University Press. Discussion of distinctiveness and mental availability in brand building. 4. Kohavi, Ron, Diane Tang, and Ya Xu. Trustworthy Online Controlled Experiments. Cambridge University Press. Practical guidance on experiment design and pitfalls. 5. Ogilvy, David. Ogilvy on Advertising. Vintage Books. Practitioner perspective on concept, headline, and offer.
"},{"id":"chapter-50-landing-pages-and-conversion-rate-optimization","title":"Chapter 50: Landing Pages and Conversion Rate Optimization","html":"By the end of this chapter, the reader will be able to: 1. Explain why a landing page must keep the specific promise made by the ad that sent the visitor. 2. Diagnose page speed problems and rank the fixes by their likely effect on conversion. 3. Design a lead form that collects only what the business will actually use in the first follow-up. 4. Apply the concept-level requirements of WCAG to a landing page without a specialist audit. 5. Sequence a testing program from obvious defects to small refinements and calculate whether traffic can support each test. 6. Evaluate a reported test win and judge whether it is real, noisy, or an artifact of how the test ran.
\nShould you be optimizing landing pages at all?
\nHere is the case. Your ad did the hard work. It found a person at a useful moment and earned a click, and the page is only the room where that person lands. Most pages I see in client accounts get a few hundred visits a week. At that volume, a test that moves the button text produces a result that looks like information and is mostly noise. Polishing the page while the offer is weak is like repainting the locker room while the team keeps losing at home.
\nI think that argument is about half right, and the half that is right matters. If your traffic is thin, you should not run clever experiments. If your offer is unclear, no layout will rescue it. If the ad and the page disagree, fix that first and call it a day.
\nThe half that is wrong is the idea that the page is a passive container. It is the only part of the funnel you fully own. You do not control the auction, the feed, the algorithm or the competitor who just raised bids. You do control what loads, how fast it loads, what it says and what it asks for. Every click you pay for arrives at that page, and every defect on it is charged to your media budget a second time.
\nSo this chapter treats the page as a promise to keep, a performance budget to meet, a form to justify, and an experiment to run honestly.
\nA visitor arrives with a short memory. They saw a line, an image and an offer a few seconds ago, and they expect to see that same thing again. Marketers call this message match. I think of it as the handshake. The ad extends a hand, and the page either takes it or leaves the visitor hanging.
\nSuppose a regional outdoor retailer, Marrow Creek Supply (an illustrative composite), runs an ad for a winter boot with free exchanges. The click lands on a footwear category 60 products deep, with the exchange policy buried in the footer. The visitor asked about one boot and one policy, not a catalog. Most will leave, and the click becomes a cost without a return.
\nThe fix is rarely dramatic. Send each ad group to a page whose headline repeats the ad's main phrase and states the offer in the first screen, with a link to the terms.
\nMatching has a limit. The page must not promise more than the ad, and it must not hide a condition the ad left out. A \"free\" trial that asks for a card on the second screen, or a \"no obligation\" quote that requires a phone call from three vendors, breaks the promise and creates compliance exposure. Operational guidance only here; your counsel decides what your disclosures must say. The conservative default is to put every material condition on the same screen as the claim it qualifies.
\nSlow pages are not a technical footnote. They are a choice somebody made when they added a video background, four tracking scripts and an uncompressed hero image.
\nVisitors on a phone, on a mediocre connection, will not wait politely. The longer the first screen takes to become usable, the more people leave before the page has said anything, and paid visitors have no loyalty to you yet.
\nThree measurements matter most, and a free tool can give you all of them. Largest Contentful Paint is the time until the main visible element appears. Interaction responsiveness is how quickly the page reacts when someone taps. Layout shift is how much the page jumps around while loading, the thing that makes you tap the wrong button.
\nThe usual culprits: 1. Hero images that are several times larger than the screen that displays them. 2. Third-party scripts for chat, heat tracking, retargeting and testing that all load before the content. 3. Web fonts that block text from appearing.
\nSet a performance budget before the design starts: a maximum total page weight, a maximum number of third-party scripts, and a target load time on a mid-range phone with a throttled connection. Then test against it on every revision. A budget turns speed from an opinion into a pass or fail check, and it settles arguments quickly when a stakeholder wants one more widget.
\nMost weak pages fail at something simple. The visitor cannot tell, within about five seconds, what is offered, who it is for, what it costs or what to do next.
\nRun the squint test. Show the page to a stranger for five seconds, take it away, and ask three questions. What is this? Who is it for? What would you do next? If they stumble on any of the three, the problem is clarity, and no button change will help.
\nA clear first screen has a headline that names the outcome, one supporting line, one primary action, and one image or proof element. Everything else belongs lower.
\nTwo choices deserve a decision rather than a habit. Navigation: a paid page usually does better with the site menu reduced to a logo and a phone number, because every extra link is an exit you paid for. Length: short pages suit simple, low-cost actions, and longer pages suit expensive or risky decisions where objections need answers. Length that answers real questions is good. Length that repeats a claim is padding.
\nA form is where intent turns into data, and where most lead pages lose people who had already decided.
\nEvery field must pay rent. If a field does not change what the business does in the first contact, it does not belong on the first form. Ask for company size only if a different person responds to a big company than to a small one. Ask for a phone number only if someone will call, and say so beside the field.
\nShorter forms usually produce more submissions, but can produce lower quality leads, because friction screens out the mildly curious. Neither version is automatically better. The right measure is cost per qualified outcome, such as a booked appointment, not cost per form fill.
\nPractical details that matter more than their size suggests: 1. Label every field above the input, not as faint placeholder text that disappears when someone types. 2. Use the correct input type so a phone keyboard shows numbers for a phone field. 3. Allow autofill, and write error messages that say how to fix the problem. 5. Make the submit button say what happens next, such as \"Get my quote\".
\nFamily-first standard: collect the least personal data you can, never ask about children, health or finances unless the offer requires it, and link the privacy notice beside the button.
\nTrust signals are evidence, and evidence beats adjectives.
\nUseful trust elements include genuine reviews with the date and the reviewer's first name and city, clear pricing or a plain price range, a visible return or cancellation policy, a physical business address and phone number, recognizable payment options, and named people. Certifications and affiliations belong on the page only when you actually hold them and can document them.
\nBe careful with the rest. Fabricated testimonials, invented viewer counters, fake scarcity timers and borrowed logos can be deceptive practices, and endorsement rules require testimonials to reflect honest opinions and disclose material connections. Counsel decides what applies to you. The operational rule: if you could not defend a claim to a skeptical customer in person, remove it.
\nPlace proof next to the claim it supports, not in one strip at the bottom.
\nA page that works for people using a screen reader, a keyboard, a magnifier or a phone in bright sunlight works better for everyone, and it often reduces legal exposure.
\nThe Web Content Accessibility Guidelines, usually abbreviated WCAG, organize the work around four ideas: content must be perceivable, operable, understandable, and compatible with assistive technology. You do not need to memorize success criteria to apply these at the concept level. Ask a handful of practical questions.
\nCan someone use the whole page and form with only a keyboard, and can they always see where they are? Does every image that carries meaning have a text description, and does decorative imagery stay quiet? Is text large enough and strong enough against its background to read outdoors? Does every form field have a real label tied to it in the code? Do error messages explain themselves in text, rather than relying on a visual cue alone? Does the page still work when someone zooms to twice the size? Do videos have captions?
\nAutomated tools catch only the mechanical problems. Run the check, tab through the page yourself, and bring in a qualified auditor if your page is regulated or public. This is operating guidance, not a conformance opinion.
\nHere is where I return to my opening argument. Testing is valuable, but only when the order is right and the traffic can carry it.
\nStart with a sequence rather than a wish list: 1. Fix defects. Broken forms, slow loads, mismatched messages and unreadable text do not need tests. They need repairs. 2. Apply known good practice without a test, such as readable text and a working mobile layout. 3. Test big, meaningful differences first: a different offer, a different headline angle, a shorter page against a longer one. 4. Test mid-size structural choices, such as form length, proof placement and the main image. 5. Only then test small items like button wording, and only if traffic is high.
\nCheck whether a test can finish before starting it. Take a worked example. Suppose a page gets two thousand visits a month and converts at three percent, which is 60 conversions. You want to detect an improvement from three to four percent. Detecting a gap that small, with confidence, typically needs on the order of thousands of visitors per version. At one thousand per version per month, you would wait many months, and the market will have moved by then. The honest options: test a bigger change, combine traffic, measure an earlier step such as a form start, or ship the best-judged version.
\nWrite the hypothesis before the test begins, in one sentence: \"Because visitors ask about exchanges in reviews, putting the exchange policy in the first screen will raise form starts.\" Also write the single primary metric, the minimum meaningful lift, the planned duration, and what you will do under each outcome. Writing it down in advance stops you from picking whichever metric happened to improve.
\nRun tests across whole weeks, randomize visitors rather than days, and show each person the same version on every return.
\nMost disappointments in optimization are not bad luck. They are predictable errors.
\nPeeking is the first. Stop the moment one version looks ahead and you will declare winners far more often than chance justifies. Fix the duration in advance, or use a method built for continuous monitoring.
\nRunning many tests or reading many metrics at once is the second. Check enough segments and something will light up. A lift that appears only for one device in one city on one weekday is usually a coincidence. Treat segment findings as new hypotheses, and test them separately.
\nThe third is a novelty effect. A new version can get extra attention just for being different, and the lift fades. Another is the difference between click rate and outcome. A headline that gets more clicks on the button may bring people who never buy. Always follow the result through to the business outcome you care about.
\nA test that ran during a promotion measured the promotion, not the page. Once in a while, run an A/A check: split traffic between two identical pages and confirm the tool reports no difference. If it finds one regularly, your setup is broken.
\nUse this rubric before launch and at every quarterly review. Score each line pass, partial or fail, with a reason for anything short of a pass.
\n| Area | Question | Pass looks like | |---|---|---| | Message match | Does the headline repeat the ad's main promise? | A visitor can see the same offer within one second | | Speed | Does the first screen load quickly on a mid-range phone? | Meets the written performance budget | | Clarity | Can a stranger state the offer, audience and next step after five seconds? | All three answered correctly | | Primary action | Is there one main action, visible without scrolling? | One clear button, one purpose | | Form | Does every field change the first follow-up? | No unused fields, clear labels, specific button text | | Trust | Is proof real, dated and placed beside the claim? | Verifiable and near the doubt | | Disclosure | Are all material conditions on the same screen as the claim? | Nothing hidden below or behind a link | | Accessibility | Can the page and form be completed by keyboard, with labels and captions? | Passes the practical questions above | | Mobile | Are buttons tappable and text readable without zooming? | Tested on a real device | | Measurement | Are the primary metric and tracking verified? | A test submission appears in the data |
\nMessage match: The degree to which a landing page repeats the specific promise, wording and offer of the ad that sent the visitor there. Largest Contentful Paint: A speed measurement of how long the biggest visible element on the first screen takes to appear. Minimum detectable effect: The smallest improvement a test is designed to reliably notice, set before the test starts and used to calculate the required traffic. A/A test: A check in which traffic is split between two identical versions to confirm the testing setup does not invent differences. Novelty effect: A short-lived lift in response caused by a version being new, which fades as visitors get used to it.
\n1. Sending every ad to the homepage or a category page, so visitors must search for the thing the ad promised. 2. Running tests on pages with too little traffic to finish, then acting on noise as if it were a finding. 3. Judging a test by form fills or clicks alone, without checking whether those leads became customers. 4. Stopping a test the first day one version pulls ahead, which inflates false winners. 5. Treating accessibility and page speed as a launch checklist item instead of a standing requirement checked on every revision.
\n1. Write the hypothesis, the single primary metric, the minimum meaningful lift and the duration before any test starts. 2. Set a performance budget for page weight, third-party scripts and load time, and test every revision against it. 3. Build a few locked templates with swappable headlines and proof points instead of many one-off pages. 4. Keep one primary action per page and justify every form field by its use in the first follow-up. 5. Keep a test log recording the result, the uncertainty and the decision, and review it before planning new tests.
\n1. Take the phone number off the form unless someone will genuinely call within a day, and say that beside the field. 2. Check the page on the cheapest phone you can borrow, on a throttled connection, before trusting any desktop speed score. 3. Place the refund or cancellation promise beside the price and the button, where doubt is highest. 4. Defer or remove third-party scripts that do not serve the first screen, since each one competes with your content for load time. 5. Review the form submission path end to end every month by submitting a real test and confirming it arrives where the sales team will see it.
\nLighthouse (Apache 2.0) audits a page for speed, accessibility and basic quality, and gives repeatable scores you can hold against your performance budget. sitespeed.io (MIT) runs those checks on a schedule across many pages and keeps the history, so regressions show up before visitors notice them. GrowthBook (MIT core) handles experiment assignment and statistical analysis for your tests. Unleash (AGPL) is a feature flag service that can show different page versions to different visitors and switch a bad change off instantly. Penpot (MPL 2.0) is a design and prototyping tool for drafting page layouts and variants before anyone writes code. All five can be self-hosted without paid subscriptions.
\nA landing page earns its keep by keeping the ad's promise, loading quickly on ordinary phones, stating one clear offer, asking only for data the business will use, and backing its claims with evidence a skeptical customer would accept. Testing works only when the sequence runs from repairs to big differences to small refinements, and only when traffic can finish the test. False wins come from peeking, too many metrics, novelty and outside events, so a written plan and an occasional A/A check protect the budget. The review rubric turns these ideas into a repeatable check.
\n1. The page is the part of paid media you fully control, so every defect on it is paid for twice. 2. Message match means the headline repeats the ad's specific promise, with every material condition on the same screen as the claim. 3. A performance budget turns page speed into a pass or fail requirement that stakeholders cannot argue away. 4. Every form field must change the first follow-up, and form length should be judged by cost per qualified outcome, not cost per submission. 5. Calculate whether traffic can finish a test before you start it, and write the hypothesis, metric and duration down in advance.
\nBefore you fund another test, run the five-second stranger test and the phone speed check on your highest-spend landing page, and fix whatever fails. Then test only changes big enough for your traffic to detect.
\n1. World Wide Web Consortium. Web Content Accessibility Guidelines (WCAG) 2.1. Recommendation defining perceivable, operable, understandable and compatible requirements for web content. 2. Federal Trade Commission. Dot Com Disclosures: Information about Online Advertising. Guidance on clear and conspicuous placement of online disclosures. 3. Federal Trade Commission. Guides Concerning the Use of Endorsements and Testimonials in Advertising, 16 CFR Part 255. 4. Kohavi, R., Tang, D., and Xu, Y. Trustworthy Online Controlled Experiments: A Practical Guide to A/B Testing. Cambridge University Press. 5. Lighthouse project documentation. Performance and accessibility auditing, Apache 2.0 licensed software.
"},{"id":"chapter-51-experimentation-programs-and-statistical-discipline","title":"Chapter 51: Experimentation Programs and Statistical Discipline","html":"By the end of this chapter, the reader will be able to: 1. Define a testable hypothesis with a stated metric, minimum worthwhile effect, and decision rule. 2. Calculate the approximate sample size and duration an experiment needs before launch. 3. Explain why peeking and multiple comparisons inflate false positives, and apply a fix for each. 4. Select guardrail metrics that protect the business while a test chases a primary metric. 5. Evaluate whether a proposed test is worth running or should be replaced by a direct decision. 6. Design an experiment log and review meeting that turns null results into usable knowledge.
\nPicture a regional furniture brand, fictional and illustrative, called Harbor Oak. The brand manager runs a landing page test for a spring promotion. Version B carries a bigger offer banner. On day three, B leads by 18 percent on lead form completions. The dashboard shows a friendly \"significant\" label. The owner is in the room, the budget is moving, and everyone agrees the answer is obvious. B ships to all traffic that afternoon.
\nThree weeks later, lead volume is flat, and the sales team reports thinner leads: fewer answer the phone, fewer book showroom visits. The decision was sensible. It was also wrong, because the test was stopped the moment it looked good, nobody defined \"good\" beforehand, and the only metric on screen was the one B happened to win.
\nMost brand managers believe that experimentation is a tool problem. Buy the platform, split the traffic, read the winner. That belief is where programs go to die. A testing tool will happily compute a result from any data you feed it, including data that cannot support a conclusion. The tool does not know you peeked. It does not know you ran nine other variants last month. It does not know sales quality fell.
\nThis chapter treats experimentation as a discipline with a handful of rules, most of them decided before launch. The statistics stay at a plain level, and free open source libraries do the arithmetic.
\nA hypothesis is a prediction with a price tag and an exit. \"Test a new headline\" is not a hypothesis. It is an errand. A usable hypothesis names the change, the audience, the primary metric, the smallest improvement worth acting on, and the decision you will make under each outcome.
\nHere is a template I use with teams. \"If we change X for audience Y, then metric Z will improve by at least W, because of reason R. If it does, we will do A. If it does not, we will do B.\" The because clause matters more than it looks. It forces you to say why you believe the change works, and that reason is what you learn about when the test ends, win or lose.
\nConsider an illustrative case. A regional credit union, fictional, runs paid search for auto loan applications. The team suspects that mentioning \"same day decision\" in ad copy attracts shoppers who are further along. The hypothesis reads: \"For auto loan search ads in the metro area, adding same day decision language will raise completed applications per click by at least 15 percent, because it reduces fear of a slow process. If it does, we roll it into every auto loan ad group. If it does not, we keep current copy and test the application form instead.\"
\nNotice the smallest worthwhile effect. Fifteen percent is a business judgment, not a statistical one. A target of two percent sounds humble, yet it demands enormous traffic and rarely changes a decision.
\nPick one primary metric per test. Secondary metrics can be watched, but they are for context and for generating the next hypothesis. If you promote a secondary metric to winner status after the primary one disappoints, you have started a new test with the data from the old one, and the math no longer applies.
\nStatistical power is the chance that your test detects a real effect of the size you care about. Power of 80 percent means that if the true lift equals your target, you will catch it four times in five. Underpowered tests are the quiet scandal of marketing: they miss real winners, and when they do show a winner, the estimate is often inflated. A small test that crosses the threshold usually did so because noise happened to point upward.
\nThree things set the sample you need. The baseline rate of your metric, the smallest lift you want to detect, and how much false positive risk you accept (usually five percent).
\nHere is a worked example with round illustrative numbers. Suppose a page converts at four percent and you want to detect a lift to five percent. With 80 percent power and a five percent false positive rate, you need roughly seven thousand visitors in each arm. If the page gets two thousand visitors a day split evenly, that is about seven days. Now suppose you care about a lift to only four and a half percent. The needed sample jumps to roughly 26,000 per arm, or about 26 days. Halving the effect you want to see multiplies the sample by about four.
\nstatsmodels computes this in a few lines for two proportions, and scipy can check your work. Run it before the test and write the required sample in the hypothesis document.
\nRun tests in whole weeks. Weekday and weekend visitors behave differently, and a test covering only Tuesday through Friday describes only those days. Avoid known events such as a holiday sale, or cover them fully in both arms.
\nIf the arithmetic says you need more traffic than you have, that is an answer. Test a bigger change, use an earlier funnel metric with a higher rate, pool markets, or skip the test.
\nPeeking means checking results during a test and stopping when they look good. It feels like diligence. It is a trap. The significance threshold assumes you look once, at a planned sample size. Every extra look is another chance for random noise to cross the line. Picture a free throw contest where you may stop whenever you are ahead: even a shooter with no edge will crown a winner. Teams that check daily and stop at the first significant reading can see their real false positive rate climb to several times the stated five percent.
\nThe Harbor Oak story at the start is exactly this. Day three results on a small sample swing wildly. The early lead was a coin flip that happened to land well.
\nThere are three practical defenses, from simplest to most advanced.
\nFirst, fix the sample in advance and look only at the end. This is the plain answer and it works. Put the stop date in the hypothesis and treat the dashboard as off limits until then. Monitoring for broken tracking or a technical failure is fine, as long as nobody reads the lift number.
\nSecond, use a sequential method when you truly need the option to stop early. Sequential testing, at a plain level, means the test is designed from the start with a stricter bar early on that relaxes as data accumulates. The total false positive risk stays at five percent even though you look many times. The price is that you need somewhat more traffic to reach the same power than a fixed test that you look at once. If a decision cannot wait, say a promotion that must lock by a set date, this tradeoff is worth paying.
\nThird, set a stop-on-harm rule. If a guardrail metric falls past a stated threshold, the test halts. That is a safety brake, not a license to declare winners early.
\nEvery additional comparison is another lottery ticket. If you test one variant against control at a five percent false positive rate, you have one ticket. Test 20 variants, or one variant across 20 audience segments, and the chance that at least one shows a spurious win is roughly two in three. Slice by device, age band, city, and hour, and something will light up. It will not replicate.
\nThe more ways you can cut the data after the fact, the more likely you find a story. Here are the rules I give to teams.
\nSuppose a fictional restaurant group tests seven ad creatives across four markets. That is 28 combinations, and the single best cell is almost certainly a lucky one. Pool the markets for the primary read and label market differences as exploration.
\nA primary metric tells you whether the change worked. A guardrail metric tells you whether it worked at an unacceptable cost. Harbor Oak had no guardrail on lead quality. A bigger banner pulled in curious clickers, and the form completions looked healthy while the sales pipeline thinned out.
\nAsk what an overeager version of this change could damage.
\nEach guardrail needs a threshold written before launch. \"Unsubscribe rate may not rise by more than a stated amount\" is a rule. \"Keep an eye on unsubscribes\" is a hope. A guardrail checked with too little data cannot detect harm, so a flat reading often means \"unknown\" rather than \"safe.\"
\nA guardrail breach has a clear rule. The primary win does not ship until you understand the breach. Sometimes the answer is a modified version of the change. Sometimes it is that the win was bought at too high a price.
\nA program is a habit, and habits live in records. The experiment log is a simple shared table, spreadsheet or database, with one row per test.
\nEach row holds the hypothesis in the template form, the primary metric and guardrails, the required sample and planned end date, who owns the test, launch and stop dates, the result with an estimate and an interval (not just a verdict), the decision that followed, and what the team learned. Mark tests that were stopped early, broken by tracking issues, or contaminated by a promotion.
\nLog failed and inconclusive tests with the same care as wins. A log of only winners tells the next team that winners are common and cheap. They are neither.
\nHold a short review on a regular schedule, every two weeks works for many teams. The agenda is simple. Which tests ended, what did each find, what decision follows, what comes next. Keep three habits at the meeting.
\nRead the interval, not only the verdict. A result of \"lift between negative two and plus nine percent\" says something very different from \"lift of three percent, significant.\" The first tells you the data could not settle the question. The second pretends it did.
\nAsk what would make each new result untrustworthy: tracking gaps, overlapping campaigns, a holiday, a bidding change mid-test. Better to find these before launch.
\nA null result means the test could not tell your variant from the control. It is not the same as \"the idea is bad,\" and it is not wasted. If the test was properly powered, a null result tells you the lift, if any, is smaller than the minimum you set. That is a real finding. It lets you stop spending effort on that idea and redirect the team.
\nCulture makes this possible. Where a failed test marks its author, people stop proposing risky ideas and run safe tests that teach nothing. Strong programs ask after a null result: \"What did we learn that we did not know before?\" and reward a clear answer.
\nBuild that habit with small practices. Celebrate the clean stop, where a test ended as planned and a decision followed on schedule. Share a one-paragraph summary of every test, nulls included. Track the share of tests that ended with a documented decision, not the share that won.
\nA good baseball hitter fails seven times in 10 and is paid handsomely. The strikeouts are the cost of swinging at enough pitches. A testing program works the same way.
\nSome questions should not become tests, and knowing which ones is part of the discipline.
\nSkip the test when the traffic cannot support it. If the power calculation says a year, decide on judgment.
\nSkip it when the downside is tiny and the change is easily reversed. A typo fix needs no split.
\nSkip it when the decision is already made. If leadership ships regardless, the test is theater.
\nSkip it when the change is required by law, policy, or brand safety. Disclosure language, consent wording, and age-appropriate targeting for family audiences are fixed rules, not optimization subjects. Counsel decides legal questions; this is operational guidance only.
\nSkip it when a more direct method exists. Five short usability sessions will tell you more about a confusing form than a conversion split.
\nThe test of the test: if I could see the answer now, would I act differently, and can I get the answer in a reasonable time with the traffic I have? If either answer is no, put the idea somewhere else.
\nStatistical power: The probability that a test detects a real effect of a stated size. Eighty percent is a common planning target.
\nMinimum detectable effect: The smallest lift a test is designed to reliably detect, usually set by what the business would act on.
\nFalse positive: A result that looks like a real effect but is produced by chance. The significance threshold caps how often this happens.
\nGuardrail metric: A secondary measure with a pre-set limit that protects against harm while the primary metric is being improved.
\nSequential testing: A test design that allows planned early looks at the data while keeping the overall false positive risk at the stated level.
\n1. Stopping a test the moment the result turns significant, which turns random early swings into false winners. 2. Launching without a power calculation, then treating an undersized test's null result as proof that an idea does not work. 3. Slicing results by every segment after the fact and reporting the one cell that lit up as a finding. 4. Optimizing a single primary metric with no guardrails, so a lift in clicks or form fills quietly damages lead quality or margin. 5. Logging only the winners, which erases the null results and lets the next team repeat the same failed ideas.
\n1. Write the hypothesis, primary metric, minimum worthwhile effect, sample size, and stop date before launch, and save the document where the team can see it. 2. Run tests in whole weeks, with a planned end date, and keep the lift number hidden until that date arrives. 3. Declare your segments in advance and adjust the threshold when many comparisons are unavoidable. 4. Set a written threshold for every guardrail metric and a rule for what happens when one is breached. 5. Hold a regular review where every ended test gets a documented decision, including the nulls and the stopped tests.
\n1. Before spending effort on a design, ask the traffic question first: can this page or audience supply the sample the effect needs in a reasonable time? 2. Test bigger, more distinct ideas. A small tweak has a small effect, and small effects need huge samples to detect. 3. Keep a calendar of live tests by audience and surface so that overlapping experiments do not contaminate each other. 4. For metrics that arrive slowly, such as refunds or repeat purchase, schedule a follow-up read after the test closes and record it in the log. 5. When a result matters for a large budget decision, rerun the test or confirm it with a second method before scaling the spend.
\nThree free, open source tools cover most of what a working team needs. GrowthBook (MIT core) handles experiment assignment, feature flags, and result reporting against your own data. scipy (BSD) offers standard statistical tests and distributions for checking a result by hand. statsmodels (BSD) provides power and sample size calculations for proportions and means, along with multiple testing corrections such as Bonferroni and Holm. Together they let a small team plan, run, and audit tests without a paid subscription.
\nA testing program earns trust through decisions made before the data arrives. A good hypothesis names its metric, minimum worthwhile effect, and exit rules. A power calculation tells you whether the traffic can support the question, and a fixed sample or a sequential design protects you from peeking. Multiple comparisons are controlled by declaring the analysis in advance and adjusting thresholds when many looks are unavoidable. Guardrail metrics keep a win from costing more than it earns. A shared log and a regular review turn every result, including nulls, into institutional memory, and a clear sense of when not to test keeps the program focused on questions worth answering.
\n1. A hypothesis needs a named metric, a minimum worthwhile effect, and a decision for each outcome before launch. 2. Halving the lift you want to detect roughly quadruples the sample you need, so run the arithmetic first. 3. Peeking and unplanned segment slicing produce false winners, and the fixes are a fixed plan or a sequential design. 4. Guardrail metrics with written thresholds keep a primary win from hiding damage elsewhere. 5. A log that records nulls and stopped tests is worth more than a log of wins, and some questions should skip testing entirely.
\nNever launch a test without one page that states the hypothesis, the minimum lift worth acting on, the required sample, the stop date, and two guardrail limits, and do not look at the lift until that date.
\n1. Kohavi, R., Tang, D., and Xu, Y. Trustworthy Online Controlled Experiments: A Practical Guide to A/B Testing. Cambridge University Press. Practitioner reference on design, guardrails, and pitfalls. 2. Johari, R., Pekelis, L., and Walsh, D. Always Valid Inference: Bringing Sequential Analysis to A/B Testing. Published research on peeking and continuous monitoring. 3. Benjamini, Y., and Hochberg, Y. Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. Journal of the Royal Statistical Society, Series B. 4. Cohen, J. Statistical Power Analysis for the Behavioral Sciences. Lawrence Erlbaum Associates. Foundational text on power and effect size. 5. statsmodels documentation, power and multiple testing modules. Open source project documentation.
"},{"id":"chapter-52-agency-freelancer-and-in-house-operating-models","title":"Chapter 52: Agency, Freelancer, and In-House Operating Models","html":"By the end of this chapter, the reader will be able to: 1. Describe how agency, freelancer, and in-house models differ on control, speed, cost visibility, and skill depth. 2. Apply a weighted scoring matrix to compare operating models for a specific budget and workload. 3. Draft transparency requirements that let a finance partner reconcile media spend to invoices. 4. Construct a RACI chart that assigns clear ownership for each recurring advertising task. 5. Design a staged plan for moving work in-house without losing performance during the handoff. 6. Evaluate warning signs in agencies, freelancers, and internal teams before they become expensive.
\nHere is a scene that repeats in marketing departments everywhere. The agency's monthly report shows twelve hundred leads and a cost per lead the team is proud of. The finance partner opens the same month's invoice and the bank statements and asks a simpler question: how much did the media platforms actually charge us, how much did the agency keep, and where is the paperwork that proves it? Silence. The dashboard is a story. The invoice is a fact. Only one of them survives an audit.
\nMost brand managers treat the choice between agency, freelancer, and in-house team as a talent question. Who is smartest, who is cheapest, who returns calls. That framing misses the real issue. The operating model decides who holds the account logins, who sees the raw invoices, who can change a bid at nine at night, and who walks away with the knowledge when the relationship ends. Those are control questions, and they matter more than any pitch deck.
\nThere is also an assumption worth challenging on the other side. Many in-house advocates treat agencies as a margin tax. That is unfair and usually wrong. A good agency spreads the cost of senior specialists across many clients, sees patterns across categories that no single company can see, and absorbs the pain of platform changes. A weak agency does none of that. The model is not the problem. The terms and the transparency are.
\nI have run work through all three arrangements, in my own companies and for others, and I have been burned by each of them. The goal of this chapter is not to crown a winner. It is to give you a method for choosing, a way to write the contract or the job description so you can see what you are paying for, and a staged path for changing course without gutting performance in the process. Everything here is operational guidance. Your counsel and your finance team decide the legal and accounting details.
\nStrip away the marketing language and you can compare the three models on four questions.
\nControl: who holds the accounts, the data, and the authority to change spend? Speed: how many hours pass between a decision and a live change? Cost visibility: can you trace every dollar from your bank to a media platform and to a person's time? Skill depth: how much specialist knowledge sits behind the work, and how fast can it be replaced if someone leaves?
\nThink of it like a football roster. An agency is a rented veteran line: deep talent, but they play for a contract and rotate between teams. A freelancer is a specialist you sign for one position: sharp, quick to deploy, but one injury leaves a hole. An in-house team is your own roster: slower to build, fully loyal to your playbook, and entirely dependent on your coaching. No roster wins every game.
\nAgencies earn their fees when the work needs breadth. If your campaign calls for strategy, creative, paid search, paid social, analytics, and landing page work in the same month, a team of eight specialists at a retainer is hard to match. Agencies also bring pattern recognition. They have seen a tracking setup break after a site redesign dozens of times, and they fix it fast. The strain shows up in three places: attention (your account may be staffed by junior people while senior people sell), incentives (a fee tied to spend rewards more spend, not better results), and handoff (when you leave, the playbook often stays behind).
\nFreelancers are strongest when the need is narrow and clear. A paid social specialist who knows one platform deeply, a conversion copywriter, a video editor, an analytics engineer. You pay for focused skill without overhead. The strain is capacity and continuity. Freelancers get sick, take vacations, and win larger clients. If one person holds your entire paid search account in their head, you have bought a single point of failure.
\nIn-house teams are strongest where context matters more than breadth. They know why last spring's offer flopped, which product ships late, and what the sales team will actually follow up on. They move fast because there is no brief to write. The strain is skill depth and hiring. A team of two generalists cannot match an agency's bench, and a resignation can erase a year of learning. In-house also carries costs that never appear on the marketing budget line: recruiting, software, benefits, management time, and training.
\nA matrix turns an argument into a decision. Pick six criteria, weight them to reflect your situation, score each model from one to five, and multiply. The point is not mathematical precision. The point is forcing your team to say out loud what they value.
\nHere is an illustrative example for a regional retailer spending about 40,000 dollars a month on paid media with a two-person marketing team. The weights are suppose-we-decide numbers, not a benchmark.
\n| Criterion | Weight | Agency | Freelancer | In-house | |---|---|---|---|---| | Control of accounts and data | 25 | two | three | five | | Speed of change | 15 | three | four | five | | Cost visibility | 20 | three | four | four | | Skill depth | 20 | five | three | two | | Continuity and risk | 10 | four | two | three | | Management time required | 10 | four | two | two |
\nMultiply each score by its weight, then add. Agency: 50 plus 45 plus 60 plus 100 plus 40 plus 40 gives 335. Freelancer: 75 plus 60 plus 80 plus 60 plus 20 plus 20 gives 315. In-house: 125 plus 75 plus 80 plus 40 plus 30 plus 20 gives 370. In-house wins on paper, but notice how close the totals are. A gap that small tells you the answer is probably a hybrid, and that the weights matter more than the scores.
\nTransparency is not an insult to your partner. It is the entry fee for any relationship involving someone else's money. Agencies that run clean operations usually accept these terms without argument. Agencies that resist them have told you something.
\nStart with account ownership. The advertising accounts, analytics properties, tag containers, and pixels should be created under your company's legal name and your administrators. The agency or freelancer gets delegated access. If the relationship ends, you remove their access and keep your history. This single clause prevents more pain than any other.
\nNext, media cost reporting. Require monthly statements showing actual platform charges, separately from fees, with line-level detail you can match to platform invoices. If the partner buys through a reseller or an aggregated account, require disclosure of that structure and of any discount, rebate, or volume incentive earned on your spend and whether it is passed through. Many agencies pass these along and some keep them under disclosed terms. Either can be fair. Undisclosed is not.
\nThen principal-based buying. If the agency buys media in its own name and resells it to you, it is acting as a principal rather than your agent. The inventory source, the markup, and the reason should be stated in writing. Ask whether the partner owns any of the inventory or technology it recommends, since a conflict of interest is easier to manage when it is on the table.
\nFinally, the supply path. For programmatic buys, ask the partner to explain which exchanges and resellers touch your ads. Public industry files such as ads.txt and sellers.json let you check whether a seller is authorized for a given publisher. Your contract should give you the right to ask and to audit.
\nAdd a right-to-audit clause, a data and creative handover clause on termination, and a requirement that all work product, including audiences, scripts, and naming conventions, belongs to you. Have counsel review the final language.
\nRACI stands for Responsible, Accountable, Consulted, Informed. For each task, exactly one party is Accountable (the neck on the line), one or more are Responsible (they do the work), others are Consulted before a decision, and others are Informed after. The most common failure in mixed models is not bad work. It is two parties each assuming the other one was Accountable.
\nA workable starting version for a hybrid arrangement looks like this.
\n| Task | Brand manager | Agency | Freelancer | In-house analyst | |---|---|---|---|---| | Annual budget and goals | A | C | I | C | | Channel strategy | A | R | C | C | | Daily bid and budget changes | I | A/R | I | C | | Creative production | A | C | R | I | | Tracking and tag integrity | C | R | I | A | | Invoice reconciliation | C | R | I | A | | Compliance review of ads | A | R | R | C | | Monthly performance review | A | R | I | R |
\nTwo rules keep the chart honest. Every row has one A and only one. And the person who audits the numbers should not be the person who produces them, which is why tracking and reconciliation sit with the in-house analyst even when the agency does the daily work. Separation of duties is a boring phrase that finance people love, and for good reason.
\nMany engagements fail in the first quarter, and not because the partner is weak. The brand did not hand over what the partner needed.
\nIn the first two weeks, give the partner a short brief with business goals, margins by product, seasonality, the sales team's real follow-up behavior, and anything that failed before. Margin is the piece most often withheld, and it is the piece that lets a partner optimize for profit rather than for volume. Share account access in the structure described earlier, and have both sides confirm in writing who can do what.
\nIn weeks three to six, run a controlled start. Pick one or two campaigns, define success in advance, and agree on the reporting cadence and format. Ask for a tracking audit before any money moves. A partner who inherits broken conversion tracking will optimize toward noise.
\nIn weeks seven to 12, hold a formal review. Compare results against the targets set at the start, review the first transparency statements against actual invoices, and check the RACI against reality. If the review reveals problems, fix them now. Problems ignored in the first quarter become habits in the second.
\nSwitching from an agency to an in-house team overnight is how brands lose a season of performance. The safer path is staged, with each stage proven before the next begins.
\nStage one is visibility. Without changing who does the work, take ownership of the accounts, set up your own reporting, and reconcile invoices. Often this stage alone fixes the problems that triggered the conversation, and it costs almost nothing.
\nStage two is shadowing. Hire or assign one in-house specialist who sits in every strategy call, reviews every change, and learns the account. The agency continues to execute. Expect a few months here.
\nStage three is the split. Move the work that is most tied to your context, such as first-party data, email, and reporting, to the in-house team. Leave the heaviest specialist work, perhaps programmatic buying or video production, with the agency or a freelancer. Define the interfaces in the RACI.
\nStage four is the test. Assign one campaign or one channel to the in-house team with the agency running a comparable one. Compare results on the agreed metrics for at least a full buying cycle. Suppose the in-house channel matches the agency within a tolerance you set in advance, say within 10 percent on cost per acquisition. That is your signal to continue. If it trails badly, you have learned something cheaply.
\nStage five is the handover. Transfer the remaining work on a calendar with overlap, keep the agency on a reduced scope or short retainer for questions, and record every procedure in your SOP library.
\nThroughout, track two numbers: total cost of the operation including salaries, tools, and management time, and performance on the metrics you set. If cost goes down and performance holds, move on. If cost goes down and performance falls, you have not saved money. You have moved it.
\nAgencies: refusing to give you account ownership; vague answers about who buys the media and at what price; reports that show only platform-reported results with no link to your sales or finance data; senior people appearing only in the pitch; a fee structure that rewards spend growth with no efficiency target; and pressure to commit to long minimum terms before results are shown.
\nFreelancers: no written scope or deliverables; one person with exclusive access to all your accounts; reluctance to document procedures; payment requests that bypass your normal process; and claims of results with no description of how they were measured. Also watch for the opposite problem, which is overreliance on one person you like. Liking someone is not a continuity plan.
\nIn-house teams: reports built by the people being judged by them; a team that declines training or outside review; heavy dependence on a single employee; tools and credentials stored in personal accounts; and a habit of adopting every new platform feature without a test plan.
\nRetainer: a fixed monthly fee for an agreed scope of work, regardless of how much media is bought. Commission-based fee: compensation calculated as a share of media spend, which can reward higher spend whether or not results improve. Principal-based buying: a practice in which an agency purchases media in its own name and resells it to the client, often at a markup that should be disclosed. RACI: a responsibility chart that names who is Responsible, Accountable, Consulted, and Informed for each task. Statement of work: a written document that defines deliverables, timelines, fees, and acceptance criteria for an engagement.
\n1. Choosing a partner on the strength of the pitch team, then discovering that a different, more junior team runs the account. 2. Letting the agency or freelancer create the advertising accounts under their own business, so the brand has no history or bargaining position when the relationship ends. 3. Hiring one in-house generalist and expecting the skill depth of an eight-person agency team. 4. Comparing models only on fees, ignoring internal salary, tools, recruiting, and management time. 5. Moving everything in-house in one step, with no overlap period and no comparison campaign to prove the team is ready.
\n1. Own every account, tag container, and data property under your company's name, and grant partners delegated access only. 2. Score the models with a weighted matrix and rerun it each year, since the right answer changes as budgets and skills change. 3. Write transparency terms into every contract: platform-level cost statements, disclosure of reseller structures, audit rights, and handover on exit. 4. Assign exactly one Accountable person for each recurring task, and separate the person who produces numbers from the person who verifies them. 5. Document every procedure as it is learned, so knowledge lives in your company rather than in someone's head.
\n1. Ask any agency to walk you through one real invoice, from platform charge to what you were billed. The ease of that conversation tells you more than a case study. 2. Pay freelancers through the same approval process as every other vendor. Informal payment habits create informal accountability. 3. When you hire the first in-house specialist, hire for analytical skill and curiosity before platform experience. Platforms change; habits of checking do not. 4. Keep one backup partner warm, even a freelancer on a small project, so a departure never leaves you without help. 5. Set a quarterly review with fixed questions: cost per output, time to deliver, quality score, and performance against target. Consistent questions reveal trends that ad hoc meetings miss.
\nMarkdown with Pandoc (GPL license) is a simple way to build your operating manual. Write each SOP, RACI chart, onboarding checklist, and transparency checklist in plain Markdown text files. Pandoc converts them into formats such as PDF, Word, or web pages whenever a partner or auditor needs a copy. Plain text files are easy to compare after edits, easy to store, and free of lock-in. Every tool named here is free and open source. Where you also need project tracking or reporting, this volume recommends no further open source option for this task, so a shared spreadsheet maintained by your analyst is a reasonable stand-in.
\nThe choice among agency, freelancer, and in-house models is a decision about control, speed, cost visibility, and skill depth, not only about talent or price. Each model fits certain circumstances, and agencies deserve fair credit for breadth and pattern recognition when their terms are clean. A weighted matrix makes the tradeoffs explicit, while transparency clauses, account ownership, and a RACI chart keep any model honest. Onboarding well in the first 90 days prevents most early failures. When moving work in-house, staged steps with a comparison test protect performance. Watching for red flags on every side, including your own briefs, keeps relationships healthy and the finance partner satisfied.
\n1. Compare models on control, speed, cost visibility, and skill depth before you compare price. 2. A weighted scoring matrix turns opinion into a documented decision and shows when a hybrid is the better answer. 3. Account ownership, platform-level cost reporting, and audit rights belong in every contract regardless of model. 4. One Accountable owner per task, with verification separated from production, prevents most handoff failures. 5. Move work in-house in stages, and advance only after a comparison campaign shows the team can hold performance.
\nThis week, move every advertising account and tag container under your company's name, then ask your current partner to reconcile one month of platform invoices to its bill to you.
\n1. Pandoc User's Guide. John MacFarlane and contributors. Documentation for the GPL-licensed document converter. 2. IAB Tech Lab. ads.txt specification, current version. Standard for declaring authorized digital sellers. 3. IAB Tech Lab. sellers.json specification, current version. Standard for disclosing the sellers in a supply chain. 4. Federal Trade Commission. Guides Concerning the Use of Endorsements and Testimonials in Advertising, 16 CFR Part 255. 5. Project Management Institute. A Guide to the Project Management Body of Knowledge, the section on the responsibility assignment matrix.
"},{"id":"chapter-53-contracts-insertion-orders-and-fee-transparency","title":"Chapter 53: Contracts, Insertion Orders, and Fee Transparency","html":"By the end of this chapter, the reader will be able to: 1. Identify the documents that make up an advertising agreement and explain which one wins when they conflict. 2. Explain how cancellation, makegood, and discrepancy clauses move financial risk between buyer and seller. 3. Draft viewability, fraud, data ownership, and indemnity language that a lawyer can then refine. 4. Compare commission, retainer, fixed fee, and performance fee structures by the behavior each one rewards. 5. Evaluate a vendor's disclosure of principal-based trading and rebates against a written standard. 6. Build an audit right and a pre-signature checklist that fit your actual spend.
\nMost advertising disputes are lost before anyone reads the contract. I will defend that, but the objection to it deserves a hearing first.
\nThe argument goes like this. Contracts are slow, relationships are fast, and the people who buy media for a living rarely look at a clause after the first meeting. A good partner who respects you will fix a bad delivery without being asked. A bad partner will find a way around any paragraph you write. Lawyers cost money, redlines cost weeks, and a brand that is trying to hit a launch date cannot wait on a negotiation over a discrepancy window. If trust is the real control, paper looks like theater.
\nSome of that is true. Trust matters. Speed matters. A contract will not turn a careless vendor into a careful one.
\nBut look at what the argument assumes: that the relationship stays the same people. It does not. The account lead who made you promises moves to another client. The agency is acquired and the new owner has different incentives. A publisher changes its ad server and your delivery reports start telling a different story. When any of that happens, the only thing that still speaks for you is the document you signed. Trust is how you feel on day one. The contract is what you can enforce on day 300.
\nThere is a second problem. Advertising paper is written by the seller, and sellers draft for their own protection. Standard terms often allow them to bill you for impressions you never received, to refuse refunds after a short window, and to keep the discounts they earned using your money. Signing quickly does not preserve the relationship. It quietly sets the price of the next disagreement.
\nThis chapter walks through the terms that matter, the fee structures you will meet, the disclosure you should demand about how your money is used, and the audit rights that make all of it real. It is operational guidance, not legal advice. Every contract in this chapter should be reviewed by counsel before you sign it.
\nA single campaign usually sits on several documents, and they do not all carry equal weight. There is a master services agreement that governs the relationship, an insertion order (the campaign-specific purchase document, often shortened to IO) that lists what you are buying, and a set of standard terms that the seller links to from the bottom of the IO. Sometimes there is also a statement of work for services that are not media.
\nThe first question for any buyer is: if these documents disagree, which one controls? Many sellers write their standard terms to override anything on the IO. That sounds technical. In practice it means the cancellation promise you negotiated by email can be erased by a web page you never opened.
\nFix this with an order of precedence clause. State that the master agreement controls, then the signed IO, then the statement of work, and that nothing in linked or online terms changes a signed document unless you sign the change. Ask also that the seller send you the version of any linked terms in force on the day you sign, and keep a copy. Web pages change. Your file should not.
\nThree clauses decide who absorbs the cost when a campaign does not go to plan.
\nCancellation sets how long before a flight starts you can cancel without paying. Direct deals with publishers often carry notice periods measured in weeks, with heavier penalties for sponsorships and premium placements. Programmatic buying through a platform is usually more flexible, but check what happens to committed minimums. Suppose you promise 60,000 dollars of spend over a quarter and cancel half way through. Without a clean cancellation clause, you may owe the remainder as a shortfall. Ask for staged exit rights: full cancellation up to a set date, partial cancellation after that, and a plain statement of what counts as spend already incurred.
\nA makegood is the seller's remedy when delivery falls short, usually extra impressions or placements in a later period. The trouble is how vague it can be. Define the trigger (for example, delivery below a stated share of booked impressions), the timeframe (inside the flight or within a fixed window after it), and the quality standard (equal or better placement and audience). Add a clause giving you the option of a credit or refund when a makegood cannot deliver at equal value, since a replacement run in a dead week of the calendar is not a remedy at all. An ad that missed a playoff weekend cannot be repaired with a spot in the offseason.
\nThe discrepancy window is the period during which you may dispute the seller's numbers. Your count and their count will differ, because ad servers measure differently. Contracts handle this by naming a source of truth and a tolerance. Common practice is to treat the neutral ad server as the record and allow a small variance before anyone argues. Whatever the figure, write it down. Then give yourself enough time to dispute: some terms allow only a few days after an invoice, which is shorter than most finance teams need to find the invoice.
\nViewability and invalid traffic are where a vague promise becomes an expensive one. Viewability means an ad had a fair chance of being seen. Invalid traffic means views and clicks produced by bots or other non-human activity. A contract that says only \"premium inventory\" or \"brand safe delivery\" gives you nothing to enforce.
\nWrite the standard as a measurable test. Name the measurement vendor, the definition of a viewable impression you will use (the industry bodies publish accepted definitions, and your counsel can cite them), the minimum rate you expect, and the party that pays for verification. Then say what happens below the line. A workable remedy is that you do not pay for impressions flagged invalid or unmeasurable by the agreed verifier, and that the seller refunds or credits them within a set number of days.
\nFraud deserves a separate clause. Specify that the seller will not knowingly sell non-human traffic, that it will cooperate in investigations, and that you may withhold payment for traffic flagged as invalid while a dispute runs.
\nFor brand safety, list your exclusions explicitly instead of pointing at a vague category. Include categories, keywords, and any specific content you refuse, and apply a conservative default when children, health, or sensitive subjects are anywhere near the placement. Require removal within a short window of notice and a credit for impressions served against your exclusions.
\nEvery campaign produces data: who saw what, who clicked, who converted, who raised a hand. If the contract is silent, you may find your vendor claims the right to use it, combine it with other clients' data, or keep it after you leave. The principle is simple. Data you supplied, and data generated by your campaigns about your customers, belongs to you. The seller may use it to deliver your campaign and for nothing else without your written consent.
\nWrite four things into the document. First, ownership: your first-party data and campaign data remain yours. Second, permitted use: the vendor may use it only to perform the services. Third, return and deletion: on termination the vendor hands you the data in a usable format (not a PDF screenshot of a dashboard) and deletes its copies within a stated number of days. Fourth, security: the vendor protects the data to a stated standard and tells you promptly if it is exposed. Privacy rules differ by place and category, so your counsel decides what language the law requires.
\nIndemnity is the promise to cover the other side's losses if something goes wrong. Both sides usually give it, and the question is what each side covers. Sellers should cover claims that their inventory, technology, or conduct caused harm. You should cover claims arising from your own creative and your claims about your product. Ask for the seller's indemnity to include invalid traffic and data misuse, and be wary of one-sided versions in which only you indemnify. Also look for caps on liability. A cap set at the fees for one month can make a promise worth very little. Counsel will tell you what is reasonable for your exposure.
\nEvery fee structure rewards something. Your job is to notice what.
\nA commission on spend pays the vendor a share of what you spend. It is simple, and it quietly rewards spending more. A retainer is a flat monthly fee for a defined scope. It protects you from spend-driven incentives, but it can reward doing the minimum unless the scope is specific. A project fee suits one-time work such as a campaign build. A performance fee pays for a result such as qualified leads, which sounds ideal until you find that the vendor controls how the result is counted.
\nA small example makes the difference visible. Suppose you spend 100,000 dollars on media with a vendor charging a 15 percent commission. If the fee is added to the spend, you pay 115,000 dollars and the vendor earns 15,000. If the fee is taken out of a 100,000 dollar budget, the publishers receive 85,000 dollars and the vendor keeps 15,000. Both are called a 15 percent fee. They are not the same deal. Ask for the share of your total budget that reaches working media, meaning the money that actually buys ad space.
\nWhere possible, tie some fee to the outcome you care about, but use a counted event that you control, such as sales recorded in your own system. Keep the base fee large enough that the vendor stays honest about a bad month, and write the formula into the contract, with an example calculation beside it.
\nIn an agent relationship, the vendor buys media on your behalf and charges you a disclosed fee. In principal-based trading, the vendor buys inventory in its own name and resells it to you. The vendor takes the price difference as margin. That is a legal and common model. It is also one where you may not know what the vendor paid.
\nPrincipal trading is not wrong. The vendor takes inventory risk, and sometimes you get a lower price than you could have found yourself. The problem is invisibility. If you believe you hired an agent and you are in fact buying from a reseller, the vendor's interest in the price runs opposite to yours.
\nAsk for a disclosure clause with these parts: a statement of whether any portion of your buy will be sourced on a principal basis; notice before any principal inventory is placed in your plan; the vendor's cost on request or a stated maximum margin; and your right to opt out of principal buys. For rebates, volume incentives, and other money the vendor receives from publishers or platforms because of your spend, require disclosure and pass-through to you, or a written agreement that the vendor may keep specified amounts.
\nIf the vendor refuses all of this, you can still decide to proceed. Just do it knowing what you bought: a managed service priced in a way you cannot inspect.
\nAn audit right on paper that cannot be used is decoration. Write one that works.
\nState what you may examine: invoices, publisher statements, ad server logs, rebate records, and any records showing the vendor's cost for media you bought. State how often you may audit (once a year plus once more if a discrepancy is found), the notice you will give, and who may do it. Allow an independent third party under confidentiality. Say who pays: normally you do, unless the audit finds an overbilling above a stated threshold, in which case the vendor pays the cost and refunds the difference with interest.
\nYou do not need a large firm to start. A finance analyst with the invoice, the platform report, and your own server log can run a first check in an afternoon. Compare billed impressions with delivered impressions, billed rate with the contracted rate, and total fees with the disclosed fee. Differences are not proof of wrongdoing. They are questions worth asking.
\nBefore anyone signs, ask whether the documents answer each of these:
\nInsertion order: A written purchase document that lists the campaign, the dates, the rates, the volumes, and the terms of sale between a buyer and a media seller. Makegood: A remedy offered when delivery falls short, usually additional placements or impressions of equal value, or a credit. Discrepancy window: The period during which a buyer may dispute the seller's delivery or billing figures. Principal-based trading: A model in which a vendor buys media in its own name and resells it to the client at a price the vendor sets, keeping the difference. Working media: The share of a total budget that actually pays for ad space or placements, as opposed to fees, tools, and other costs.
\n1. Treating linked online terms as harmless, then discovering they override the cancellation promise made on the IO. 2. Accepting a discrepancy window of a few days, which is shorter than the time finance needs to receive and check an invoice. 3. Writing \"brand safe\" or \"premium\" into the IO without a measurable standard, a named verifier, or a remedy. 4. Comparing fees by the headline number instead of asking what share of the budget reaches working media. 5. Signing a contract that is silent on data return, then finding no right to the campaign data after the relationship ends.
\n1. Put an order of precedence clause in every agreement and archive the version of any linked terms in force on signing day. 2. Name a neutral source of truth for delivery counts, plus a stated tolerance and a dispute window long enough to use. 3. Write measurable viewability and invalid traffic standards, with refund or credit as the remedy. 4. Require written disclosure of principal trading, rebates, and vendor margin, and pass through any money earned on your spend. 5. Keep a version-controlled clause library so every contract starts from your standards instead of the seller's.
\n1. Send the seller your own terms first when your spend is large enough to justify it, since the first draft sets the baseline everyone negotiates from. 2. Ask what the vendor would have to change for you to sign without the audit right. The answer reveals more than the clause does. 3. Put a worked fee example inside the contract so nobody can argue later about what the formula meant. 4. Calendar the cancellation and renewal dates the day you sign, because auto-renewals reward the party that forgets. 5. Ask finance to compare three invoices against three platform reports before the first renewal. Patterns show up quickly.
\nPandoc (GPL) is a free document converter that turns plain text files into Word, PDF, or web formats. For contracts, it lets you keep a clause library as simple text: one file per clause, such as cancellation, makegood, data ownership, and audit. When a new deal starts, you assemble the clauses you need into one document and export it in the format counsel prefers. The same library works for every vendor, so your standards stay consistent and your redlines are easy to compare. Pandoc does not give legal advice and does not replace a lawyer. It only makes your own approved language faster to reuse and easier to maintain. No other tool is needed for this chapter.
\nAn advertising agreement is a stack of documents, and the order of precedence decides which one wins. The terms that shift money are cancellation, makegoods, discrepancy windows, viewability and fraud standards, data ownership, and indemnity, and each should be written as something measurable with a stated remedy. Fee structures reward different behavior, so the right question is how much of the budget reaches working media. Principal-based trading and rebates are legitimate when disclosed and a problem when hidden. Audit rights make every other clause enforceable, and a checklist reviewed before signing keeps the document honest. All of this is operational guidance, and counsel must review the final language.
\n1. Trust sets the tone on day one, but only the signed document speaks for you after the people change. 2. Cancellation, makegood, and discrepancy clauses decide who absorbs the cost of a campaign that goes wrong. 3. Vague promises such as brand safe or premium are unenforceable until they become measurable standards with remedies. 4. Compare fee structures by what they reward and by the share of budget that reaches working media. 5. Principal trading, rebates, and audit rights should be disclosed in writing, since you cannot negotiate what you cannot see.
\nCopy the checklist in this chapter into a one-page memo, answer every line from the actual documents, and send it to counsel with the contract before your next signature.
\n1. Interactive Advertising Bureau and Media Rating Council. Viewable Impression Measurement Guidelines. Industry definitions used in viewability clauses. 2. Interactive Advertising Bureau Tech Lab. ads.txt and sellers.json specifications. Public records of authorized sellers. 3. Association of National Advertisers. Programmatic Media Supply Chain Transparency Study. Industry discussion of fee and principal-trading disclosure. 4. MacFarlane, J. Pandoc User's Guide. Documentation for the GPL-licensed document converter.
"},{"id":"chapter-54-team-structure-intern-pipelines-and-training","title":"Chapter 54: Team Structure, Intern Pipelines, and Training","html":"By the end of this chapter, the reader will be able to: 1. Describe the three shapes of a small digital advertising team and match one to a given budget. 2. Assign ownership of channels, creative, and data to pods so that no task has two owners or none. 3. Design an intern pipeline in which supervision, review gates, and data access limits are set before the first shift. 4. Build a four-tier training curriculum and sequence it against real work. 5. Write a standard operating procedure (SOP) that a new hire can follow without asking for help. 6. Evaluate a career ladder for fairness, clarity, and retention value.
\nHere is a piece of arithmetic that should bother you. Suppose an intern gives you 15 hours a week for 12 weeks. That is 180 hours of labor at a cost of nothing, or close to it. Now suppose a senior buyer spends three hours a week supervising, and reworks half of what the intern produces at a rate of one senior hour for every two intern hours. Add it up. Supervision costs 36 hours. Rework costs another 45. You have spent 81 senior hours to receive 180 junior hours, and a good share of those junior hours were spent learning, not producing. Free labor just turned out to be a loan against your best person's calendar.
\nMost brand managers read that and conclude interns are a bad idea. That is the wrong conclusion. The right one is that unstructured interns are a bad idea. Structure changes every number in the example. When the intern works from a written procedure, rework drops. When review happens at a gate rather than after a mistake goes live, senior hours get spent once. When the intern's task is picked because it is repeatable (naming conventions, tracking audits, report assembly), the senior hour per two intern hours becomes one per six.
\nThe common assumption is that a team is a list of people. It is not. A team is a set of agreements about who decides, who checks, who can touch which data, and how a stranger learns the job. A five-person team with those agreements outperforms a nine-person team without them, and it does so quietly, without heroics.
\nThis chapter covers roles and pods for a small digital team, a training curriculum, an intern pipeline that produces real output safely, SOP writing, and career ladders. I am a veteran, and the Army taught me something that applies here: a unit that cannot function when its best person is out sick is not a unit. It is a person with an audience.
\nTeam shape follows budget, but not in a straight line. Spend alone does not tell you how many people you need. Complexity does: the number of channels, the number of markets, how many creative variants you ship a month, and how much compliance review your category demands.
\nThe generalist shape suits a small budget. One person plans, buys, builds the ads, and reads the reports. This works until the account runs more than two or three channels. The generalist's danger is not skill. It is that nobody checks the work. The person who sets up a campaign is also the person who approves it, and errors in targeting or tracking can run for weeks.
\nThe specialist shape suits a middle budget. A media buyer, a creative producer, and an analyst each own a function. Three people can look like a team on a chart and still behave like three freelancers who share a calendar. The fix is a weekly rhythm in which each person states what they shipped, what they are blocked on, and what they need checked.
\nThe layered shape suits a large budget, with a lead, channel managers, specialists, and analysts. Layers add coverage but also add handoffs, and every handoff is a place where context leaks out. Before adding a layer, ask whether a clearer procedure would solve the same problem for free.
\nA useful test is the vacation test. Pick any person, remove them for two weeks, and list what stops. If more than two things stop, you have a single point of failure, and your next hire or your next SOP should address it.
\nA pod is a small group that owns an outcome end to end. In a small team, three pods cover most of the work.
\nThe acquisition pod owns planning, buying, and optimization for paid channels. The creative pod owns concepts, production, variant naming, and the asset library. The data pod owns tracking, tag audits, the reporting layer, and the test log. A pod can be one person wearing one hat, and in a team of four or five it often is.
\nWhat matters is the rule of single ownership. Every recurring task has one named owner and one named checker, and those two are never the same person. When the buyer builds a campaign, the analyst verifies tracking before launch. When the producer exports a creative, the buyer verifies specs and disclosures before upload.
\nDraw a simple ownership grid on one page. List the 15 or so recurring tasks down the side (budget pacing, tag audit, weekly report, creative handoff, invoice reconciliation, and so on) and put an owner and a checker beside each. Gaps and duplicates become visible in about ten minutes.
\nInterns slot into pods as supporting hands, never as owners. That one rule prevents most of the trouble this chapter describes.
\nA pipeline has stages, and each stage has an exit condition. Without exit conditions, interns drift through a summer and leave knowing little more than when they arrived.
\nRecruitment starts with the schools that already have marketing, communications, sport management, and data programs, along with community colleges and military transition programs. Screen for curiosity and reliability before skill. A candidate who answers email within a day and finishes a small take-home task carefully is worth more than one with a polished resume and a slow response time.
\nOnboarding takes the first week. The intern reads the brand's one-page guide, the naming conventions, and the ownership grid, then shadows a pod and completes a practice task in a sandbox account with fake spend and fake data.
\nProduction work comes next, and it starts narrow. Good first assignments are repeatable and checkable: auditing URLs and tracking parameters across live ads, assembling the weekly report from an approved template, tagging assets in the library, drafting first-pass copy variants from a brief, and logging test results. Each of these has a right answer a supervisor can verify in minutes.
\nConversion is the last stage. Set the criteria in advance: how many tasks passed review on the first try, whether the intern followed procedure without reminders, whether the pod wants them back. Interns who meet the criteria get a first look at open roles. This is how the pipeline pays you back, because a converted intern already knows your tools, your naming, and your standards.
\nSafety is a design decision, not a hope. Three controls do most of the work.
\nSupervision means a named person is responsible for each intern, and that person has protected time. A workable pattern is a short daily check-in, one longer weekly review, and an agreed channel for questions. Suppose a supervisor holds two interns. Budget about five hours a week per intern at the start and let it shrink as trust builds. If the supervisor's calendar cannot absorb that, take fewer interns.
\nQA gates are checkpoints where work stops until someone else approves it. Place three of them. The first is pre-launch: no ad, audience, or budget change goes live without the checker's sign-off. The second is pre-send: no report leaves the building without a senior read. The third is pre-spend: any change that raises daily spend beyond an agreed ceiling requires the owner's approval. Interns can prepare work for all three gates. They cannot pass their own work through them.
\nData access limits matter most. Give interns the least access that lets them do the assigned task. That usually means read-only reporting views, no ability to export customer lists, no access to payment details, and no admin rights on ad accounts. If an intern needs a dataset, supply a sample with identifiers removed.
\nKeep a conservative default on audiences. Anyone touching targeting for products near children, health, or finance works under a senior's direct eyes. Operational guidance only here; your counsel and your privacy policy decide the legal particulars.
\nTraining fails when it is a pile of videos. It works when it is a sequence tied to the work a person will do in the next 30 days.
\nTier one is foundations: how auctions and placements work, what the core metrics mean and where each one misleads, how tracking parameters flow from ad to landing page to report, and basic spreadsheet fluency. Plan about two weeks, with a short quiz at the end of each module.
\nTier two is channel skills: search, social, video, email, and display, one at a time. Each module ends with a practical task in a sandbox, such as building a small campaign structure and defending the naming choices.
\nTier three is tools and data: the reporting layer, the tag audit routine, basic SQL (the language used to ask a database questions), and reading a test result without overclaiming. Learners should leave able to explain why a result might be noise.
\nTier four is professional skill: writing a clear brief, giving and receiving review feedback, running a short meeting, and explaining a number to someone who dislikes numbers. In my experience this tier is the one people skip, and it is the one that separates a good specialist from a future lead.
\nDeliver the curriculum as plain documents with a course shell around them. Each module needs a stated outcome, a reading, a task, and a pass standard. A new hire who finishes tier one in week two and tier two in week six is learning against real deadlines, which beats any binge of lectures.
\nAn SOP is a written procedure for a recurring task. Most are ignored because they are too long, too vague, or hidden.
\nA good SOP fits on one page, opens with its purpose in one sentence, and names its owner and checker. Then it lists numbered steps, each starting with a verb and each small enough to verify. It ends with a short list of known failure points and what to do about them.
\nConsider the tag audit. A weak SOP says \"check that tracking is working.\" A usable one says: open the live ad, click through, confirm the landing page address carries the agreed parameters, confirm the analytics tool registers the visit within five minutes, record the result in the audit sheet, and escalate any mismatch to the data pod the same day. Anyone can do it, anyone can check it, and anyone can improve it.
\nTest every SOP by handing it to a person who has never done the task and watching silently. Wherever they hesitate, the document is wrong. Fix it, and date the revision.
\nStore SOPs in one place, in a plain text format with version history, so changes are visible and reversible. Review each one on a fixed schedule, quarterly for fast-moving channels and twice a year for the rest. A procedure nobody has reviewed in a year is a rumor with a title.
\nSmall teams lose people for a predictable reason: nobody can see what comes next. A ladder answers the question before it is asked.
\nKeep it short. Four or five levels cover most small teams: associate, specialist, senior specialist, pod lead, and a team lead or director. For each level, write down the scope (what the person owns), the judgment (what they decide alone), the output (what they produce reliably), and the influence (whom they teach). Use observable behavior, not adjectives. \"Runs the weekly test review and writes up the decision\" is observable. \"Strong strategic thinker\" is not.
\nPromotion should follow demonstrated work at the next level for a sustained stretch, not tenure and not a good quarter. Publish the criteria so that any employee can read them and any manager can be held to them. Pay bands attached to levels reduce quiet unfairness, because two people at the same level doing the same work see the same range.
\nLadders also need a second track. Not every strong specialist wants to manage. Offer a path where deep expertise in analytics, creative, or channel craft carries the same status and pay as leading a pod. Otherwise you promote your best buyer into a job they will dislike and lose them anyway.
\nFor interns, the ladder starts one rung below associate. A converted intern enters at associate with credit for the training already completed, which makes the offer both fair and cheap.
\nOutside classic internships, some brand managers draw talent from structured workforce programs. Military transition programs in the style of SkillBridge place service members in civilian roles during their final months of service, and the Dakdan internship and SkillBridge-style programs are one example of an in-house route for organizations that want a trained, supervised pipeline. This is informational: the model is worth understanding whether you use that route or build your own.
\nThe operating lessons carry over to any program. Placement works best when the host organization supplies a defined role, a named supervisor, and a written training plan. Participants arrive with discipline, comfort with procedure, and a habit of asking for the standard before they begin, and those traits suit SOP-driven digital work well.
\nWhatever the source, run workforce placements under the same controls as interns: limited data access, QA gates, and a conversion decision made on evidence. Confirm eligibility rules, wage rules, and any agreements with your counsel and the program administrator before you commit. A pipeline that skips paperwork will eventually produce a problem that costs more than the labor saved.
\nPod: A small group that owns an outcome from start to finish, such as paid acquisition or creative production. SOP: A standard operating procedure, a short written set of steps for a recurring task, with an owner and a checker. QA gate: A checkpoint at which work stops until a second person approves it. Least-privilege access: The practice of giving a person only the system permissions needed for the assigned task. Career ladder: A written set of levels, each with defined scope, judgment, output, and influence, used to guide promotion and pay.
\n1. Letting the person who builds a campaign also approve it, which removes the only check that catches tracking and targeting errors before money is spent. 2. Assigning interns open-ended projects (\"help with social\") instead of narrow, checkable tasks with a right answer a supervisor can verify quickly. 3. Granting interns full ad account and customer data access for convenience, then discovering an export or an accidental change after the fact. 4. Writing SOPs as long essays, never testing them on a newcomer, and never reviewing them again. 5. Promoting the best specialist into management by default, losing a great practitioner and gaining a reluctant manager.
\n1. Draw a one-page ownership grid with a named owner and a different named checker for every recurring task, and update it with each staffing change. 2. Give every intern a named supervisor with protected time, and cap the number of interns each supervisor holds. 3. Set the exit conditions for each pipeline stage in writing before the program starts, including the criteria for a conversion offer. 4. Tie each training module to real work in the next 30 days, with a stated outcome and a pass standard. 5. Store SOPs in plain text under version control and review each on a fixed calendar.
\n1. Run the vacation test twice a year: remove each person on paper for two weeks and fix whatever stops. 2. Start new interns on audits and report assembly, because those tasks teach the account's structure while carrying little risk. 3. Use a sandbox account with fake spend for practice, so mistakes in week one cost nothing. 4. Ask a departing intern or employee to rewrite the SOP they found most confusing, since fresh eyes expose the gaps faster than any review. 5. Publish pay bands next to career levels, and watch your offer acceptance rate and early resignations for signs the ladder is not believed.
\nTwo free, open source tools cover most of what this chapter needs. Markdown with Pandoc (GPL) lets you write SOPs, training modules, and the ownership grid as plain text files, then convert them to PDF, Word, or web pages from a single source, so one edit updates every format. Moodle (GPL v3) is a learning management system you can host yourself: it delivers modules in sequence, runs quizzes, tracks completion, and records who passed which tier. For small teams, plain files may be enough at first, and Moodle earns its place once you are onboarding several people a year.
\nTeam design is a set of agreements more than a headcount. Shape follows complexity, pods own outcomes, and every recurring task has one owner and a different checker. Interns become productive when their work is narrow and checkable, when supervision is scheduled, when QA gates stand between their drafts and live spend, and when data access is limited to what each task needs. A four-tier curriculum tied to real deadlines, one-page SOPs that a newcomer can follow, and a ladder with observable criteria and two tracks keep the capability inside the organization. Workforce programs can add to the pipeline when they run under the same controls and with counsel confirming the rules.
\n1. A team is a set of agreements about who decides, who checks, who touches which data, and how a stranger learns the job. 2. Unstructured interns cost more senior time than they return; structured interns working from procedures return far more. 3. Interns prepare work for QA gates and never approve their own output. 4. Training works as a sequence tied to upcoming real work, with a pass standard for every module. 5. A career ladder built on observable behavior, published criteria, and two tracks reduces losses of your best people.
\nThis week, draw the one-page ownership grid with an owner and a different checker beside every recurring task, then take your three most repeatable jobs and write each as a one-page SOP an intern could follow on day one.
\n1. U.S. Department of Labor, Wage and Hour Division. Fact Sheet 71: Internship Programs Under the Fair Labor Standards Act. Operational reference for evaluating unpaid internships; counsel decides. 2. U.S. Department of Defense. DoD Instruction 1322.29, Job Training, Employment Skills Training, Apprenticeships, and Internships for Eligible Service Members. Program rules for military transition placements. 3. National Institute of Standards and Technology. Special Publication 800-53, Security and Privacy Controls for Information Systems and Organizations. Source for the least-privilege access principle. 4. Pandoc Project. Pandoc User's Guide. Documentation for document conversion from Markdown. 5. Moodle Project. Moodle Documentation. Administration and course design guidance for the open source learning platform.
"},{"id":"chapter-55-fraud-chargebacks-and-vendor-audits","title":"Chapter 55: Fraud, Chargebacks, and Vendor Audits","html":"By the end of this chapter, the reader will be able to: 1. Identify the main patterns of impression, click, lead, and affiliate fraud and the signals that expose each one. 2. Explain how payment disputes and chargebacks create cost exposure inside lead and affiliate programs. 3. Draft the contract terms that make a vendor audit possible before any money moves. 4. Apply a three-part audit protocol (logs, reconciliation, path checks) to a vendor invoice. 5. Design an audit cadence that matches vendor risk and spend. 6. Evaluate a fraud finding and assemble the documentation, notice, and credit request that follow it.
\nHere is the assumption I hear most often: \"My vendor runs a verification tool, so my traffic is clean.\" It sounds reasonable. It is also the assumption that costs brands the most money without ever showing up on a line item.
\nA verification tool inspects traffic against a set of rules and flags what looks invalid. It does that well for the categories it was built to catch. It does not see a lead form filled in by a paid human farm in a country the offer does not serve. It does not see an affiliate who dropped a tracking cookie on a shopper who was already at checkout. It does not see a reseller who relabeled cheap inventory as premium before it reached you. And the verification report itself is often configured, selected, or delivered by the same party you are paying.
\nSo the real situation is this. Fraud is not one problem with one detector. It is a family of behaviors that move to wherever the checking is thinnest. Chargebacks, meaning disputed payments pulled back through the card networks or recovered from a vendor through credits, are the financial echo of those behaviors. And the vendor audit is the only control that works across all of them, because it does not depend on guessing what the fraudster will try next. It simply compares what you were told with what happened.
\nThis chapter is operational guidance, not legal advice. Your counsel decides what your contracts allow and what you may assert. My job here is to show you what to look for and how to document it.
\nStart with a plain truth. Fraud follows the payment trigger. Whatever event you pay for becomes the event someone fakes.
\nIf you pay per thousand impressions, the cheapest fake is an ad loaded where no person can see it, such as a hidden frame or software that never opens a browser window. If you pay per click, the fake is automated clicking, or accidental clicks engineered by placing a button where a thumb lands. If you pay per lead, the fake is a form submission, either by script or by people paid pennies to type names. If you pay per sale, the fake is a claim on a sale that was going to happen anyway.
\nHere are the signals I train people to watch.
\nCall a sample. A phone number that rings to a dead line is a verdict no dashboard can overturn.
\nPicture a stadium on game day. Counterfeit tickets get caught because someone counts the people in the seats and compares that number with the tickets sold. Audits are seat counts.
\nAffiliate and lead programs deserve their own section because the vendor and the fraudster can be the same entity, and the contract often pays them to look busy.
\nConsider three affiliate patterns. Cookie stuffing drops your tracking cookie onto a visitor who never clicked an affiliate link, so a later purchase gets credited to the affiliate. Trademark bidding places search ads on your own brand name so the affiliate intercepts shoppers already looking for you. Coupon scraping posts your private discount codes on public pages, then collects commission when your own loyal customers search for a code at checkout.
\nLead programs have their own set. Lead stuffing floods the funnel with submissions that clear a basic validity check. Incentivized leads pay people to complete a form, so the contact is real but the interest is not. Duplicate recycling sells you the same person under slightly different spellings. Out-of-territory leads arrive from places your offer does not cover.
\nSuppose a regional services brand pays 40 dollars per qualified lead and buys 500 leads in a month. That is a 20,000 dollar invoice. If a phone sample of 50 leads finds that 20 are unreachable or duplicated, a large share of the batch will likely never become revenue. I would not extrapolate that into a precise loss. I would use it to justify a deeper pull of the full file.
\nSet a conservative default wherever the audience may include minors or a sensitive category. If a lead source cannot show where consent came from, pause it. Bad consent is a reason to stop buying, whatever the conversion rate says.
\nThe word chargeback carries two meanings in this business, and mixing them up causes confusion in vendor meetings.
\nThe first meaning is the consumer payment dispute. A cardholder asks the card issuer to reverse a charge. For a brand running lead and affiliate programs, the exposure comes from the way sales were produced. An affiliate who promises a result the product cannot deliver, hides a recurring charge in the fine print, or sells to someone who never intended to buy will generate disputes. Those disputes come back to your merchant account with fees attached. Processors typically watch the ratio of disputes to transactions, and a program that pushes that ratio too high can face tighter terms or the loss of the account. Rules vary by processor, so read your merchant agreement and ask counsel what it requires.
\nThe second meaning is the vendor recovery: a credit, refund, make-good, or reversal you request from a media vendor or network for traffic that should not have been billed. Contracts differ widely on dispute windows, notice form, and whether recovery is cash or future credit. You cannot negotiate those after the fact.
\nA spike in consumer disputes can be an early fraud signal for an affiliate or lead program, before any verification report notices anything. Ask finance for dispute counts by acquisition source each month. When one source carries a share of disputes far out of line with its share of sales, audit it first.
\nKeep the family-first lens here too. Subscriptions that renew without clear notice create disputes honestly earned. Fix the offer.
\nThe best audit I ever ran succeeded because of two sentences in a contract signed months earlier. Before you sign with any vendor, look for these terms and ask counsel to confirm the wording.
\nVendors sometimes resist log-level data over privacy concerns. Propose records without personal identifiers if privacy is the stated concern. A vendor who will not show any row-level support for a bill is telling you something about the bill. Name an owner for each vendor audit. Unowned audits never happen.
\nA vendor audit needs only three moves, performed in order.
\nLogs. Collect the raw records behind the invoice period: impressions served, clicks, conversions or leads, timestamps, placements, and the cost attached to each. Collect your own records too: your server logs, your tag data, your customer relationship system, and your payment processor reports. Never edit the originals.
\nReconciliation. Compare the two sets. Start with the biggest number and work down. Does the vendor's billed volume match the log volume? Does the log volume match your own count of arrivals? Do the conversions the vendor claims exist in your system as real customers or real qualified leads? Two systems always count slightly differently, so decide your tolerance before you look. A gap that varies in both directions is ordinary. A gap that always runs in the vendor's favor is a pattern.
\nSuppose a vendor bills nine hundred thousand impressions at six dollars per thousand, a 5,400 dollar invoice. Your ad server counted eight hundred ten thousand deliveries to that vendor's placements. The gap is ninety thousand impressions, worth 540 dollars. It might be ordinary counting loss, or it might not. You now have a number to ask about.
\nPath checks. Verify that the money traveled the route you were told it traveled. Ask where the ad actually ran, which domains and apps, and who sold the inventory at each step. Then check the public records of authorization. The sellers file and the authorized-seller file that publishers post describe who may sell their inventory and in what role. If a vendor claims premium placement on a publisher, but that publisher's published files do not list the seller in the chain, you have a mismatch to resolve.
\nHere is the worksheet I hand to new buyers.
\n| Question | Source A | Source B | Gap | Action | |---|---|---|---|---| | Billed volume equals logged volume? | Invoice | Vendor logs | Compare | Ask vendor to explain | | Logged volume equals our count? | Vendor logs | Our ad server | Compare | Check tag firing and filters | | Claimed conversions exist? | Vendor report | Our system | Compare | Sample and verify contact | | Domains match the plan? | Placement report | Insertion order | Compare | Flag unplanned sites | | Seller authorized by publisher? | Chain record | Public files | Compare | Request the seller's role |
\nDone monthly, it teaches you more about your vendors than any sales presentation.
\nHow often should you audit? The answer depends on how much a vendor handles and how much visibility you have into its work, not on a calendar tradition.
\nA starting cadence has three layers. The light layer is monthly: reconcile billed against logged volume for every vendor above a spend level you set, checking only the headline numbers. The deep layer is quarterly for your riskiest vendors: full logs, sampled contact verification for leads, domain review, and the path check. The triggered layer happens whenever something breaks pattern: a sudden drop in cost per lead, a jump in volume without a matching jump in sales, a spike in disputes, or a vendor who goes quiet when you ask for data.
\nRank vendors by risk first: spend, distance from the end customer, how little data they share, and how new they are. New vendors deserve more attention in the first 90 days. Long-standing ones deserve it too, since familiarity quietly turns into trust.
\nFinally, track your audit finding rate: of the audits you ran, how many surfaced a discrepancy worth a conversation? If the rate sits at zero for a long stretch, either your vendors are excellent or your audits are shallow.
\nThe first rule is the hardest one: stop and preserve before you accuse.
\nDo not call the vendor in anger. Do not delete anything. Work in this order.
\n1. Freeze the evidence. Export the logs, reports, screenshots, and invoices for the period, with dates and the names of the people who pulled them. Store copies somewhere that the original system cannot overwrite. 2. Quantify carefully. Calculate the affected volume and the billed amount for that volume, using the vendor's own rate card. Separate what you can prove from what you suspect. Present the first number as the claim and label the second as a question. 3. Stop the leak. Pause the affected sources while you investigate. You can resume later; you cannot un-spend money. 4. Send written notice. Follow the notice rules in your contract: who receives it, how, and within what window. State the facts, attach the evidence, name the invoice lines affected, and request the remedy the contract provides. A calm letter with attached rows is harder to dismiss than an angry call. 5. Ask for their data. Request the vendor's own logs for the disputed lines and a written explanation of the traffic source. 6. Agree the resolution in writing: a refund, a credit, a make-good with verified traffic, or removal of the offending source. 7. Close the loop inside your company. Log the finding and raise that vendor's audit layer.
\nBe ready for disagreement. A vendor may claim the traffic was valid or that your window has closed. Some arguments will be fair, so rest the claim on records you hold. If the amount is large or the vendor refuses to engage, bring in counsel before you decide what to do next.
\nIf you discover that your ads ran beside content aimed at children, or that leads included minors, treat that as a compliance matter first and a refund matter second.
\nInvalid traffic: Impressions, clicks, or events that do not come from a real person with a real chance to see or act on an ad, including automated traffic and accidental activity engineered by the placement. Cookie stuffing: Dropping an affiliate tracking cookie on a visitor who did not click the affiliate link, so that a later sale credits the affiliate. Reconciliation: Comparing two independent records of the same activity, such as a vendor invoice and your own logs, and explaining each gap. Make-good: A remedy in which the vendor delivers replacement verified traffic or exposure instead of refunding the cost of the invalid portion. Dispute ratio: The share of card transactions that customers contest, which payment processors watch when deciding whether a program is acceptable.
\n1. Trusting a vendor-supplied verification report as the audit, when the report is one input from a party with an interest in the result. 2. Reading the contract only after a problem appears, then learning the dispute window closed 30 days after the invoice date. 3. Auditing only click and impression counts while ignoring the quality of leads and the sales credited to affiliates. 4. Calling the vendor first and gathering evidence second, which gives them time to adjust their records and leaves you with an argument instead of a claim. 5. Treating consumer chargebacks as a finance-department problem, so the marketing team never sees which sources produce them.
\n1. Negotiate audit rights, log-level data, and a defined dispute remedy into every media and affiliate contract before the first insertion order is signed. 2. Reconcile billed volume against your own records every month for every vendor above your spend threshold, using a tolerance you set in advance. 3. Call a sample of leads from every lead source each cycle and record how many are reachable and in territory. 4. Keep an original, untouched copy of every export you use in an audit, with the date and the name of the person who pulled it. 5. Review dispute counts by acquisition source each month, and investigate any source whose share of disputes outruns its share of sales.
\n1. Plant one test lead or test purchase in each new lead or affiliate source and check whether it appears correctly in the vendor report. 2. Read the unplanned domains and apps in a placement report first, since those lines hold most of the surprises. 3. Ask each vendor to describe in writing who gets paid at each step if it buys your inventory through another seller. 4. Check affiliate search activity on your own brand name each week, because trademark bidding is quick to spot and slow to stop once it spreads. 5. When a vendor offers a credit to settle, ask what source it came from and confirm that source has been removed, or the credit simply buys you the next round of the same problem.
\nThree free tools cover most of an audit desk. DuckDB (MIT license) is a database that runs on your laptop and reads large log files directly, so you can join a vendor export against your own data with a short query and no server. Python is a free programming language, and its open standard libraries handle parsing sellers files, flagging duplicate leads, and calculating gaps. Metabase (AGPL license) is a self-hosted reporting tool that turns your queries into a monthly audit dashboard anyone on the team can read. Together they cost nothing to license. A full verification service that scores traffic in real time has no open source equivalent at parity, so treat that part of the work as manual sampling and reconciliation.
\nFraud follows whatever event you pay for, which means impressions, clicks, leads, and sales each attract their own fakes, and verification tools alone cannot see all of them. Affiliate and lead programs carry added exposure, both through manipulated attribution and through consumer payment disputes that return to your merchant account. The control that works across every pattern is the vendor audit, built from contract terms agreed before spending starts and run as logs, reconciliation, and path checks on a cadence set by vendor risk. When a finding appears, preserve the evidence, quantify what you can prove, send written notice under the contract, and settle the remedy in writing, with counsel involved when stakes or disagreement rise.
\n1. Fraud targets the event you pay for, so audit the payment trigger first and the dashboards second. 2. A verification report from a party you pay is an input to your audit, never the audit itself. 3. Audit rights, log-level data, and a defined dispute remedy must exist in the contract before the money moves. 4. Reconciliation works only when you set a tolerance in advance and watch for gaps that always run in one direction. 5. Evidence comes before accusation: freeze records, quantify the provable amount, and send written notice under the contract.
\nThis month, pick your three largest vendors, request the raw logs behind their last invoice, and reconcile billed volume against your own count with a tolerance you wrote down first.
\n1. Media Rating Council. Invalid Traffic Detection and Filtration Guidelines. Standards for identifying and filtering invalid traffic. 2. Interactive Advertising Bureau Tech Lab. ads.txt specification, current version. Authorized digital seller declarations by publishers. 3. Interactive Advertising Bureau Tech Lab. sellers.json specification, current version. Seller identity disclosure for supply chain verification. 4. Interactive Advertising Bureau Tech Lab. SupplyChain object specification, current version. Transparency of intermediaries in a bid request. 5. Federal Trade Commission. Dot Com Disclosures: Information about Online Advertising. Guidance on clear disclosure in digital advertising.
"}]},{"title":"Part IX: Strategy and the Road Ahead","items":[{"id":"chapter-56-budget-allocation-frameworks","title":"Chapter 56: Budget Allocation Frameworks","html":"By the end of this chapter, the reader will be able to: 1. Describe the four main allocation methods and the situation each one fits. 2. Explain how a response curve turns spend into diminishing returns. 3. Calculate and compare marginal return across channels using a simple worksheet. 4. Set floors, ceilings and a protected testing budget for a media plan. 5. Evaluate reallocation triggers and decide when a shift is justified. 6. Defend an allocation to a finance leader using plain evidence and stated limits.
\nA media buyer I respected once did the most sensible thing in the world. Search was our best channel, so he gave search more money. The next quarter it was still our best channel on paper, so he gave it more again. Then the results stopped growing, and nobody could say why. The dashboard still showed a healthy average return. The average was the problem.
\nThe assumption I want to challenge is this: the channel with the best average return deserves the next dollar. It does not. The average tells you what the whole pile of money did. The next dollar behaves differently, because every channel runs out of cheap customers. Search runs out when the people actually looking for you have all been reached. Social runs out when the audience that responds has seen your ad too often. The first 10,000 dollars in a channel and the last 10,000 dollars are not the same product.
\nA second belief does quiet damage too: that allocation is a spreadsheet exercise you finish in January. In practice it is a running argument between what you know, what you guess, and what the market is doing this month. The strongest plans I have seen were never the cleverest. They were the ones that admitted what was guessed, protected some money for finding out more, and had rules agreed in advance for when to move.
\nThis chapter gives you four methods, a way to think about response curves without advanced math, a worked example with illustrative figures, and a script for the finance meeting where someone asks why you want to move money. Treat every number as a teaching figure. Your own accounts will differ, and your own data should replace mine.
\nPercent of revenue is the oldest method. You set the total budget as a fixed share of sales, say nine percent of last year's revenue. It is simple, finance understands it, and it scales with the business. It also behaves backward in a downturn: sales fall, the budget shrinks, and the brand goes quiet exactly when it needs to be heard. Use it to set the size of the pot, not to divide it.
\nObjective and task works from the goal backward. If you need three thousand new customers and your current cost to win one is 80 dollars, the task costs 240,000 dollars. This method forces a link between money and a result, which is its great virtue. Its weakness is that the cost assumption hides inside one cell, and a stale cost makes the whole plan stale. It suits new launches and fixed targets.
\nMarginal return asks one question: what does the next block of money earn in each channel, and where is it highest? This is the method that actually divides the pot. It needs response curves, covered next, and it needs enough data to estimate them honestly. Where data is thin, you borrow rough slopes from experiments and adjust.
\nTest and scale is the practical cousin. You commit a small slice to controlled trials, find what works, and move money toward proven winners in steps. Think of a coach running a scrimmage before the real game. Nobody learns the second-string lineup in the championship.
\nMost sound plans use all four in sequence. Percent of revenue sizes the pot. Objective and task checks that the pot can reach the goal. Marginal return divides it. Test and scale keeps the division honest over time.
\nA response curve is a line showing how much result you get as spend rises. Picture filling a bathtub from a tap that slows down as the water rises. Early water arrives fast. Later water barely moves the level. That slowing is diminishing return, and every paid channel has it.
\nThe shape differs by channel. Search tends to rise steeply and flatten early, because the pool of people typing your category terms has a ceiling set by the market, not by your budget. Paid social often climbs more gently and keeps climbing longer, because the reachable audience is larger, though frequency eventually wears it down. Online video and retail media sit somewhere between, depending on inventory and how tightly you target.
\nYou do not need to fit a perfect curve to use the idea. Three data points will do. Suppose a channel returned 1.8 dollars of gross profit per dollar when you spent 100,000 dollars, 1.4 when you spent 160,000, and 1.0 when you spent 220,000. The average is still above break-even at the third point. The marginal dollars between 160,000 and 220,000, however, returned far less than the average suggests. You can compute the marginal figure directly: total gross profit at 220,000 dollars is 220,000, at 160,000 dollars it is 224,000. The extra 60,000 dollars bought less than nothing. The channel was already past its useful point.
\nThat small calculation is the heart of the chapter. Averages flatter. Differences between spend levels tell the truth.
\nBe careful about where the data comes from. Platform reports credit conversions generously and can overlap across channels. Controlled experiments, such as holding out a region or pausing a campaign for a set period, give cleaner slopes. When the two disagree, I trust the experiment and treat the platform number as a ceiling.
\nThe rule of the method fits on one line: move money from channels with low marginal return to channels with high marginal return until they roughly match, or until the return falls to your hurdle rate. The hurdle rate is the minimum return a dollar must earn to be worth spending at all.
\nFinance will care about this number more than any other, so set it deliberately. If your product earns a 40 percent gross margin, then each dollar of revenue is worth 40 cents of gross profit. A marginal dollar of ad spend that produces one dollar of revenue loses 60 cents. This is why raw revenue return figures mislead. Convert everything to gross profit before comparing, and set the hurdle at one dollar of gross profit per dollar spent, or higher if the business carries a payback requirement.
\nConsider a rule of thumb for steps. Do not move more than 10 to 15 percent of a channel's budget in one shift. Curves are estimates, and a big move on an estimate can overshoot into a different part of the curve. Move, wait one full buying cycle, re-measure, then move again.
\nThere is one more discipline. Marginal return is measured per time window. A video channel may look weak at 14 days and strong at 90, because it builds memory that search later collects. When a channel works upstream, judge it on assisted outcomes or on holdout tests, not on last-click credit alone.
\nPure marginal optimization has a flaw: it will happily starve a channel that is quietly doing important work. Floors and ceilings are the guardrails.
\nA floor is the minimum you will spend in a channel regardless of this month's numbers. Brand search is the classic case. Protecting your own name from competitors bidding on it costs little, and an empty seat there invites a rival in. A floor also protects learning. Cut a campaign to a trickle and you lose the data needed to judge it fairly.
\nA ceiling is the maximum you will spend, set by something real: limited search demand, available inventory, a frequency cap, or a customer service team that cannot handle more orders. Ceilings stop a hot channel from absorbing money it cannot use well.
\nThen comes the protected testing budget. I set aside 10 percent of the total, and I defend it harder than any single channel line. It is ring-fenced, which means no one raids it to cover a shortfall in a quarter that is going badly. Without this rule the testing money disappears first, every time, because tests do not pay back inside the quarter. A plan with no testing line is a plan that stops learning and starts guessing.
\nSpend the test money on questions that change allocation. Does a new video format beat the old one? Is a new channel worth a floor? Where is the real ceiling on search? Write the question and the decision it will drive before the money moves.
\nMeet Birchwater Outdoors, a fictional mid-sized retailer of camping gear. This is an illustrative composite, not a real company. Suppose its annual revenue is 12 million dollars and its digital budget is set at 720,000 dollars.
\nStep one: reserve the testing budget. Ten percent is 72,000 dollars, ring-fenced. That leaves 648,000 dollars to allocate.
\nStep two: set the starting plan and its limits.
\n| Channel | Start | Floor | Ceiling | Estimated marginal gross profit per dollar | |---|---|---|---|---| | Search | 300,000 | 200,000 | 320,000 | 0.9 | | Paid social | 200,000 | 120,000 | 260,000 | 1.3 | | Online video | 98,000 | 60,000 | 130,000 | 1.6 | | Retail media | 50,000 | 30,000 | 80,000 | 1.2 |
\nThe hurdle is 1.0. Search sits below it at the margin, even though its average return is the best in the plan. That is the veteran's trap from the opening. Video and social sit well above the hurdle and have room under their ceilings.
\nStep three: move in measured steps. Shift 40,000 dollars out of search, which leaves it at 260,000 and still above its floor. Send 25,000 dollars to video, bringing it to 123,000, under its ceiling. Send 15,000 dollars to social, bringing it to 215,000. The plan still totals 648,000 dollars.
\nStep four: re-measure after one buying cycle. Suppose search marginal return rises to 1.0 now that the weakest spend is gone, while video eases to 1.4 and social to 1.2. The gap has closed, so you stop. Had search still been below the hurdle, a second step would follow.
\nNotice what did not happen. Nobody abandoned search, and nobody bet the plan on a guess. The shift was modest, written down, and reversible.
\nMost bad reallocations happen in a meeting, under stress, after a bad week. Triggers written in advance take emotion out of it.
\nA good trigger names a metric, a threshold and a time window. For example: if a channel's marginal gross profit per dollar falls below the hurdle for two consecutive measurement periods, release 10 percent of its budget. If a channel exceeds its hurdle by a set margin for the same span and sits under its ceiling, it earns the released money. If a test beats its control with a result you pre-agreed to trust, it graduates into the main plan with a floor of its own.
\nSeparate signals from noise. A single weak week is noise. A seasonal dip that appears every year is a pattern, not a trigger. A new competitor entering your auctions, a platform policy change, or a tracking break can all distort numbers without changing true response. Check for these before moving a dollar.
\nAlso write triggers that run the other way. If a channel is cut and the total results collapse, restore it. Reversal rules protect you from your own conclusions.
\nKeep a log. Each reallocation records the date, the reason, the amount and what you expect to see. Three months later you can judge your own judgment. That habit alone improves allocation more than any model.
\nFinance people are not your opponents. They have seen marketing budgets justified by confident storytelling and then missed. Meet them with structure.
\nOpen with the size of the pot and the rule that set it. Then show the hurdle rate, in gross profit terms, and explain why you chose it. This single move often changes the tone of the meeting, because you are speaking their language about whether a dollar earns its keep.
\nShow the allocation as a table with floors and ceilings, and say plainly which numbers are measured and which are estimated. Admit the estimated ones. Finance trusts a stated uncertainty far more than false precision, and a range with a reason beats a single figure with none.
\nExplain the testing line as a research budget with named questions, not a slush fund. Give the decision each test will drive.
\nThen show the triggers. The point to land is that the plan has rules for its own correction. A leader who sees pre-agreed thresholds worries less about surprise requests later.
\nFinally, bring one scenario: what happens to the plan if revenue falls 15 percent. Show which channels hit floors first and what you would protect. This operational guidance does not replace your finance team's judgment, and any contract commitments with media vendors should be reviewed by counsel before you rely on flexibility you may not have.
\nResponse curve: A line showing how an outcome such as gross profit changes as spend in a channel rises, typically steep at first and flat later. Marginal return: The extra result earned by the next block of spend, as opposed to the average result of all spend so far. Hurdle rate: The minimum return a marginal dollar must earn, usually set in gross profit terms, to justify spending it. Floor: The minimum budget a channel keeps regardless of short-term results, set to protect defense and learning. Ceiling: The maximum useful budget for a channel, set by demand, inventory, frequency or operational capacity.
\n1. Feeding the channel with the best average return, which ignores that the next dollar earns less than the dollars before it. 2. Judging channels on revenue instead of gross profit, so a marginal dollar that loses money on margin looks like a winner. 3. Moving large shares of budget in one step on an estimated curve, then being unable to tell which change caused the result. 4. Letting the testing budget absorb shortfalls, so the plan stops learning within two quarters. 5. Cutting an upstream channel because last-click credit is low, then watching search and direct traffic quietly decline.
\n1. Size the total with percent of revenue, check it with objective and task, and divide it with marginal return. 2. Convert every channel result to gross profit before comparing, and set the hurdle rate in writing. 3. Ring-fence about 10 percent for testing and attach a named question and decision to every test. 4. Write floors, ceilings and reallocation triggers before the period starts, with metric, threshold and window. 5. Log every reallocation with its reason and expected result, then review the log each quarter.
\n1. When platform numbers and a holdout test disagree, treat the platform figure as an upper bound and plan from the test. 2. Run three spend levels in a channel when you can, because even a rough three-point curve beats an average. 3. Set the search ceiling by looking at impression share on your core terms, since a channel that already wins most eligible auctions has little room left. 4. Revisit floors at the start of each quarter, because a floor that made sense for a launch can become dead weight. 5. Bring finance one downside scenario with the plan, since the first question in the room is almost always about a revenue miss.
\nYou can build every part of this framework without paid software. LibreOffice Calc (MPL) handles the allocation worksheet, the marginal return table and simple scenario cases. DuckDB (MIT) runs fast queries over exported spend and outcome files on your own computer, which is useful for comparing spend levels across weeks. PyMC-Marketing (Apache 2.0) fits Bayesian marketing mix models and gives you response curves with stated uncertainty, though it takes some Python comfort and enough history to be reliable. Start with Calc and graduate to the modeling tool when your data justifies it.
\nBudget allocation works best as a sequence of methods, not a single formula: percent of revenue sizes the pot, objective and task checks it against the goal, marginal return divides it, and test and scale keeps the division honest. Response curves explain why the best average channel is rarely the best place for the next dollar, and a hurdle rate in gross profit terms gives every channel the same pass mark. Floors, ceilings and a ring-fenced testing budget guard against both neglect and overspending, while written triggers and a decision log keep reallocation calm and reversible. Defending the plan to finance comes down to stating what is measured, what is estimated, and how the plan corrects itself.
\n1. The next dollar, not the average dollar, decides where money should go. 2. Compare channels in gross profit per dollar against a written hurdle rate. 3. Move budget in steps of 10 to 15 percent and re-measure after a full buying cycle. 4. A ring-fenced testing budget with named questions keeps the plan learning. 5. Finance trusts plans that label estimates honestly and carry pre-agreed correction rules.
\nRebuild your channel report in gross profit per marginal dollar before your next planning meeting, set a hurdle rate of one, and move no more than 15 percent from the first channel that falls below it.
\n1. Brodersen, K.H., Gallusser, F., Koehler, J., Remy, N., Scott, S.L. Inferring causal impact using Bayesian structural time-series models. Annals of Applied Statistics. Published methodology for estimating the effect of an intervention without a randomized control. 2. PyMC-Marketing project documentation. Bayesian marketing mix modeling and media response curve estimation. Apache 2.0 licensed software. 3. DuckDB project documentation. In-process analytical SQL database for local analysis of exported campaign files. MIT licensed software. 4. LibreOffice Calc documentation. Spreadsheet functions, scenario tools and charting. MPL licensed software. 5. Media Rating Council. Minimum Standards for Media Rating Research. Reference for measurement quality expectations.
"},{"id":"chapter-57-channel-mix-by-business-stage","title":"Chapter 57: Channel Mix by Business Stage","html":"By the end of this chapter, the reader will be able to: 1. Identify the stage a business is actually in, using evidence rather than revenue size or founder optimism. 2. Describe the channel mix that fits startup, growth and mature businesses, and the archetype adjustments for local service, B2B and ecommerce. 3. Apply a stage gate to decide whether a new channel has earned a place in the plan. 4. Calculate whether a channel is still paying its way using a simple marginal return comparison. 5. Evaluate a current mix and name the channel that should be paused, cut or capped. 6. Design a 90 day adoption sequence that adds one channel at a time and records what was learned.
\nHere is a question. Your brand has a monthly budget and a list of eight channels that every vendor says you need. How many of those channels should be live right now?
\nMost managers answer with a number that is too high. They hear \"mix\" and picture a pie chart with eight or nine slices, each slice sized by some formula. A pie chart is a tidy picture. It is also a trap, because it implies that a healthy business buys everything at once and merely adjusts the proportions.
\nReal businesses do not work that way. A kitchen does not open with a 40 item menu. It opens with six dishes it can cook well, learns what sells, and adds the seventh dish only when the first six run smoothly. Advertising is the same. A mix is not a pie. It is a sequence, a story about which channel earned the right to be added next, and which one lost the right to stay.
\nI have watched small teams spread a modest budget across search, social, video, email and influencers in one quarter, then conclude that digital advertising does not work for them. It worked fine. It never had enough money in any one place to produce a readable result.
\nThere is a second wrong answer, common at the opposite end. A mature brand sees the same two channels carry it for years and assumes the mix is finished. Those channels are usually showing diminishing returns by then, and nobody has looked, because the reports still say the numbers are fine.
\nThis chapter gives you a way to answer the opening question honestly. You will learn what each stage tends to need, how three common business types bend the pattern, how to set a gate that a channel must pass before it is added, and how to recognize when a channel has finished its job. Every example here is a composite archetype with a fictional name, and every figure is illustrative. Your own accounts will supply the real numbers.
\nFounders tend to label their stage by age or revenue. Both are weak signals. A 12 year old company launching a new product line is a startup in that line.
\nA better test asks three questions. Do you know what a customer costs to win through at least one channel? Do you know roughly what that customer is worth over time? Can you repeat the result when you add money? If all three answers are no, you are at the startup stage, whatever the calendar says. A repeatable channel that scales with predictable returns means growth. Saturated main channels, with growth now coming from retention, price or a second market, mean maturity.
\nThis matters because the stage decides what the budget is for. At the startup stage, money buys learning. In growth, money buys scale. In maturity, money buys defense and efficiency. A manager who funds a startup like a mature brand will demand efficiency from a business that has not yet learned what works. A manager who funds a mature brand like a startup will keep paying for experiments that the account already answered.
\nAt this stage you have an offer, a small budget and a pile of untested guesses. The goal is not efficiency. The goal is a first repeatable path from stranger to buyer.
\nPick one primary channel. Choose it by asking where people already show intent for what you sell. If buyers type the problem into a search engine, start with search. If the product is visual, impulse driven and cheap enough to buy on a whim, a social platform may be the better starting point. Resist the urge to start with the channel that is most fun to talk about.
\nSuppose a fictional company, Birchline Pet Supply, sells a specialty dog harness. It has six thousand dollars a month to spend. The tempting plan is one thousand five hundred dollars each on search, social, video and influencers. The better plan puts the entire six thousand dollars on one channel for eight weeks, with a simple measurement setup and a landing page that has been tested by a few real customers.
\nWhy one channel? A result needs volume to be readable. Split the money four ways and a weak month could come from the channel, the creative, the page or the price, and you cannot tell which.
\nThe startup mix also needs three supports that are not channels: a tracked conversion event, a place to collect email addresses, and a first-party record of who bought. Chapter 41 treats the data side in depth. Avoid brand awareness campaigns with no visible response and any channel that needs a long window to judge. You can afford those later.
\nGrowth begins when one channel has become predictable. You know roughly what an order costs, and when you raise spend a little, results rise a little. That is a good problem to have, and the temptation now is to add channels too fast.
\nThe growth stage has two jobs. First, push the first channel until it starts to tire. Second, add a second channel that reaches people the first one cannot.
\nConsider Larkspur Home Services, a fictional lawn and yard care company that has done well with paid search. Search captures people who already want a lawn service. It cannot reach the homeowner who has not yet thought about it. Larkspur's second channel might be a local video placement or a neighborhood social campaign, aimed at the people who will search next season. That addition is not about efficiency. It is about feeding the demand search later collects.
\nWatch for three signs that the first channel is tiring: cost per result climbs at flat spend, impression share is already high for the terms that matter, and each extra one thousand dollars brings visibly less than the last. When two show up together, plan the second channel.
\nRetargeting and email also earn their place now, because there are finally enough visitors and buyers to talk to. Wait until the audience is large enough that you are not hammering the same few people. Add one channel, run it for a full cycle, and read it alone before adding the next.
\nA mature brand has strong awareness, a loyal base and channels that mostly perform. The risk is not failure to grow. It is slow erosion of efficiency that nobody notices because the totals look acceptable.
\nMature mixes lean on retention, defended brand terms in search, and steady reach through video and email. The larger lesson is measurement. When most sales come from people who already know you, last-click reports credit channels that merely stood near the checkout (Chapter 44 covers that). A mature brand should reserve a fixed slice of budget, say five to 10 percent, for channels it does not use today, and hold out a region now and then to see what happens when a channel goes quiet.
\nSuppose Harborlight Outdoor Goods, a fictional retailer, spends 100,000 dollars a month. It keeps 10,000 dollars for experiments and every quarter pauses one established channel in a few matched regions for four weeks. If sales in those regions barely move, the channel was taking credit rather than creating demand, and the money can move. If sales fall noticeably, the channel earned its budget, and the test paid for itself by showing it.
\nMature teams also tend to freeze the mix. Years later the audience has aged, a competitor has found a newer channel, and nobody has practice adding anything. A fixed testing slice keeps that skill alive.
\nStage tells you what the money is for. Business type tells you which channels fit. These composites are illustrative, not recipes.
\nLocal service. A plumber, a dentist or a roofer wins work from people with an urgent need in a small area. Demand is local and often immediate. The first channel is almost always paid search tied to location, supported by a well-kept business listing and a phone number that is tracked. Reviews matter as much as ads here, because a searcher compares three businesses in under a minute. A modest local social presence keeps the name familiar before the emergency, and email handles repeat customers. Three channels is usually plenty.
\nB2B. A company selling to other businesses faces long cycles, several decision makers and small audiences. Search captures buyers who already know the problem. A professional network platform reaches job titles with precision, at a high price per click. Content and email carry the long middle of the cycle. The mistake to avoid is judging a B2B channel by a 30 day window when the sale takes six months. Agree on interim signals, such as qualified meetings booked, and judge on those.
\nEcommerce. An online store has fast feedback and plentiful data, which is both a gift and a hazard. Search shopping placements and social catalogs often form the base, with email and text as the retention layer. Stores tend to chase the cheapest first order and neglect repeat purchase. One with a strong second order rate can afford to pay more for a new customer than a rival with a weak one, so size the mix to lifetime value.
\nIn all three, live channels stay few until the budget supports more. A business spending five thousand dollars a month rarely needs five channels.
\nA stage gate is a short list of conditions that must be true before you add the next channel. It protects you from your own enthusiasm and from persuasive sales calls. Write the gate down before the pitch arrives.
\nA workable gate has four tests: 1. The current channel is established. It has run for at least one full buying cycle with stable tracking, and its cost per result is known. 2. The current channel shows strain. Marginal returns are falling, or reach is largely used up. 3. The new channel has a defined job. You can say in one sentence what it does that the current channel cannot, such as reaching early-stage buyers or recovering abandoned carts. 4. There is a funded, readable test. The test budget is big enough to produce a result in the planned window, and the success rule is written down in advance.
\nIf any test fails, the answer is \"not yet.\" Write down why.
\nThe fourth test deserves emphasis. A test that cannot fail teaches nothing. Before spending, decide what result would make you expand, what would make you hold, and what would make you stop. Suppose Birchline Pet Supply tests a video campaign with a four thousand dollar budget over six weeks. The rule might be: if the campaign produces a cost per new customer within 25 percent of search, expand; if it is within 50 percent but shows rising branded search, hold for another cycle; otherwise stop. Those thresholds are illustrative. The point is that the thresholds exist before the first dollar moves.
\nThe rule can also allow for a channel that loses on direct cost but lifts another one. Awareness channels often do, which is why you track branded search and direct visits alongside the channel's own numbers.
\nAssume a business with one working channel and tracking it trusts. In the first 30 days, add nothing. Audit the existing channel, confirm conversion tracking, and record the cost per result and the marginal return from the last two budget changes. That is your baseline.
\nFrom day 31 to 60, run one new test under the written rules, funded at a level that can produce a reading. Leave the baseline channel alone so changes can be attributed cleanly. From day 61 to 90, read the result against the rules: expand, hold or stop. Log the date, budget, result and the reason. A year later that log tells you what you already tried.
\nThen repeat. A team following this sequence adds about four channels a year at most. That is the right speed.
\nAdding is easy. Stopping is hard, because every live channel has someone who set it up, defends it, and can produce a favorable chart.
\nUse a marginal return test. Take the last increment of budget and divide the extra results by the extra dollars. Suppose Larkspur Home Services spends four thousand dollars a month on a social campaign and gets 40 leads, a 100 dollar average cost. When it raises spend to five thousand dollars, leads rise to 44. The extra one thousand dollars bought four leads, a 250 dollar marginal cost. The average looks acceptable, but the margin is poor. The next dollar should go elsewhere, or back into the pocket.
\nAverage cost hides the problem because it blends early efficient dollars with late wasteful ones.
\nOther signals say stop or cap. The channel's results vanish in a holdout test. Frequency climbs as the audience tires. Or tracking breaks and you can no longer verify what you are buying. Stopping need not mean deleting: cap the channel at a maintenance level and keep the creative and settings, so a restart is cheap.
\nDecline is the other reason to shift. If the business is shrinking, tilt the mix toward email and loyalty offers, because winning new customers costs more than keeping current ones.
\nChannel mix: The set of advertising channels a business runs at one time and the share of budget each receives. Stage gate: A written list of conditions that must be met before a channel is added or expanded. Marginal cost: The extra cost of the extra results produced by the last increment of spend, as distinct from the overall average. Holdout test: A test that pauses a channel for a matched group of regions or audiences to see how outcomes change without it. Diminishing returns: The pattern in which each additional dollar in a channel buys fewer results than the one before.
\n1. Launching five or six channels in the first quarter with a small budget, which leaves every test too thin to read. 2. Copying a competitor's visible mix without knowing the competitor's stage, margins or customer value. 3. Judging a long-cycle B2B channel by a 30 day window, then cutting it before the sales cycle can show results. 4. Looking only at average cost per result and missing that the latest dollars are far less productive than the first ones. 5. Freezing a mature mix for years, with no testing slice, so the team forgets how to add a channel and the audience drifts away.
\n1. Place each product line at its stage using the three questions about cost to win, customer value and repeatability. 2. Write the stage gate before any vendor conversation, and apply it the same way every time. 3. Add one channel at a time, run it for a full buying cycle, and read it alone before adding another. 4. Fix the success rule in writing before spending, including what result means expand, hold or stop. 5. Keep a dated decision log so the reasoning behind every channel outlives staff changes.
\n1. If you can afford only one channel, put the whole budget on the one where buyers already show intent. 2. Track branded search and direct visits next to any awareness channel, since its value often appears there first. 3. Local service businesses should treat reviews and listings as part of the mix, not as extras. 4. Pause a tired channel rather than deleting it, and keep creative and settings ready for a cheap restart. 5. Reserve a fixed testing slice in mature accounts so new channels get tried before a competitor forces the issue.
\nTwo free tools cover most of what this chapter asks. Matomo (GPL) is a self-hosted web analytics platform that records visits, sources and conversions on your own server, giving you channel level results without sending visitor data to an outside vendor. DuckDB (MIT) is an embedded analytical database that runs on a laptop and queries spend exports and Matomo exports directly, so you can calculate marginal cost by channel and week in a short query. Together they support the stage gate tests, the holdout comparisons and the decision log. Neither buys media, and no open source product replaces a full demand-side platform, so treat them as measurement and record keeping, not as ad delivery.
\nA channel mix is better understood as a sequence than a pie chart. Startups should concentrate on one channel until they learn what a customer costs and what a customer is worth. Growing businesses add a second channel only when the first shows strain and the new one has a defined job. Mature brands defend what works, hold out regions to check whether channels create demand or merely collect credit, and keep a small testing slice alive. Local service, B2B and ecommerce businesses bend the pattern in different ways, but all of them benefit from a written stage gate, a funded readable test and a marginal return check. Channels should be stopped or capped when their last dollars stop paying, and a decision log keeps the lessons.
\n1. A mix is a sequence of channels that each earned their place, not a set of proportions to be filled in. 2. Stage is determined by evidence about cost to win, customer value and repeatability, not by age or revenue. 3. Concentrated spend produces readable results, while thin spend across many channels produces confusion. 4. A stage gate with a written success rule keeps new channels honest and protects the budget from sales pressure. 5. Marginal cost, not average cost, tells you when a channel should be capped or stopped.
\nBefore adding any channel this quarter, write its gate and its stop rule on one page, and compare the cost of the last one thousand dollars in your best channel to its average. If the margin is more than double the average, fund the next test from that channel.
\n1. Matomo documentation, self-hosted web analytics. Project documentation under the GPL license. 2. DuckDB documentation, embedded analytical database. Project documentation under the MIT license. 3. Binet, L. and Field, P. The Long and the Short of It: Balancing Short and Long-Term Marketing Strategies. Institute of Practitioners in Advertising. 4. Sharp, B. How Brands Grow: What Marketers Don't Know. Oxford University Press. 5. Brodersen, K. H., Gallusser, F., Koehler, J., Remy, N. and Scott, S. L. Inferring causal impact using Bayesian structural time-series models. Annals of Applied Statistics.
"},{"id":"chapter-58-seasonality-launches-and-crisis-pivots","title":"Chapter 58: Seasonality, Launches, and Crisis Pivots","html":"By the end of this chapter, the reader will be able to: 1. Describe how seasonal demand changes auction cost and why the direction of that change matters more than any single number. 2. Build a planning calendar that separates fixed dates, movable dates, and unknown dates. 3. Sequence a launch into pre-launch, launch, and settling phases, with a spending rule for each. 4. Execute a crisis pause and a staged resume using written triggers and named owners. 5. Evaluate brand safety exposure during a public crisis and set conservative placement defaults. 6. Design a contingency reserve with release rules that protect the core plan.
\nPlanning calendars are overrated. Every brand manager I have met owns a beautiful one, color coded by quarter and approved by six people, and most of them abandon it by the second week of the busy season. Reality does not read the calendar. A competitor drops prices early. A supplier slips. A news story lands on a Tuesday and your carefully timed message now sounds tone deaf.
\nSo why write a chapter on seasonality, launches, and crises at all?
\nBecause the calendar was never the point. The failure is not that people plan. The failure is that they plan the dates and forget to plan the decisions. A date tells you when something happens. A decision rule tells you what you will do when it does not happen on time, or happens twice as hard, or does not happen at all.
\nThink about a football coaching staff. Nobody believes the script for the first 15 plays survives contact with the defense. They write it anyway, then carry a second sheet with if-then rules for fourth down, a turnover, a key injury. The script is the calendar. The second sheet is where the money is saved.
\nThis chapter is about building both sheets for paid media. You will learn how seasonal demand pushes auction prices around, how to stage a launch so spend follows evidence instead of excitement, and how to pause and resume spend when something goes wrong in the world. None of this requires a forecast you cannot make. It requires rules written down before the pressure arrives, because decisions made in the middle of a crisis tend to be expensive, fast, and wrong.
\nStart with the plain mechanics. In an auction, price is set by competition for a limited supply of attention. When more advertisers want the same audience at the same moment, the cost to win rises. Seasonality is simply the calendar version of that pressure.
\nThree kinds of seasonal pressure show up in most accounts. Retail demand swells ahead of gift giving and back to school. Business demand follows budget cycles, with buying committees waking up at the start of a fiscal period and going quiet near the end of one. Event demand clusters around games, conferences, and cultural moments, and it can be intensely local.
\nHere is the part brand managers miss. Cost and demand do not peak together. Shoppers start researching before they buy, so search volume and intent often build ahead of the busiest buying days. Advertisers pile in as the peak approaches, so cost per click and cost per thousand impressions often climb faster than the conversion rate does. At the very top of the season you may pay the highest price for a user who was already going to buy somewhere. After the peak, prices fall quickly, sometimes while buyers are still around.
\nThis is why I tell teams to stop asking what the busy season will cost and start asking where the marginal dollar earns the most. Suppose an account spends 10,000 dollars a week in a quiet period and gets a certain return. Doubling spend at the peak will not double results, because the extra impressions are bought at higher prices and from less efficient pockets of the audience. Often the better move is to spend earlier, while intent is building and competitors are still asleep, and to hold the peak at a level your best audiences can absorb.
\nQualitative reading beats false precision. Look at your own history week by week and note which way cost, conversion rate, and volume lean. You do not need a prediction.
\nBuild the calendar in three layers. Mixing them is how calendars become decorations.
\nFixed dates are events you cannot move: public holidays, league schedules, regulatory deadlines, a trade show the whole industry attends. Plan backward from these. Creative lead time, approvals, and platform review all take longer than people admit, so work out the date assets must be live and subtract the review time, then the production time, then a buffer.
\nMovable dates are moments you control: your own promotions, product drops, email sends, price changes. These are your strongest tool. If a competitor owns a fixed peak with heavy bidding, you can often shift your own push a week earlier or later and pay less for the same audience. A movable date that happens to land on the most expensive week of the year deserves a hard look.
\nUnknown dates are the things you know will happen but cannot place: a supply shortage, a viral moment, a competitor price war, a news event touching your category. You cannot schedule these. You can budget for them, which is what the contingency reserve later in this chapter is for.
\nFor each fixed and movable date, record four things: the audience, the spend range, the metric that tells you it is working, and the person who can change the plan. That last field is the one people skip. When the date arrives and results are off, a calendar with no named owner produces a meeting, not a decision.
\nMark the quiet weeks on purpose. Quiet weeks are where testing belongs, because a bad result costs less when auctions are calm.
\nA launch is a planned event, which makes it the easiest of the three situations to control and the one most often wasted. The usual error is front loading: a huge budget on day one, a flat line afterward, and a post-mortem that blames the creative.
\nStage it instead.
\nPre-launch builds the audience you will later ask to buy. Spend here is small and aimed at collecting something you own: waitlist signups, email addresses, followers who opted in. Teaser creative, creator partnerships, and early access offers all belong here. The measurable output is not sales. It is the size and quality of the list, and how fast it grew. If the list is thin two weeks before launch, you have learned something valuable while it is still cheap to fix.
\nLaunch is the capture phase. Search, social, and any retail or marketplace placements should be live and tested before the date. Your pre-launch list gets first access, which means your first sales come from the cheapest audience you have. Paid prospecting then widens the net. A good rule is to set a launch ceiling and a launch floor in advance. The ceiling is the most you will spend per day regardless of enthusiasm. The floor is the spend below which the data will be too thin to read.
\nSettling begins when the novelty fades, usually within a few weeks. Spend drops toward a steady level, and the job shifts to retention: email sequences, onboarding messages, cross-sell, and review collection. This is where the margin is. A launch that wins customers and then ignores them has paid the full acquisition price for a single purchase.
\nConsider an illustrative brand, Birchwater Outfitters, launching a camp stove. The team collects three thousand waitlist emails over three weeks, spends modestly on teaser video, and opens sales to the list 48 hours before the public. Paid search goes live on public day. Because the list absorbed the early demand, the paid campaigns are not asked to carry the whole launch, and the team can see within a week which search terms deserve more money.
\nWhen something serious happens in the world, you have two questions. Should your ads stop? And when can they start again?
\nAnswer the first with a written trigger, not a feeling. A trigger is a specific event class that automatically puts a pause on the table: a national tragedy, a safety incident connected to your product, a public scandal involving your brand or a named spokesperson, an outage that breaks your checkout. Decide ahead of time who may call the pause. One person, with one backup. Group chats are where crises go to become slow.
\nPausing is not a single switch. Sort activity by risk: 1. Anything adjacent to the event, such as news placements, trending content, or topical creative, stops first. 2. Anything that cannot deliver on its promise, such as ads sending traffic to a broken checkout or out of stock products, stops next. 3. Evergreen search for people actively looking for you usually continues, because turning off your own brand terms hands the shopper to a competitor. 4. Scheduled social posts and email sends get reviewed by a human before the next one goes out.
\nResuming is harder than pausing, so stage it. Bring back the lowest risk activity first, watch for a defined stretch of time, then add the next layer. Set resume criteria in advance: the story has left the top of the news cycle, comments and messages are no longer dominated by the event, the problem you were pausing for is fixed, and the person who called the pause signs off. Keep a time stamped log of every change, so anyone who asks what you did and when gets an answer in a minute.
\nDo not confuse a pause with silence. If the crisis is yours, a short, plain statement on owned channels, written with legal review, beats paid reach every time. Operational guidance only here; counsel decides what you say.
\nBrand safety means keeping your ads away from content that would embarrass you or harm the audience. In a crisis it becomes live, not theoretical. Placement lists that were fine last week can be dangerous this week, because the same publishers and video channels may suddenly carry graphic or distressing coverage.
\nSet a conservative default. If your category touches children, health, money, or grief, assume the stricter stance and widen it only with evidence. During a major event, tighten exclusions by keyword and topic, remove open exchange inventory you cannot inspect, and move spend to placements you have reviewed by hand. Yes, reach drops. That is the price of not appearing beside a tragedy.
\nTwo checks are worth building into the routine. First, review placement reports daily during an active crisis, not weekly. Second, check your own creative against the news of the day before every scheduled flight. A cheerful message about a flash sale that lands next to coverage of a disaster costs far more in goodwill than it earns in orders.
\nKeywords are blunt tools. Blocking one word can remove safe inventory, and missing a variant leaves a gap, so pair lists with manual review. Where creators or influencers carry your message, call them before they post, not after. Give them a written hold instruction and a named contact, and remember that anything involving minors deserves extra caution.
\nA contingency reserve is money set aside at the start of a period for events you cannot schedule. Without one, every surprise becomes a raid on the core plan, and the raid usually happens in a panic.
\nSize the reserve by exposure, not by custom. A brand with a single peak season and thin margins may hold back a larger slice than a brand with steady demand. Round, illustrative numbers help: suppose a quarterly plan is 300,000 dollars. A reserve somewhere around five to ten percent is a reasonable starting range, to be adjusted after you see how often you actually draw on it.
\nWrite the release rules before you need them. 1. Define what qualifies: a defended opportunity, a competitor move, a recovery from a pause, or a time-limited event you could not have scheduled. 2. Require a one-line justification and the expected evidence of success. 3. Cap any single release at a share of the reserve, so one hasty decision cannot empty it. 4. Return unspent reserve to the core plan on a fixed date, so the money does not sit idle.
\nResist the opposite temptation, raiding the reserve to chase a good week. If you never touch the money, the reserve is too large. If you drain it every month, you underbudgeted the plan.
\nYou do not need forecasting software to learn from last year. You need your own history organized so you can compare like with like.
\nExport daily spend, clicks, impressions, and conversions from each platform and store them in a single table. Add a column for known events: your promotions, competitor actions you noticed, holidays, and any crisis pauses. Then ask simple questions. In which weeks did cost per click rise before conversions did? When did conversion rate drop even though traffic held? How long did it take for results to recover after a pause?
\nFor a first pass, a database that runs on a laptop is plenty, and plain summaries by week answer most of these questions. If you want to separate a repeating seasonal pattern from a trend and from random noise, a statistical decomposition does that cleanly. It splits the series into a pattern that repeats, a direction that drifts, and what is left over. If the leftover is large, you have an unexplained event worth investigating.
\nBe honest about limits. One or two seasons of data is a small sample, and a pattern seen twice may be coincidence or a forgotten platform change. Test seasonal findings in quiet periods before betting a peak on them. Where you ran a launch or a pause, compare against a similar region or period that did not change, and say plainly that the comparison is imperfect.
\nFlighting: Scheduling ads to run in bursts with gaps in between, instead of at a steady level, so spend matches demand or events. Dayparting: Limiting or raising bids by hour or day of week to match when your audience is most likely to act. Contingency reserve: A set-aside portion of budget held for unscheduled events and released only under written rules. Brand safety: The set of controls that keep ads out of content or placements that could embarrass the brand or harm viewers. Holdout: A comparison group or region where ads are withheld or left unchanged, used to estimate what would have happened anyway.
\n1. Planning dates but not decisions, so the calendar shows when a peak arrives and says nothing about what to do when results miss expectations. 2. Spending the biggest share at the most expensive moment of the season, when earlier weeks with building intent and calmer auctions would have cost less per result. 3. Putting nearly the whole launch budget on day one, which leaves nothing to react to what the first week of data reveals. 4. Switching off every campaign in a crisis, including brand search and working customer service paths, then restarting everything at once with no staging and no log. 5. Treating brand safety as a setting changed once at account setup, rather than a daily check during live events.
\n1. Keep fixed, movable, and unknown dates in separate layers of the calendar, and assign a named owner to every entry. 2. Move your own promotions away from the most competitive weeks whenever the business can allow it. 3. Stage every launch in three phases with a written daily ceiling and a floor for readable data. 4. Write pause triggers, resume criteria, and the name of the person who can call a pause before any crisis occurs. 5. Hold a contingency reserve with written release rules, and review how it was used at the end of each period.
\n1. Reserve the quiet weeks for testing, because experiments cost less and fail more gently when auctions are calm. 2. Keep brand search running during most pauses, unless the pause is caused by a problem with the brand itself, so shoppers do not land on a competitor. 3. Give creators a written hold instruction and one contact person before a crisis, not during it. 4. Time stamp every pause, change, and resume in a shared log that leadership and partners can read in a minute. 5. Compare any launch or pause against a similar region or period left unchanged, and state the comparison's weaknesses out loud.
\nTwo free tools cover most of what a small team needs here. DuckDB (MIT license) is an analytical database that runs on a laptop and lets you query exported platform reports with ordinary SQL, which makes weekly seasonal comparisons quick. statsmodels (BSD license) is a Python library for statistical analysis, including decomposition of a time series into its repeating pattern, trend, and leftover noise. Together they let you study your own history without buying anything. Neither replaces judgment, and a small sample of seasons still deserves skepticism.
\nSeasonality changes auction cost because competition for attention changes, and cost and conversion rarely peak together, so the useful question is where the marginal dollar earns the most. A good calendar separates fixed, movable, and unknown dates, and it names an owner for every entry. Launches work best when staged into pre-launch, launch, and settling phases with spending limits set in advance. Crisis response depends on written triggers, a staged pause, a staged resume, and a log of every change. Brand safety tightens during live events, and a contingency reserve with release rules absorbs surprises without raiding the core plan.
\n1. Seasonal pressure on cost often builds before and peaks after the point where conversions are strongest. 2. Movable dates are the strongest lever, because you can step away from the most expensive weeks. 3. A staged launch lets spend follow evidence, with the pre-launch list supplying the cheapest early sales. 4. A crisis pause should be sorted by risk, resumed in stages, and recorded in a time stamped log. 5. A contingency reserve only works when release rules are written before the pressure arrives.
\nWrite the second sheet before your next busy season: one page naming who can pause spend, which campaigns stop first, what must be true to resume, and how much reserve you can release without asking permission.
\n1. Interactive Advertising Bureau Tech Lab. OpenRTB specification, current version. Primary specification for how programmatic auctions determine price. 2. Hyndman, R.J. and Athanasopoulos, G. Forecasting: Principles and Practice. Open online textbook covering seasonal decomposition of time series. 3. Seabold, S. and Perktold, J. statsmodels: Econometric and statistical modeling with Python. Documentation for the seasonal decomposition tools. 4. DuckDB Foundation. DuckDB documentation. Reference for querying local data files with SQL. 5. Brodersen, K.H., Gallusser, F., Koehler, J., Remy, N., Scott, S.L. Inferring causal impact using Bayesian structural time-series models. Annals of Applied Statistics. Methodology for comparing outcomes against a holdout baseline.
"},{"id":"chapter-59-ai-assisted-creative-and-targeting-with-open-models","title":"Chapter 59: AI-Assisted Creative and Targeting with Open Models","html":"By the end of this chapter, the reader will be able to: 1. Identify the advertising tasks where an open-weight language model or speech model saves real labor. 2. Explain why a model described as open can still fail this book's license rule. 3. Configure a local hosting plan that keeps brand files and customer data on hardware you control. 4. Design a human review gate with named reviewers, a checklist and a stop rule. 5. Evaluate disclosure and intellectual property exposure before an AI-assisted asset goes live. 6. Plan around platform-side AI automation that you cannot inspect.
\nPicture a regional furniture brand, fictional and illustrative, called Harbor Oak. The marketing manager is sensible. She wants more ad variants than her two-person team can write, so she downloads a free language model, asks it for 200 headlines, and has an assistant paste the best 40 into a social campaign. Nobody checks the license. Nobody reads every line.
\nThree weeks later a customer posts a screenshot. One headline promised \"lifetime free repairs,\" a policy Harbor Oak has never offered. The model invented it, it sounded plausible, and a tired reviewer skimmed past it. Then the legal contact notices the download page for the model carried a use restriction that excluded advertising for some commercial categories. The decision to use AI was not the mistake. The missing process around it was.
\nHere is the assumption worth challenging: that the hard part of AI-assisted advertising is getting the model to produce good output. It is not. Producing output is the cheap step. The expensive steps are verifying that the model is yours to use, confirming that what it wrote is true, deciding who is accountable for each asset, and knowing what happens to your data when you feed it a customer list.
\nThis chapter treats open models the way a good equipment manager treats a new piece of gear. Useful, yes. Also something that can injure the team if it is handled carelessly. I will show where these models earn their place, how to run them on your own machine, how to read a license the way a lawyer's checklist would, and how to build review gates that catch the Harbor Oak problem before a customer does. I will also be honest about the part you cannot control: the automated AI that ad platforms run inside their own systems. Nothing here promises performance gains. Treat any such promise with suspicion.
\nStart with a plain rule. Use a model for work where a human can check the result faster than doing the work from scratch. That rule sorts the field quickly.
\nDrafting variants is the obvious win. If you already have an approved message, a model can restate it in 15 different tones, lengths and hooks. You supply the claim and the facts. The model supplies phrasing. Your reviewer then picks from a menu instead of staring at a blank page, and the facts never leave your control because you fed them in.
\nTranscription is the second win. Podcast sponsorship reads, creator videos, customer call recordings and radio spots all become searchable text. A brand manager who can search an hour of audio for every mention of a competitor has a research tool that used to cost an agency intern a week.
\nTagging and classification are the quiet workhorses. Suppose you have 5,000 customer comments across your social posts. A classification pass can sort them into themes such as shipping complaints, product praise and pricing questions. The sort does not need to be perfect, because you are looking for patterns to investigate, not issuing verdicts on individual people.
\nNow the places to be careful. Do not ask a model to invent facts about your product, to write claims that need substantiation, or to produce anything resembling a customer testimonial. Do not ask it to guess sensitive traits about individuals. A model that writes fluent text is not a model that knows your return policy. Fluency and accuracy are separate properties, and the gap between them is where Harbor Oak got hurt.
\nA good test: could a wrong answer from this task reach a customer without a human reading it? If yes, the task needs a gate. If the answer is \"no, a person reads every output,\" the task is a candidate for automation.
\nThe word open gets used loosely. Some models publish their weights, meaning the trained numbers that make the model work, and call themselves open, yet attach terms that restrict who may use them, for what, or at what company size. Some forbid use in certain industries. Some require you to display a brand notice. Some reserve the right to change terms later.
\nThis book's rule is strict and simple. Only permissive licenses pass: Apache 2.0 or MIT. If the terms include a field-of-use restriction, a company-size threshold, a usage-reporting duty or a custom acceptable-use policy, the model fails, however popular it is and however good its demos look. A community favorite that fails the rule is still a failure.
\nVerification is a short, repeatable routine, and I want it done by a person with a checklist, not by a hunch.
\nOne more wrinkle. A permissive license covers the weights and code. It does not tell you what was in the training data. That is a separate question, and I will return to it under intellectual property. Nothing in this chapter is legal advice. It is operational guidance, and your counsel decides what your company can accept.
\nRunning a model on your own hardware solves a problem that shows up the moment you paste a customer list into a web chat window: where does that data go? With a locally hosted model, prompts and outputs stay on your machine. No third party logs them, trains on them, or changes terms on you next quarter.
\nThe host software does the unglamorous work of downloading a model, loading it into memory and answering requests. Ollama, released under the MIT license, is the host this chapter uses. You install it on a workstation or a small server, pull a model that passed your license check, and talk to it through a local address. Your team can reach it from a browser interface or a simple script.
\nBe realistic about hardware. Small models run on an ordinary modern laptop, slowly. Larger models want a computer with a strong graphics card and plenty of memory. Match the model to the task. A short headline rewrite does not need the biggest model you can find, and a smaller one that runs quickly will get used more often than a larger one that makes people wait.
\nA small hosting checklist:
\nThese three tasks deserve their own attention because they are where the safest returns sit.
\nFor transcription, Whisper, released under the MIT license, converts speech to text on your own machine. Consider an illustrative sponsorship team at a fictional trail-running brand called Thistlewick. They sponsor eight podcasts. Each delivers sponsor reads in a different style, and the team must confirm every read included the required legal line. Transcribing each episode and searching for that line takes minutes. Before, someone listened to hours of audio. The model does not judge compliance. It produces text, and a person reads the matches and the gaps.
\nFor tagging and classification, spaCy, released under the MIT license, is a language processing library that can pull out names, places and organizations, and can sort text into categories you define. It is not a chatbot. It is closer to a very fast, very literal filing clerk. You can use it to sort comments into themes, flag posts that mention competitors, or screen creator captions for words on your avoid list.
\nA caution on audience work. Classification of text content is one thing. Classification of people is another. Do not use models to infer health, finances, religion or similar traits about individuals, or build targeting from such guesses. Keep audience work to what people did with your brand and the context where your ad appears. Where children may be involved, set the most conservative default and let counsel set any exceptions.
\nA useful discipline is to measure the sorter before trusting it. Pull 100 comments, label them by hand, run the classifier, and compare. If it disagrees with your people often, fix the categories or drop the task. No performance claim in this book should substitute for your own spot check.
\nA review gate is a checkpoint where a named person must approve an asset before it moves on. It sounds bureaucratic. Think instead of a referee. The game can run fast because someone holds the whistle.
\nHarbor Oak had no gate. A good gate has four parts.
\nFirst, a source rule. Every factual claim in an AI-assisted asset must trace to an approved document: a price list, a policy, a product spec. If the reviewer cannot find the source, the line comes out.
\nSecond, a checklist. Keep it short enough that people actually use it. Is every claim sourced? Does the tone match the voice guide? Does it mention any person, place or event that needs verification? Does it resemble another company's slogan? Does the asset need a disclosure?
\nThird, a named owner. \"The team reviewed it\" means nobody did. Put one person's name on each asset, with a backup for vacations.
\nFourth, a stop rule. Any reviewer, at any level, can pause a campaign for a suspected error without asking permission. Decide this in advance, because arguments about authority are worst when something is already live.
\nRules on how AI-generated content must be labeled are changing, and they vary by jurisdiction and by platform. Operational guidance only: ask counsel what applies to you. Some ad platforms ask advertisers to label synthetic or altered media, especially when it depicts realistic people or events. Endorsement rules in the United States still require that material connections be disclosed, and a synthetic person delivering a testimonial raises problems that a real person with a real experience does not.
\nMy conservative default is simple. Never present a model-generated person as a real customer. Never fabricate a testimonial, a review, a quote or an expert. Never imitate the voice or likeness of a real individual without written permission from that person. When in doubt, label the asset as AI-assisted and let counsel decide whether the label is required.
\nOn intellectual property, three questions deserve attention. Who owns the output? In many jurisdictions, the answer for purely machine-generated material is unsettled, which can weaken your ability to protect it. What went into the training data? A permissive model license does not promise the training material was cleared for every use. And did the model reproduce something it learned too well? Models occasionally echo phrases from existing work. Run key lines through a plain web search before they ship, and keep a human author involved in anything you hope to protect, such as a signature campaign line.
\nFor children's audiences, sensitive product categories and anything touching health, finance or safety, keep AI drafting to internal brainstorming and have a human write the final copy. Creators who work with you deserve the same courtesy. If you ask a creator to deliver a read, do not quietly swap their voice for a synthetic one.
\nEvery large ad platform now builds automation into its buying tools. These systems may pick audiences, mix creative, set bids and place ads across inventory with little input from you. You cannot swap in your own license-checked model, and you usually cannot see why the system made a given choice.
\nBe honest with yourself about what that means. These tools are not open source, and this book does not recommend them as software. Where you must use a platform to reach an audience, you are buying media there, not adopting a tool. The question is how to buy it with your eyes open.
\nFour habits help. Hold back a control. Keep a portion of spend, or a geographic region, running with the automation limited or off, so you have a baseline to compare against. Guard your inputs. The automation learns from what you give it, so the creative, the conversion events and the audience signals you provide are your steering wheel. Record changes. When the platform flips a default setting, write down the date, so you can explain a shift in results later. And set hard limits: budget caps, brand safety exclusions and placement exclusions, checked weekly.
\nSuppose a fictional meal-kit company called Fennel Crate turns on a fully automated campaign type and sees spend drift toward one placement it never chose. Without a control and a change log, the team cannot tell whether the shift helped or hurt. With them, the team can compare, decide, and push back. Where the platform's reporting is the only evidence of success, remember that trackable is not the same as measurable. Treat platform claims as leads to test.
\nThe deeper strategic risk of leaning on systems you cannot inspect belongs to Chapter 60, and I will leave it there.
\nDo not roll this out across a whole department. Run a four-week pilot on one low-risk task, such as variant drafting for one product line. Verify the license and write the voice guide in week one, run the gate and log reviewer changes in week two, test human-written against AI-assisted copy in week three, and decide in week four. If reviewers spend as long fixing drafts as they would have spent writing them, the pilot says stop, and that is a good result.
\nOpen-weight model: A model whose trained parameters are published for download, though the license may still restrict how you use it. Permissive license: A license such as Apache 2.0 or MIT that allows commercial use with minimal conditions. Hallucination: A confident, fluent model output that is false or unsupported by any source. Classification: Sorting items such as comments or captions into categories you define. Human-in-the-loop: A process in which a named person must approve model output before it is used.
\n1. Treating a free download as a licensed one, and skipping the license read because the model is popular. 2. Letting a model write factual claims, prices or guarantees that nobody checks against a source document. 3. Pasting customer lists or private files into hosted web chat tools when a local host would keep them in-house. 4. Reviewing a sample of finished assets instead of every asset that actually goes live. 5. Using a model to guess sensitive traits about people and building targeting from the guesses.
\n1. Keep a license register listing model name, version, source, license and date checked, and recheck on every upgrade. 2. Give every AI-assisted asset a named owner, a short checklist and a stop rule. 3. Feed the model approved claims and a voice guide, so it rewrites facts you supplied rather than inventing its own. 4. Hold a control group or control region whenever platform automation touches your spend. 5. Pilot one low-risk task for four weeks and record reviewer changes before expanding.
\n1. Use a smaller model that responds in seconds before reaching for a larger one that people avoid. 2. Run key headlines through a plain web search to catch accidental echoes of someone else's line. 3. Caption every video asset, and have a person correct brand names and product terms in the transcript. 4. Keep a short change log of platform default shifts, with dates, so a results swing can be explained later. 5. Ask counsel once, early, which labeling rules apply to your categories, and write the answer into the checklist.
\nFour open tools cover the tasks in this chapter. Ollama (MIT license) hosts models on your own machine and answers requests locally. Whisper (MIT license) turns audio into text for transcripts and captions. spaCy (MIT license) extracts names and sorts text into categories you define. For language models themselves, use only those released under Apache 2.0 or MIT licenses, and verify the license file yourself before download. Models with custom terms, usage restrictions or size thresholds do not pass this book's rule. No open option offers the managed convenience of platform automation, and that gap is a tradeoff to accept knowingly.
\nOpen models earn their place in advertising on tasks a human can check quickly: drafting variants from approved claims, transcribing audio, and tagging text. They fail when they are asked to invent facts, impersonate real people or guess sensitive traits. A model qualifies for this book only under a permissive license, verified from the official source and recorded in a register. Hosting locally keeps customer and brand data in-house, at the cost of hardware and staff time. A review gate with a source rule, checklist, named owner and stop rule is what separates useful assistance from the Harbor Oak mistake. Disclosure and intellectual property questions go to counsel, with conservative defaults for children and sensitive categories. Platform automation stays a black box, so hold controls, guard inputs and log changes.
\n1. A model called open is not automatically free to use; only Apache 2.0 or MIT licenses pass. 2. The cheap step is generating output, and the costly steps are verification and accountability. 3. Local hosting keeps prompts, files and customer data on hardware you control. 4. Every live asset needs a named owner, a source-checked claim list and a stop rule. 5. Platform-side AI cannot be inspected, so protect your control group and your change log.
\nBefore any AI-assisted asset goes live, name one owner, trace every claim to a source document, and record the model's license in a register. Run a four-week pilot on a single low-risk task before you expand.
\n1. Federal Trade Commission. Guides Concerning the Use of Endorsements and Testimonials in Advertising, 16 CFR Part 255. 2. Open Source Initiative. The Open Source Definition, and the text of the MIT License and Apache License 2.0. 3. Apache Software Foundation. Apache License, Version 2.0, official license text. 4. Ollama, Whisper and spaCy project repositories: official license files and documentation.
"},{"id":"chapter-60-emerging-formats-and-platform-risk","title":"Chapter 60: Emerging Formats and Platform Risk","html":"By the end of this chapter, the reader will be able to: 1. Explain why a healthy dashboard can coexist with an unacceptable business risk. 2. Identify the five ways a platform can change the economics of a channel without warning. 3. Calculate spend, revenue and audience concentration for a real media plan. 4. Build a platform risk register with owners, triggers and pre-agreed responses. 5. Design a scenario plan for an account suspension or an answer-layer shift. 6. Judge an emerging format against a capped test budget and a written kill rule.
\nPicture two documents on the same table. The first is your channel dashboard. It shows one platform delivering 70 percent of new customers at a steady cost, week after week, with a tidy trend line. The second is what a finance partner would write if you asked for a risk memo on the same business. It would say that 70 percent of acquisition runs through a counterparty that can change price, targeting, measurement or your access with no vote from you and no notice period.
\nBoth documents are accurate. Only one of them gets presented in the quarterly review.
\nMost brand managers treat platform concentration as a sign of efficiency. You found the channel that works, so you fed it. That instinct is not foolish. Efficiency is the whole job. But a channel that works under today's rules is a lease, not a deed, and the landlord rewrites the lease on a schedule you do not control. I have watched operators in my own portfolio feel brilliant for 18 months and then spend a bad quarter rebuilding from nothing because one account review went the wrong way.
\nThe honest framing is that the platforms are not villains. They are businesses optimizing their own results, and sometimes your results and theirs point in different directions. A policy tightening protects their regulatory position. A pricing shift protects their margin. An answer box that resolves a question without a click protects the user experience they sell. None of it is personal. All of it lands on your cost per customer.
\nThis chapter is the canonical statement of a thesis the rest of the book leans on: dependence on a platform is the central business risk in digital advertising, and the hedge is a balanced channel mix plus audiences you own. It also covers the new formats everyone is pitching you, and how to treat them with calm skepticism and a small, capped budget.
\nA dashboard reports performance. A balance sheet reports exposure. They answer different questions, and brand managers are usually trained only on the first.
\nSuppose a regional furniture retailer, which I will call Juniper Lane (an illustrative composite), spends 60 percent of its media budget on one social platform. Return on ad spend looks excellent. A finance partner reading the same plan would ask a different set of questions. What happens to revenue if the account is disabled for nine days? What is the contractual notice if the platform changes how conversions are counted? Which customers could we still reach tomorrow if we lost access tonight?
\nIf the honest answer to the third question is \"almost none,\" then the retailer does not own a customer acquisition engine. It rents one. The rent is currently low, which is why nobody is worried.
\nFinance people think this way because they have seen supplier concentration before. A manufacturer with one source for a critical part qualifies a second. Nobody calls that waste. Media is the same, except the supplier can also set the auction price and disable your login.
\nYou do not need to abandon your best channel. You need to price the exposure and decide, in writing, how much you will carry.
\nPlatform risk shows up in more than one form, and each form needs a different response. Naming them keeps the register honest.
\nThe first is account action. A suspension, a payment hold or a policy flag can stop delivery completely, sometimes for reasons that are wrong and sometimes for reasons that are fair but hard to appeal. The damage is the pause itself. Many appeals take days, and in a retail calendar a week is not a rounding error.
\nThe second is policy change. Targeting options narrow, a category becomes restricted, an ad format is retired, or a measurement window shortens. Your campaigns still run, but the machine that made them efficient is different.
\nThe third is price change. Auctions get more crowded, floors rise, or a new placement absorbs impressions that used to be cheap. Cost per acquisition drifts upward without any dramatic announcement.
\nThe fourth is measurement change. The platform alters how it counts, attributes or reports conversions. Your trend line breaks, and you cannot tell whether performance moved or the ruler did.
\nThe fifth is competitive change. The platform launches a product that competes with yours, or promotes its own offering in the same placements where you buy. Retail marketplaces do this often, and so do app stores.
\nOnly the first is dramatic. The other four arrive quietly, and no one calls a meeting about them.
\nSearch used to be a list of links. Increasingly it is an answer, written by the engine, at the top of the page. For advertisers this creates a strange new position: the user's question may be resolved before any ad or organic listing is clicked.
\nThis area changes quickly, and anyone offering precise predictions is guessing. What can be said is structural. When an answer layer summarizes a topic, fewer visits go to the sources it summarizes, and the sources that get cited gain visibility out of proportion to their ranking. Paid placements around answers may shift in format and price. Organic traffic, which feels free and stable, becomes exposed to the same rule changes as a paid channel.
\nPlan for it as a scenario, not a prophecy. If search clicks fell by a third over two quarters, which parts of the business would feel it first? Branded search often holds up better than category search. Comparisons, local answers and transactions tend to keep sending people to sites. Simple factual questions are the most exposed.
\nThe defense is the same as everywhere in this chapter. Be findable by name, keep a direct line to people who already bought, and keep your owned pages clear and worth citing.
\nTwo formats deserve a clear-eyed look because the money is moving toward them and the pitches are loud.
\nRetail media means advertising inside a retailer's own properties, and increasingly on other sites using the retailer's shopper data. The attraction is real. The retailer knows who bought what, so targeting can be tied to purchase behavior. The strings are also real. You are buying inside a closed environment where the seller reports its own results, sets its own attribution window and often controls the placement rules. Measurement across different retailers rarely lines up. If a single retail network becomes your main route to shoppers, you have swapped one concentration for another. Hold it to the same standard as any channel: ask for the reporting method in writing and run a holdout test where you pause spend in one comparable region and compare.
\nCreator media means paying individuals with an audience to speak to it. At its best it carries trust that a banner never will. At its worst it is a rented relationship with a person whose own platform can vanish. Here the family-first standard applies firmly. Default to creators whose content you would be comfortable having a child see, require clear disclosure of paid relationships, and keep contracts specific about usage rights, approval and what happens if the creator's account is suspended. This is operational guidance only; counsel should review the actual agreements.
\nBoth can earn a place in a balanced plan. Neither fixes dependence. Often they add a second dependence with a more charming sales team.
\nYou cannot manage a risk you have never put a number on. Three ratios do most of the work, and each takes an afternoon to calculate.
\nSpend concentration is the share of your media budget going to your largest platform. Revenue concentration is the share of attributed revenue that platform reports, and you should treat that figure with the usual caution, since platforms tend to claim generously. Audience concentration is the share of your reachable customers whom you could contact only through that platform.
\nHere is an illustrative worked example for a fictional outdoor equipment brand, Marrow Creek Supply, with a monthly plan of 100,000 dollars.
\n| Channel | Monthly spend | Share of spend | |---|---|---| | Largest social platform | 55,000 dollars | 55 percent | | Search | 25,000 dollars | 25 percent | | Retail media | 10,000 dollars | 10 percent | | Creators and other tests | 10,000 dollars | 10 percent |
\nSpend concentration is 55 percent. That alone is not alarming. Now add the third ratio. Suppose the brand has 90,000 past customers, and 60,000 of them are reachable by email, which the brand owns. Then 30,000 people, one third of the base, can be reached only through platform audiences. If the platform account went dark, the brand would still reach two thirds of its customers directly.
\nSet a threshold in advance, for instance a review whenever any single platform passes 50 percent of spend or when reachable owned audience falls under half of the customer base. The exact numbers are your call. The point is that crossing the line triggers a conversation, not a shrug.
\nA scenario plan is not a forecast. It is a rehearsal. You pick four plausible bad mornings and write down, calmly, what you would do before you are panicking.
\nThe first bad morning is a suspension of your main ad account. Your plan names who files the appeal, who holds the backup payment method and verified business documents, and what budget moves to search and email in the first 48 hours. It also names which campaigns pause and which keep running through a second account that was set up legitimately and in advance.
\nThe second is a measurement break, where reported conversions drop 30 percent overnight with no change in sales. Your plan says who checks your own analytics and order system first, so you know whether demand changed or only the report did.
\nThe third is a price spike that lifts your cost per acquisition by a quarter. Your plan states the maximum cost you will pay, which channels absorb the displaced budget, and what the finance partner approves without another meeting.
\nThe fourth is an answer-layer shift that cuts search clicks. Your plan prioritizes branded demand, email and the pages most likely to be cited.
\nWrite each scenario on one page: trigger, first 24 hours, first week, owner. Run a 30 minute tabletop exercise twice a year.
\nA risk register is a simple table the whole team can read. It turns vague worry into assigned work. Keep it short enough that people will actually open it.
\nEach row holds the risk, the platform, the likelihood and impact on a plain high, medium, low scale, the early warning signal, the owner and the pre-agreed response. A sample row for a fictional pet supply brand, Biscuit Hollow Pet, reads like this: risk is account suspension; platform is the main social network; likelihood is low, impact is high; early warning is a rise in rejected ads or a warning notice; owner is the paid media lead; response is the written appeal packet plus the 48 hour budget shift from the scenario plan.
\nReview the register monthly, alongside your concentration ratios. Retire risks that have faded and add ones that appeared. A register that never changes is a decoration.
\nKeep a dated log of every policy notice and of your own account health: ad rejection rates, payment status, verification status. Many bad mornings are preceded by a slow drift you could have seen.
\nNew formats will keep arriving: new placements, new commerce features, new audio and video environments, new ways to appear inside an AI-written answer. Some will matter. Most will not, or will matter only after the early pricing disappears. Treat all of them with skepticism and a small test budget.
\nA practical rule is to cap tests at five to 10 percent of the total media budget, taken from the highest-uncertainty part of the plan rather than from proven performers. Before spending a dollar, write the hypothesis, the success metric, the test length and the kill rule. For example: \"We will spend 12,000 dollars over six weeks. If cost per first purchase is not within 30 percent of search by week six, we stop.\" Writing the kill rule in advance protects you from the common failure of a test that quietly becomes a permanent line item because someone grew attached to it.
\nInsist on measurement you can verify with your own data. If a new format only reports results inside its own dashboard, add a holdout, a unique landing page or a post-purchase survey question so you can compare against something the seller does not control. Urgency is a sales tool, so be wary of any pitch that says the early window is closing.
\nThe most reliable protection against platform change is a list of people who agreed to hear from you and a place where you store what you learned about them. Email, SMS where consent is clean, a community you host and an app are all examples. None are glamorous. All survive a bad morning.
\nOwned audience building is a long game. Every purchase, signup and event is a chance to ask for permission to continue the conversation, with a clear promise about what you will send. Collect only what you need, record consent and keep sensitive categories out of the profile. Parents and young audiences call for a conservative default: if you cannot be sure about age and consent, do not collect.
\nThink of a stadium. The walk-up crowd is exciting and moves with the weather. Season ticket holders, your list, are why the building pays its bills in a losing year. A platform is the walk-up crowd.
\nMeasure your owned audience the way you measure spend: reachable contacts, the share of customers you can contact directly, and revenue from owned channels.
\nConcentration risk: The exposure created when a large share of spend, revenue or audience depends on a single platform or partner. Answer layer: The summary or direct response a search engine or assistant writes on the results page, which can resolve a question without a click. Retail media: Advertising sold inside a retailer's own sites and apps, and sometimes on outside sites, using the retailer's shopper data. Holdout test: An experiment where a comparable group or region receives no ads, so the effect of advertising can be measured against it. Owned audience: The people you can contact directly, with their permission, without paying a platform for access.
\n1. Reading a strong dashboard as proof of a safe business, when the same dashboard shows more than half of acquisition riding on one account. 2. Opening a backup ad account only after the first account is suspended, which is too late and looks like evasion to reviewers. 3. Buying an emerging format because the pitch says the early window is closing, with no hypothesis, no cap and no kill rule. 4. Treating a retail media network or a creator relationship as a cure for dependence, when it often adds another closed system with self-reported results. 5. Building a risk register once, filing it, and never reviewing it against actual policy notices and account health.
\n1. Calculate spend, revenue and audience concentration every month and set a written threshold that triggers a review. 2. Keep verified business documents, payment methods and appeal contacts organized so an account problem becomes a paperwork task, not a crisis. 3. Cap new-format tests at a small share of budget, with a written hypothesis, test length and kill rule. 4. Grow an owned audience on a permission basis, with a conservative default around age and sensitive categories. 5. Rehearse two or three bad-morning scenarios each year with the people who would actually have to act.
\n1. Read the actual text of a policy change before reacting to commentary about it, since summaries often exaggerate or miss the clause that affects you. 2. Ask any seller of a closed environment how results are counted, over what window, and whether you can run a holdout. 3. Log ad rejection rates and account verification status weekly, because slow drift often precedes a sudden suspension. 4. Write creator contracts to cover disclosure, usage rights, approvals and what happens if the creator's account disappears, then have counsel review them. 5. Keep your own order data as the final check on any platform report, so a measurement change never leaves you guessing.
\nThree free, self-hosted tools cover the core of this chapter. Matomo (GPL) records visits and conversions on your own site, so you hold a measurement record that does not depend on any ad platform's reporting. Listmonk (AGPL) sends and manages permission-based email to your own list, which is the backbone of an owned audience. PostgreSQL stores your spend, revenue and contact data in one place, so concentration ratios and your risk register can be calculated with plain queries. None of the three requires a paid subscription. You host and maintain them yourself, which takes real time.
\nA dashboard shows performance, while a finance partner asks about exposure, and a healthy brand answers both. Platforms can change your economics through account actions, policy, price, measurement and competing products, and an answer layer in search adds exposure to traffic once thought stable. Retail and creator media bring genuine opportunity with closed systems attached, so they belong in a balanced plan with verification. Concentration ratios, scenario rehearsals and a short risk register convert worry into work. New formats get small, capped tests with written kill rules, and owned audiences remain the hedge that survives a bad morning.
\n1. Concentration on one platform is a business exposure even when the dashboard looks excellent. 2. Platform change arrives in five forms, and the quiet ones usually cost more over time than the dramatic ones. 3. Spend, revenue and audience concentration are three simple ratios you can calculate this week. 4. Scenario plans and a short risk register turn a bad morning into a rehearsed procedure. 5. Emerging formats earn budget through capped tests with a written kill rule, not through urgency.
\nThis week, calculate your top platform's share of spend and the share of your customers you can email directly, then write a one-page plan for the day that platform account goes dark.
\n1. Interactive Advertising Bureau. Digital advertising standards and guidelines, as published on the organization's site. 2. Federal Trade Commission. Guides Concerning the Use of Endorsements and Testimonials in Advertising, 16 CFR Part 255. 3. Media Rating Council. Minimum Standards for Media Rating Research. 4. Children's Online Privacy Protection Rule, 16 CFR Part 312. 5. Matomo, Listmonk and PostgreSQL project documentation, as published by each open source project.
"},{"id":"chapter-61-the-90-day-digital-advertising-operating-plan","title":"Chapter 61: The 90-Day Digital Advertising Operating Plan","html":"By the end of this chapter, the reader will be able to: 1. Describe the four phases of a 90-day plan (audit, instrument, test, scale) and the deliverable each phase must produce. 2. Assign a single named owner to every workstream and explain why shared ownership fails. 3. Apply a pass, adjust or stop rule at the day 30, day 60 and day 90 decision gates. 4. Construct a staffing model that uses an intern pipeline without putting budget or compliance at risk. 5. Evaluate a half-finished plan and identify which week is slipping and why. 6. Design a weekly operating rhythm that connects the companion spreadsheet tracker, the workbook and a free reporting dashboard.
\nA 90-day plan is too short to prove that your advertising works. Ninety days will not settle a brand's long-term pull, it will not smooth out seasonality, and it will barely cover one full buying cycle for a considered purchase. Anyone who tells you the plan produces a verdict on the whole program is selling comfort.
\nThe objection is a good one, though. Most advertising programs do not fail because the first 90 days were too short. They fail because nobody ever ran a first 90 days. The budget was approved in a meeting, the agency or the intern started buying, and a year later the team still could not say which tags fired, which audiences were real, or which test taught them anything. Waiting for the perfect long horizon is how a program drifts for four quarters.
\nSo I hold both positions. Ninety days cannot prove everything. It can prove a few specific things, and those few things are worth more than a year of vague activity. It can prove that your measurement is trustworthy enough to act on. It can prove that you have run at least one clean test and written down what it taught you. It can prove that money moved toward what worked and away from what did not. It can prove that someone other than you could run the program next week.
\nThink of it as a preseason, not the championship. Nobody wins a title in preseason, but the coach who skips it walks into game one with an untested roster.
\nA plan that produces activity is not a plan. A plan produces decisions, and each decision needs evidence that was not available on day one. That framing changes how you build the schedule.
\nFour phases run in order. Audit covers the first two weeks. Instrument runs through the end of week four. Test runs from week five through week eight. Scale runs from week nine through week 13. Week 13 ends on day 91, one day past the plan, so treat that last day as overflow and schedule nothing critical on it.
\nEach phase ends in a deliverable you can hold in your hand, and each deliverable is a file, not a feeling.
\n| Phase | Weeks | Deliverable | Gate | |---|---|---|---| | Audit | one to two | Written findings: accounts, tags, spend, creative, compliance | Feeds day 30 | | Instrument | three to four | Working tags, a baseline report, a live dashboard | Day 30 | | Test | five to eight | Two completed experiments, each with a one-page readout | Day 60 | | Scale | nine to 13 | Rebalanced budget and a drafted next-quarter plan | Day 90 |
\nBefore anything else, write one sentence that states the objective. It needs a number, a date and a boundary. \"Improve performance\" fails all three. \"Reduce cost per qualified lead on paid search and paid social combined by a meaningful margin we define in week two, while holding spend flat, by day 90\" passes. The margin gets set after the audit, because before the audit you are guessing at a baseline you have never seen.
\nThe audit has one purpose: find out what you actually own and what is actually happening. Most brand managers discover that the answer differs from the slide deck.
\nStart with access. List every ad account, analytics property, tag container, domain, pixel, payment method and creative folder, and for each one record who holds admin rights. If a former agency or a departed contractor is the only admin on an account, that is your first finding and it goes to the top of the page. Recovering control of an account you pay for is not optional work.
\nNext, trace the money. Pull 90 days of spend by channel from each platform's export, then pull what your own analytics recorded for the same period. The two will not match. Write down the gap's size and direction, and do not try to close it yet.
\nThen inventory creative. Count how many distinct ads ran, which ones carried most of the spend, and how old they are. A program that put most of its spend behind two ads that are months old has a fatigue problem waiting to surface.
\nFinally, check compliance basics: privacy notice language, consent handling on forms, disclosure on any endorsement or creator content, and a conservative default wherever the audience might include minors. This is operational guidance only, and your counsel decides what the law requires. Your job in week two is to list the questions to bring to counsel, not to answer them.
\nThe audit ends with a findings document of no more than five pages. Rank the findings by cost of leaving them alone. Pandoc can turn the same plain-text source into a PDF for leadership and a Word file for the team, so you write it once.
\nYou cannot test what you cannot measure, so these two weeks belong to instrumentation.
\nWeek three is tagging. Stand up your analytics, define the handful of events that matter (a lead submitted, a purchase completed, a call longer than a set number of seconds), and confirm each fires exactly once, on a phone, on a laptop and in a browser that blocks third-party cookies. Many teams find a form that fires its conversion twice, which quietly makes every campaign look better than it is.
\nWeek four is the baseline and the dashboard. Build one page that shows spend, clicks, leads or sales, cost per result and creative age, by channel, refreshed automatically. Put the same fields in the spreadsheet tracker so the two can be reconciled by hand once a week. The baseline is the average of what the program produced before your changes, and you will compare every later week against it.
\nThe day 30 gate. Pass means all of the following are true: every named conversion event has been verified as firing once, the dashboard has run for at least seven consecutive days without a manual fix, and the baseline is written down and signed off by the owner of measurement. Adjust means one item is missing and you give it one more week while the test design proceeds in parallel. Stop means the data cannot be trusted after the extra week. If that happens, do not start the test phase. Testing on broken measurement produces confident wrong answers, which are worse than no answers.
\nTesting is where discipline erodes fastest, because the temptation to peek and call a winner early is strong.
\nRun two experiments, no more. One should be a creative test, since creative is usually where the most room hides (Chapter 49 covers the reasoning). The second should be a structural test, for example changing the audience definition, the bid approach, or the landing page. Running more than two experiments with a small budget means none of them reaches a readable result.
\nEach experiment gets a written design before launch: the question, the single variable changed, the primary metric, the minimum number of conversions you need before you will read it, and the date you will read it. Write the stopping rule first. Suppose you commit to reading the test only after each arm has at least 100 conversions or after four weeks, whichever comes later. Then a lucky first week cannot talk you into a decision.
\nKeep a control, the unchanged version that runs alongside the new one. If a holiday weekend lifts every campaign, a test without a control takes credit for it.
\nGrowthBook can handle experiment assignment and basic statistical readouts for web tests, and it is free to run yourself.
\nThe day 60 gate. Pass means both experiments reached their stopping rule and each has a one-page readout stating what happened, what you believe caused it, and what you will do next. Adjust means one experiment is unreadable, in which case you document why and carry the budget into the scale phase unchanged. Stop applies to a channel, not the plan: if the audit and the tests both show a channel producing nothing you can trace, propose pausing it. A readout that says \"inconclusive\" is a legitimate result, and it is better than an invented win.
\nScaling does not mean spending more. It means moving the same dollars toward what you now have evidence for, and then increasing spend only where the evidence is strong.
\nBegin week nine by writing the allocation memo. For each channel, state the baseline result, the tested result if any, your confidence, and the proposed budget change. Keep moves modest: shift no more than a quarter of any channel's budget at once, wait two weeks, and read again.
\nUse weeks 10 and 11 to roll the winning creative and structure into more campaigns, and to retire what the tests showed was weak. Week 12 is for a second look at measurement. Compare the platform-reported results to your own analytics again. Has the gap changed? If it widened, find out why before you scale further.
\nWeek 13 is the planning week. Draft the next quarter: what you learned, which two tests come next, who owns what, and what you will stop doing. This draft is the real product of the 90 days.
\nThe day 90 gate. Pass means the allocation memo was acted on, a next-quarter plan exists with named owners, and the dashboard still runs without hand repair. Adjust means you extend one workstream by a defined number of weeks with a written reason. Stop means the objective sentence from week one can no longer be met under the current budget, and you take that finding to leadership with the data attached, not with an apology.
\nEvery workstream gets exactly one owner. Not a pair, not a committee. When two people own something, each assumes the other is checking, and the gap between them is where the double-firing tag lives.
\nFour workstreams cover the plan: measurement, creative, media buying and operations (which includes compliance and vendor contact). In a small team one person may own two of them. That is fine. What is not fine is a workstream with a department's name in the owner column.
\nStaffing is where many plans quietly collapse. You are the brand manager, you are in meetings, and the weekly work will not do itself. An intern pipeline can carry a lot of this weight if you build it with care. Interns and early-career staff are well suited to the audit inventory, tag verification checklists, creative naming and tagging, dashboard upkeep, and drafting readouts for your review. They should not hold payment methods, approve spend, or sign off on compliance wording. Give them a written checklist for each task, a named reviewer, and a short training session in week one.
\nPlan for turnover. A strong intern leaves in a semester, and the knowledge leaves with them unless every task ends in a document or tracker entry.
\nFor readers who want an in-house execution path, the Dakdan and Sports Media workforce and internship programs, including SkillBridge-style placements, are one informational option for sourcing trained early-career staff. Treat that as a staffing question to evaluate against your own needs, like any other.
\nThe companion spreadsheet tracker and the workbook exist so the plan lives in files you control. The tracker holds the schedule, the owners, the gate criteria and the weekly numbers. The workbook holds the exercises: the audit questions, the experiment design template and the allocation memo form.
\nThen run a fixed weekly rhythm. Keep it short, because a long meeting is one people start skipping.
\n1. Monday, 30 minutes: each owner states what shipped last week, what ships this week, and one blocker. Nothing else. 2. Midweek, async: owners update the tracker. A task not updated is treated as not done. 3. Friday, 15 minutes: the plan owner compares dashboard numbers to the tracker, notes any mismatch, and writes three lines of status for leadership.
\nAt the end of each month, the gate review replaces the Monday meeting. Bring the criteria, the evidence and the pass, adjust or stop call, in that order. Decide before the meeting what evidence would change your mind, so the review is not a performance.
\nMatomo can serve as your own analytics source, and Metabase can read the tracker data and the analytics database to display progress on a single page.
\nConsider a composite archetype, Northline Outfitters, a fictional regional retailer with a modest monthly budget. The numbers are illustrative.
\nIn the audit, the owner finds that a former agency still holds admin rights on the main search account, and that the site's checkout event fires twice on a slow connection. By week four the double-firing is fixed. The baseline, now trustworthy, shows a cost per sale about a fifth higher than the old dashboard claimed.
\nThat is an uncomfortable day 30 meeting, and it is a good one. The team passes the gate on measurement and starts testing honestly.
\nIn the test phase, a customer-story ad beats a product-focused ad on cost per sale, while an audience test is inconclusive and gets documented as such. In the scale phase they move a quarter of the weakest channel's budget toward the customer-story format and confirm the gain held. The program did not become perfect. It became legible, and that is what 90 days can buy.
\nDecision gate: A scheduled checkpoint where the team compares evidence to written criteria and decides to pass, adjust or stop. Baseline: The verified average of program results before any change, used as the comparison point for every later result. Control: The unchanged version of an ad or setup that runs alongside a test version to separate real effects from outside noise such as seasonality. Stopping rule: A pre-written condition, such as a minimum number of conversions or a date, that decides when a test may be read. Single owner: The one named person accountable for a workstream's deliverables, as distinct from the people who contribute to it.
\n1. Launching new campaigns during the instrumentation weeks, which fills the baseline with data from a setup you are about to change. 2. Naming a team or department as the owner of a workstream, so that nobody in particular answers when it slips. 3. Reading a test early because the first week looks good, then reversing course when the result regresses. 4. Running five small experiments at once, which leaves every one of them too thin to read. 5. Letting interns hold payment methods or approve compliance wording because there was no one else free.
\n1. Write the one-sentence objective with a number, a date and a boundary, and set the number only after the audit. 2. Write the gate criteria in week one, before any results exist, so nobody can adjust the standard to fit the outcome. 3. Keep every plan document in plain text or a spreadsheet you own, so it survives changes of staff and software. 4. Reconcile platform-reported results with your own analytics every week and record the gap, even when you cannot close it. 5. End every task with a written artifact: a checklist, a tracker entry or a readout.
\n1. Do the access audit first. A surprising number of accounts turn out to be controlled by someone who no longer works with you. 2. Pick two experiments and protect their budgets. A third idea can wait for next quarter. 3. Schedule the gate reviews on the calendar in week one, with leadership invited, so the dates are fixed before work begins. 4. Record inconclusive results in the same format as wins. A library of honest nulls prevents you from repeating a bad idea. 5. When a channel misses its gate, cut the budget by a fixed fraction first rather than switching it off, and read the effect before the next move.
\nFour free tools carry most of this plan. Matomo (GPL) provides self-hosted web analytics for the instrumentation phase. Metabase (AGPL) turns your tracker and analytics data into the shared dashboard. GrowthBook (MIT core) assigns visitors to experiment arms and produces readouts for the test phase. Pandoc (GPL) converts one plain-text source into the PDF and Word versions of your audit, readouts and allocation memo. None of these requires a paid subscription. A full-scale demand-side buying platform has no open source equivalent at parity, so the media buying itself still happens inside the ad marketplaces.
\nA 90-day plan cannot prove that a program works over the long run, but it can prove that measurement is trustworthy, that at least one test was run cleanly, that budget moved toward evidence, and that the work can survive a change of staff. Four phases carry it: audit, instrument, test and scale, with gates at day 30, day 60 and day 90, each with written pass, adjust and stop criteria. Every workstream has a single owner, an intern pipeline supplies capacity under review, and a short weekly rhythm keeps the tracker, workbook and dashboard consistent with one another.
\n1. Ninety days will not prove a whole program, but it can prove measurement, one clean test, one budget move and one handoff. 2. Write the gate criteria before results exist, then apply pass, adjust or stop without renegotiating. 3. Keep experiments to two, give each a written stopping rule, and run a control alongside. 4. Assign one owner per workstream, and use interns for inventory, checks and drafts, never for spend or compliance sign-off. 5. Finish with a next-quarter plan, since the draft is the most durable product of the 90 days.
\nBefore you spend another dollar this quarter, put the three gate dates on the calendar and write the pass, adjust and stop criteria for each. Then name one owner per workstream by the end of this week.
\n1. Interactive Advertising Bureau Tech Lab. OpenRTB specification, current version. Primary specification for programmatic auction mechanics. 2. Media Rating Council. Minimum Standards for Media Rating Research. Measurement standards. 3. World Wide Web Consortium. Web Content Accessibility Guidelines, current version. Specification for accessible advertising creative. 4. Kohavi, R., Tang, D. and Xu, Y. Trustworthy Online Controlled Experiments. Cambridge University Press. Practical guide to running and reading online tests.
"},{"id":"chapter-62-conclusion-building-a-durable-advertising-engine","title":"Chapter 62: Conclusion: Building a Durable Advertising Engine","html":"By the end of this chapter, the reader will be able to: 1. Define a durable advertising engine as three connected parts: repeatable tests, clean data, and trained people. 2. Explain why reported campaign efficiency and true incremental efficiency can differ by a wide margin. 3. Design a 30 day plan for standing up one open source measurement stack. 4. Construct a simple incrementality test that can be finished before the next budget review. 5. Evaluate an advertising operation using test velocity, data quality, and team capability. 6. Assemble an execution plan that assigns every task to a named owner and a due date.
\nLet me start with arithmetic, because arithmetic is where comfortable stories go to die.
\nSuppose a brand spends 60,000 dollars in one quarter on retargeting, the ads shown to people who already visited the site. The platform reports 600 purchases. That is 100 dollars per purchase, a number that looks wonderful in a slide deck. Now suppose the team finally holds back a random slice of those visitors and shows them nothing. The holdout group buys at a rate that implies 450 of those 600 purchases would have happened anyway. The ads produced 150 extra purchases. The real cost per extra purchase is 400 dollars, four times the number on the slide.
\nNothing in that example involves fraud or a broken pixel. The platform did exactly what it was built to do: count every purchase it could touch. The brand simply never asked the harder question.
\nHere is the assumption I want to challenge across this last chapter. Most brand managers believe a good advertising result is something you achieve. A strong quarter, a winning campaign, a lucky creative. I have run enough operations to say that a good result is something you can lose just as fast as you found it, because the person who found it moves on, the tracking breaks after a site update, or the platform changes a default setting overnight.
\nA result is an event. An engine is a machine that produces events on schedule. The difference decides who still has a job, a budget, and a negotiating position in three years. This chapter is the canonical home of a simple idea: the durable engine is made of repeatable tests, clean data, and trained people. Take any one of them away and you have a campaign with good luck attached. Then I will ask you to make three commitments before you close this book.
\nA test is repeatable when a stranger could run it again next quarter using only your written notes. That sounds like a low bar. Most teams fail it.
\nThink of a baseball club that keeps a scouting file on every player. The file does not make any one game winnable. It makes the next decision cheaper, because someone already did the homework. A test library works the same way. Each entry holds the question, the hypothesis (a specific guess about what will happen and why), the audience, the spend, the dates, the result, and what the team decided afterward.
\nThree kinds of tests belong in the library. Creative tests compare messages, offers, and formats. Audience tests compare who sees the ad. Incrementality tests ask whether the ads caused anything at all. The third kind is the rarest and the most valuable, and I will return to it in its own section.
\nA few rules keep the library honest. Change one thing at a time when you can. Decide before launch what result counts as a win, and write it down. Set the end date in advance, so nobody stops the test the moment it looks good. Record the losers with the same care as the winners, because a failed test that is written down saves the next manager from paying for the same lesson twice.
\nSuppose a team runs one test a month for a year. Even if each winning test improves one metric by only a few percent, and only half the tests win, the gains stack up, while a team running no tests stays exactly where it began.
\nAsk three people in a meeting what last month's cost per lead was. If you get three answers, you do not have a data problem. You have a trust problem, and no test result will survive it.
\nClean data means one agreed definition for each number, one place where each number lives, and a record of how it got there. The first definition fight is almost always about what counts as a conversion. A form fill? A phone call that lasts over a minute? A purchase that is later refunded? Settle it once, in writing, and apply it to every channel.
\nThe second piece is first-party data, meaning information your own properties collect directly from people who visit or buy from you. You own it, you can audit it, and no platform can quietly change how it is counted. Platforms will always report their own version of events, usually the most flattering version. Your records are the referee.
\nThe third piece is a habit of checking. Compare ad platform clicks to site visits. Compare reported purchases to orders in the store system. Expect some gap, since no two systems agree perfectly, but know the size of the gap and watch it move. A gap that suddenly doubles usually means something broke, and you want to find that on a Tuesday, not at the quarterly review.
\nFamily-first note: when your data includes children or any sensitive category, collect less, keep it for a shorter time, and let counsel review the policy. Operational guidance only. Legal decisions belong to your attorney.
\nTools do not run themselves, and neither does a dashboard nobody understands. I learned this in the Army long before I learned it in marketing. A unit with great equipment and untrained soldiers loses to a modest unit that drills until the basics are automatic.
\nAdvertising teams need the same drilling. Everyone who touches spend should be able to explain, in plain words, what a conversion is in your shop, how a test is set up, and why a platform report is a claim and not a fact. That is a college freshman level of understanding, and it is enough to prevent most expensive mistakes.
\nCross-training matters more than most managers admit. If only one person knows how the tracking works, your engine has a single point of failure and a vacation schedule. Rotate people through the data work. Make the buyer sit in on a test review. Make the analyst read a creative brief.
\nWrite things down as you go. A short runbook for each recurring task (launching a test, checking tag health, refreshing a report) turns one person's skill into the team's skill. Interns and new hires become productive in days instead of months when the runbooks exist. Teams that run structured internship or workforce programs have a built-in advantage here, because teaching the process forces the team to document it. Skills that live only in one head leave when that head does.
\nOver 61 chapters I have made a number of arguments. I do not want to list them again. I want to show how they fit together, like the parts of a car.
\nThink of the tests as the steering. They tell you where the money should go next. Clean data is the fuel line, and nothing works if it clogs. The people are the drivers. Everything else in the book is a condition of the road or a warning light.
\nThe warnings are worth reading in that light. Platform reports are trackable but not truly measured, so you need tests to find the real effect. Every attribution model leans in a known direction, so you triangulate instead of trusting one. Owning your path to the audience protects you when a platform changes terms. Compliance is part of daily operating routine, which means it belongs in the runbook, not in a once-a-year scramble.
\nThe open tools matter for a practical reason. They keep your cost of learning low, so you can afford to run more tests, and they keep your records in your hands. A team that rents every capability from a vendor negotiates with its back to the wall.
\nNotice the common thread. Each lesson in the book makes one of three things stronger: your ability to learn, your ability to trust what you learned, or your ability to act on it. That is the machine.
\nCommitment number one: build one open source measurement stack inside the next 30 days. One. Do not build three half-working ones.
\nA workable stack has three layers. Matomo, a self-hosted web analytics tool, records visits and conversions on your own server. A database holds the cleaned records. Metabase, a dashboard tool that connects to that database, shows the team the same numbers every morning. If you plan to test for causal effects, CausalImpact analyzes the results afterward.
\nHere is a sane schedule. In week one, write your conversion definitions and name an owner for each number. In week two, install the analytics tool and tag your most important conversion on the site. In week three, connect the database and build one dashboard with no more than six numbers on it. In week four, reconcile your numbers against the store, billing, or call records, document the gap, and write the first runbook.
\nKeep the first dashboard small on purpose. A single page the whole team actually reads beats a 40 chart monster nobody opens. Hosting costs are real (a small server is not free), and so is the staff time, so budget both honestly instead of pretending open source means zero cost.
\nCommitment number two: before your next budget review, run one incrementality test. This is the experiment that opens the chapter, and it is the cleanest way to find out which dollars are earning their keep.
\nThe simplest design is a holdout. Pick your largest spend channel, or the one you trust least. Split the audience or the markets into a test group that sees ads and a holdout group that does not. Run both for a fixed period, long enough to catch a full buying cycle. Compare outcomes.
\nIf you cannot split people, split geography. Choose matched regions with similar sales patterns, pause ads in a few of them, and compare against the rest. CausalImpact, a free statistical package, builds an expected baseline from your pre-test data and shows how far the real results moved from it.
\nWork out the stakes in advance. Suppose a channel eats 25,000 dollars a month and the test suggests only one third of its reported results are truly caused by ads. You now have a factual basis to move part of that budget to a channel or a creative approach you can test next. You also walk into the review with evidence instead of an argument.
\nExpect resistance. Some people will say pausing ads for a holdout group costs sales. It costs a small, planned, temporary amount, set against a large unexamined one. Agree on the loss limit before the test begins, and counsel should review any contract terms that restrict pausing placements.
\nCommitment number three is where most plans die. A plan with no builder is a hobby.
\nIf you have the staff, the time, and the skills to build the stack, run the test, and train the team yourself, do it. Many brands do not have all three. For those readers, my own in-house execution path runs through Dakdan, Sports Media, and ChaseDaddy, and I offer it as the practical way to get the work built instead of just planned. Dakdan supplies the books, the training material, and the workforce programs that develop your people. Sports Media supplies sports-anchored digital placement and the athlete and creator relationships that go with it. ChaseDaddy supplies the creative and design production your tests will keep demanding. That is a call to action, and I mean it plainly: bring your 30 day plan and your test design, and let us put names and dates on each task.
\nWhichever route you take, bring the same discipline. Assign every task one owner. Give every task a due date. Hold a short check-in each week and a longer test review each month. Plans fail from missing owners far more often than from bad ideas, and a calendar is the cheapest fix there is.
\nAn engine needs a gauge. Three numbers will do.
\nTest velocity is the number of finished tests per quarter, counting only tests that reached a written decision. Two or three clean tests per quarter is a solid start for a small team. A big number means nothing if the tests were sloppy, so inspect quality alongside count.
\nData quality is a simple check, run monthly. Is each core number complete, current, and in agreement with an outside source within a gap you have agreed on? Score it pass or fail for each number and keep the history.
\nTeam capability is the hardest to measure and the easiest to fake. Ask a new teammate to launch a test using only the runbook. If they succeed, the capability is in the system. If they have to find the one person who knows, it is not.
\nReview all three at the same meeting where you review spend. The meeting should feel like a coach reviewing film, asking what happened and what changes next week, rather than a trial. When these gauges move the right way for four quarters running, you will no longer need to hope for a good result. You will have built the thing that makes them.
\nIncrementality: The extra results caused by advertising, beyond what would have happened with no ads at all. Holdout group: A slice of the audience or set of markets deliberately kept from seeing ads, used as the comparison point in a test. First-party data: Information a business collects directly from its own customers and visitors, on properties it controls. Runbook: A short written procedure for a recurring task, detailed enough that a new teammate can follow it alone. Test velocity: The number of tests a team completes and documents in a set period, usually a quarter.
\n1. Treating a strong campaign result as proof of skill, then failing to document how it happened, so the result cannot be repeated. 2. Building a dashboard before agreeing on definitions, which produces a polished screen full of numbers nobody trusts. 3. Running a test with no written decision rule, then arguing afterward about whether the result counts as a win. 4. Letting one person hold all the tracking knowledge, which turns a vacation or a resignation into a measurement outage. 5. Writing a thorough plan with no owners and no dates, then wondering at the next review why nothing was built.
\n1. Write every conversion definition once, assign an owner to it, and apply it to every channel. 2. Log every test in a shared library, losers included, with the question, setup, result, and decision. 3. Reconcile platform reports against your own records every month and track the size of the gap. 4. Keep a runbook for each recurring task and test it by handing it to someone who has never done the task. 5. Set the loss limit and the end date of every holdout test before it begins.
\n1. Start with the channel you trust least, because that is where a holdout test is most likely to change a decision. 2. Keep the first dashboard to six numbers or fewer, and add a seventh only when someone asks for it twice. 3. Run a short test review the same week every month, so the habit survives busy seasons. 4. Pair a creative test with every audience test whenever you can, since a weak message can hide a strong audience. 5. Budget server hosting and staff time for open source tools up front, because free software still needs someone to run it.
\nThree free tools cover the core of the engine. Matomo (GPL license) is a self-hosted web analytics platform that records visits and conversions on a server you control. Metabase (AGPL license) connects to your database and turns the cleaned records into dashboards the whole team can read without writing code. CausalImpact (Apache 2.0 license) is a statistical package that estimates what would have happened without your ads, which makes it a natural partner for geographic holdout tests. None of these tools carries a license fee, though each needs hosting and a person who owns it. Where a task needs more than these three provide, such as large-scale automated media buying, no open source option matches the paid platforms today, and I would rather say so plainly.
\nA durable advertising engine is repeatable tests, clean data, and trained people working as one machine, and removing any one of them leaves you with a campaign that depends on luck. Reported efficiency can overstate the true effect of advertising by a wide margin, which is why a holdout test belongs in every serious program. The book's separate lessons all strengthen your ability to learn, to trust what you learned, or to act on it. Three commitments turn this into work: build one open source measurement stack in 30 days, run one incrementality test before the next budget review, and give the plan builders, owners, and due dates. Test velocity, data quality, and team capability tell you whether the engine is improving.
\n1. A result is an event, and an engine is the system that produces results on a schedule you can plan around. 2. Platform-reported efficiency and true incremental efficiency can differ by multiples, and only a holdout test reveals the gap. 3. One agreed definition per number is the foundation, since a dashboard built on disputed definitions only spreads the dispute. 4. Skills documented in runbooks outlast any individual, which makes training and documentation the same activity. 5. A plan becomes real only when every task has one owner and one date, whether you build it alone or with a partner.
\nWithin 30 days, put one measurement stack live with one dashboard and one owner per number, and before your next budget review, hold back a slice of your largest channel and learn what it truly earns.
\n1. Brodersen, K.H., Gallusser, F., Koehler, J., Remy, N., and Scott, S.L. Inferring causal impact using Bayesian structural time-series models. Annals of Applied Statistics. Published methodology behind the CausalImpact package. 2. Matomo project documentation. Self-hosted web analytics installation and configuration guides. 3. Metabase project documentation. Database connection and dashboard building guides. 4. Federal Trade Commission. Dot Com Disclosures: Information about Online Advertising. Guidance on clear advertising disclosure. 5. Media Rating Council. Minimum Standards for Media Rating Research. Published measurement standards.
\nTools named in this book, with the license each is published under and the chapters where each appears. Open core means the main project is open source while some add-on features are not; this book recommends only the open parts. The license rule in Chapter 59 applies to models, not to the other tools listed here. Confirm the current license terms at the source before you deploy any tool.
\nMatomo: Web and product analytics. License: GPL. Source: matomo-org/matomo. Chapters 1, 5, 14 to 15, 17 to 18, 31, 33, 36, 38, 40 to 44, 57, 60 to 62.
\nMatomo Tag Manager: Tag management. License: GPL. Source: matomo-org/tag-manager. Chapter 43.
\nUmami: Web analytics. License: MIT. Source: umami-software/umami. Chapter 1.
\nPlausible Community Edition: Web analytics. License: AGPL. Source: plausible/analytics. Named in this appendix only.
\nPostHog (self-hosted core): Product analytics. License: MIT core. Source: PostHog/posthog. Chapter 38.
\nJitsu: Server-side event collection. License: MIT. Source: jitsucom/jitsu. Chapters 38, 43.
\nRevive Adserver: Ad serving for owned inventory. License: GPL. Source: revive-adserver/revive-adserver. Chapters 4, 28.
\nPrebid.js and Prebid Server: Publisher-side header bidding. License: Apache 2.0. Source: prebid/Prebid.js. Chapters 4, 28 to 29.
\nGrowthBook: A/B testing (open core). License: MIT core. Source: growthbook/growthbook. Chapters 9, 26, 49 to 51, 61.
\nUnleash: Feature flags. License: AGPL. Source: Unleash/unleash. Chapter 50.
\nMeridian: Marketing mix modeling. License: Apache 2.0. Source: google/meridian. Chapter 46.
\nRobyn: Marketing mix modeling. License: MIT. Source: facebookexperimental/Robyn. Chapter 46.
\nPyMC-Marketing: Marketing mix modeling. License: Apache 2.0. Source: pymc-labs/pymc-marketing. Chapters 27, 46, 56.
\nCausalImpact: Causal and lift analysis. License: Apache 2.0. Source: google/CausalImpact. Chapters 1, 5 to 6, 10, 13, 30, 33, 45, 62.
\nstatsmodels: Statistics. License: BSD. Source: statsmodels/statsmodels. Chapters 1, 45, 48, 51, 58.
\nscipy: Statistics. License: BSD. Source: scipy/scipy. Chapters 45, 51.
\nMetabase: Dashboards. License: AGPL. Source: metabase/metabase. Chapters 1, 10, 13, 20, 34, 36, 40, 44, 47, 55, 61 to 62.
\nApache Superset: Dashboards. License: Apache 2.0. Source: apache/superset. Chapter 47.
\nMeltano: Data pipelines. License: MIT. Source: meltano/meltano. Chapter 41.
\nApache Airflow: Data pipeline scheduling. License: Apache 2.0. Source: apache/airflow. Named in this appendix only.
\ndbt-core: Data transformation. License: Apache 2.0. Source: dbt-labs/dbt-core. Chapters 41, 47.
\nPostgreSQL: Database. License: PostgreSQL license. Source: postgres/postgres. Chapters 5, 14 to 15, 32, 36, 39, 41 to 44, 47, 60.
\nDuckDB: Analytical database. License: MIT. Source: duckdb/duckdb. Chapters 5 to 7, 13, 17, 20, 22, 25, 27, 30, 34, 37, 44, 47 to 48, 55 to 58.
\nListmonk: Email and newsletters. License: AGPL. Source: knadh/listmonk. Chapters 19, 35, 60.
\nMautic: Marketing automation. License: GPL. Source: mautic/mautic. Chapters 19, 35, 39.
\nPostal: Mail delivery server. License: MIT. Source: postalserver/postal. Chapters 19, 35.
\nPostiz: Social post scheduling. License: AGPL. Source: gitroomhq/postiz-app. Chapter 18.
\nLighthouse: Page speed and quality audits. License: Apache 2.0. Source: GoogleChrome/lighthouse. Chapters 9, 12, 31, 50.
\nsitespeed.io: Scheduled page speed testing. License: MIT. Source: sitespeedio/sitespeed.io. Chapters 9, 12, 50.
\nScrapy: Crawling and audits. License: BSD. Source: scrapy/scrapy. Chapters 11 to 12, 23, 29, 32.
\nKlaro: Consent management. License: BSD. Source: kiprotect/klaro. Chapters 15, 33, 41 to 42.
\npandas: Data analysis. License: BSD. Source: pandas-dev/pandas. Chapters 7 to 8, 11, 37.
\nscikit-learn: Machine learning. License: BSD. Source: scikit-learn/scikit-learn. Chapters 8, 32.
\nspaCy: Language processing. License: MIT. Source: explosion/spaCy. Chapters 8, 32, 59.
\nGIMP: Image editing. License: GPL. Source: GNOME/gimp. Chapter 16.
\nInkscape: Vector graphics. License: GPL. Source: project documentation. Chapters 16, 49.
\nKrita: Digital painting. License: GPL. Source: KDE/krita. Chapters 16, 49.
\nPenpot: Design and prototyping. License: MPL 2.0. Source: penpot/penpot. Chapters 16, 49 to 50.
\nKdenlive: Video editing. License: GPL. Source: KDE/kdenlive. Chapters 16, 21, 26.
\nShotcut: Video editing. License: GPL. Source: mltframework/shotcut. Chapters 16, 26.
\nBlender: 3D and video. License: GPL. Source: blender/blender. Chapter 16.
\nAudacity: Audio editing. License: GPL. Source: audacity/audacity. Chapter 24.
\nFFmpeg: Media processing. License: LGPL and GPL. Source: FFmpeg/FFmpeg. Chapters 16, 21, 24, 26.
\nWhisper: Transcription and captions. License: MIT. Source: openai/whisper. Chapters 16, 24, 26, 59.
\nOllama: Local model hosting. License: MIT. Source: ollama/ollama. Chapter 59.
\nPandoc: Documentation conversion. License: GPL. Source: jgm/pandoc. Chapters 52 to 54, 61.
\nMoodle: Learning platform. License: GPL v3. Source: moodle/moodle. Chapter 54.
\nLibreOffice Calc: Spreadsheets. License: MPL 2.0. Source: project documentation. Chapters 6, 48, 56.
\nThese are the formulas used throughout the book, written in words so they work in any spreadsheet. The companion spreadsheet has each one built as a working calculator in both directions: budget in and result out, or target result in and required budget out.
\nCost per thousand impressions (CPM) equals spend divided by impressions, multiplied by 1,000. Rearranged, impressions equal spend divided by CPM, multiplied by 1,000.
\nClick-through rate equals clicks divided by impressions. Cost per click (CPC) equals spend divided by clicks. CPC also equals CPM divided by 1,000, divided by click-through rate.
\nConversion rate equals conversions divided by clicks. Cost per acquisition (CPA) equals spend divided by conversions. CPA also equals CPC divided by conversion rate.
\nReturn on ad spend (ROAS) equals revenue attributed to the ads divided by ad spend. Break-even ROAS equals one divided by contribution margin, where contribution margin is the share of each sale left after product and fulfillment costs.
\nRequired budget for a target equals target conversions multiplied by expected CPA. Required impressions for a target equals required budget divided by CPM, multiplied by 1,000.
\nIncremental CPA equals spend divided by incremental conversions. Incremental conversions equal conversions in the exposed group minus the conversions the same group would have produced without the ads, estimated from a holdout.
\nFrequency equals impressions divided by reach. Combined reach under an independence assumption equals one minus the product of each channel's unreached share. Real audiences overlap, so treat the result as a rough planning figure.
\nDaily pacing target equals remaining budget divided by remaining days. Pace ratio equals spend to date divided by the planned spend to date. A ratio below one means underspend, and a ratio above one means overspend.
\nCustomer lifetime value payback in months equals acquisition cost divided by monthly contribution margin per customer.
\nThis calendar is operational guidance, not legal advice. Your counsel decides what applies to you. Use it to make sure routine duties do not slip.
\nWeekly: review new creative for required disclosures where creators or affiliates are paid. Check that every active email and message program honors opt-outs within the time your counsel sets. Scan ad accounts for policy warnings and disapproved ads.
\nMonthly: reconcile consent records against the contact lists you send to. Review data subject requests received, closed, and overdue. Check that tags on your sites respect the consent choices recorded. Review vendor access to your data.
\nQuarterly: audit the consent banner against the tags that actually fire. Review partner and affiliate disclosures. Review retention rules and delete expired data. Test the process for a request to delete personal data from start to finish.
\nTwice a year: review your privacy notice against your real data practices. Review vendor agreements for data handling terms. Rehearse your response to a platform account suspension.
\nWhen anything changes: a new vendor, a new data source, a new audience type, a new region, or a new messaging channel should trigger a review before launch, not after.
\nWrite these answers down before the test starts and keep them with the results.
\nQuestion: what decision will this test change?
\nHypothesis: what do you expect to happen, and in which direction?
\nDesign: audience holdout, geography holdout, matched markets, or creative split. Why this design?
\nUnit of assignment: people, households, or regions. How are units assigned?
\nPrimary metric and guardrail metrics: what decides the result, and what must not get worse?
\nSample size and duration: how many units, how many weeks, and how was the number chosen?
\nDecision rule: what result means scale, hold, or stop? Write the thresholds before you see data.
\nKnown risks: contamination, seasonality, promotions, outages, or changes in the budget during the test.
\nResults and interval: the effect, the range of plausible values, and what the range means for the decision.
\nDecision taken and date.
\nAsk these in writing before spending and again at every renewal.
\n1. Which entity holds the contract with the publisher, and is it a direct relationship or a reseller? 2. Where do your seller identifiers appear in the publisher's published authorization files, and as direct or reseller? 3. How many intermediaries sit between my budget and the publisher, and what does each one keep? 4. What share of my spend reaches the publisher, and will you state it in the insertion order? 5. Will you provide log-level data, or at minimum domain and app level delivery reports? 6. Which inventory is made for advertising, and how do you identify and exclude it? 7. How do you detect and credit invalid traffic, and how long is the credit window? 8. What are your viewability and completion definitions, and which standard do you follow? 9. Who owns the data collected from my campaigns, and how long do you keep it? 10. What notice do I get before fees, inventory sources, or measurement methods change?
\nWorkbook: one page set for each of the 62 chapters, with a chapter goal, three exercises, a worksheet you fill from your own records, a five point self-check, and one action commitment. Use it with your team in working sessions.
\nCompanion Tracker (spreadsheet): 23 sheets of working calculators and logs. They cover media planning by channel, a media buying calculator that works forward from spend and backward from a target cost per acquisition, budget allocation, search tier caps, the supply path fee waterfall, return tracking, incrementality and significance, test sample size, attribution reconciliation, frequency and reach, creative tests, email and text message economics, affiliate and retail media adjustments, response curves, unit economics, pacing, seasonality, vendor scoring, a compliance calendar, a 54 item audit scorecard, a 90 day tracker, and a chapter map. Input cells are shaded and every formula returns a dash when an input is blank.
\nAudit Checklist: 54 evidence-based checks, six for each part of the book, with a scoring sheet and sign-off lines. Use it at the start of every quarter and before any budget increase.
\nCourse Package: 62 modules of about 90 minutes each, with outcomes, a timed lesson plan, a practice activity, a five question quiz with an answer key, and a three criterion rubric, built for any learning platform. The package defines a completion standard but issues no credential, and it does not claim accreditation or college credit.
\nAudio Script: numbered narration text files for every chapter, a pronunciation watch list, and production notes, plus one combined script for single-file import.
\nDesign Brief: instructions for the cover, the print wrap, the companion asset branding, and launch graphics, written for the publisher's design partner.
\nA2P (application-to-person): Messaging sent by software to people, as opposed to person-to-person texts, and the category that carriers require brands to register.
\nA/A test: A check in which traffic is split between two identical versions to confirm the testing setup does not invent differences.
\nActivation: The marketing a brand does after signing a sponsorship to turn permission into attention, traffic and sales.
\nAd pod: The group of ads that plays together in one break between content.
\nAdstock: A way of modeling the lingering effect of advertising, where each week carries over a share of the previous week's impact.
\nads.txt: A public file on a publisher's domain listing the companies authorized to sell that publisher's ad space.
\nAffiliate network: A third-party platform that lists brands and publishers, tracks clicks and orders, and handles commission payments for a fee.
\nAnswer layer: The summary or direct response a search engine or assistant writes on the results page, which can resolve a question without a click.
\nAssets (extensions): Optional pieces added to a text ad, such as extra links, callouts, and structured lists, that add detail and must be honored on the destination.
\nAssisted conversion: A sale or signup where a given channel appeared earlier in the visitor's path but was not the final touch before the action.
\nAttributed sales: Sales the retailer credits to an ad according to its own rules for window, click and view credit.
\nAttribution window: The period after an ad interaction during which a conversion is credited to that ad.
\nBaseline: The verified average of program results before any change, used as the comparison point for every later result.
\nBid shading: Lowering a bid below your true value in a first-price auction to keep surplus while still winning often enough.
\nBlackout: A rights restriction that prevents an event from streaming in certain locations, often the home market of a team.
\nBrand defense: Bidding on your own name so your ad, not a rival's, appears on the results page for it.
\nBrand safety: The practice of keeping ads away from content that is nearly universally unacceptable for advertisers, such as illegal or graphic material.
\nBrand suitability: A brand-specific judgment of whether acceptable content fits that brand, as opposed to brand safety, which is the shared floor of content nobody should fund.
\nBranded query: A search that includes the company, product or person name, usually from someone who already knows the brand.
\nBranded search: Searches that include your company or product name, which usually signal existing awareness.
\nBreak-even return on ad spend: The minimum ratio of ad-attributed revenue to ad cost at which a product's contribution margin covers its advertising.
\nB-roll: Supporting footage, such as product shots or settings, laid over narration or used to build variations without new filming.
\nBurn-in captions: Text rendered permanently into the video image so it displays on every device regardless of settings.
\nCalibration: Adjusting a model so its estimates agree with the results of real experiments.
\nCannibalization: A situation in which one of your campaigns takes clicks that another of your campaigns, or your organic listing, would have earned at lower cost.
\nCanonical tag: A line of page code naming the preferred version of a page, so that duplicates do not compete with each other.
\nCareer ladder: A written set of levels, each with defined scope, judgment, output, and influence, used to guide promotion and pay.
\nCeiling: The maximum useful budget for a channel, set by demand, inventory, frequency or operational capacity.
\nChannel mix: The set of advertising channels a business runs at one time and the share of budget each receives.
\nClassification: Sorting items such as comments or captions into categories you define.
\nClawback: The reversal of a commission already paid, usually because an order was refunded, canceled, or found to violate program rules.
\nClean room: A controlled environment where two parties compare data and receive aggregate results without either side seeing the other's raw records.
\nClick injection: A fraud pattern where a fake click is fired just before a genuine install so the fraudster takes credit for it.
\nClose variant: A query the platform treats as equivalent to your keyword, such as a plural, misspelling or reworded version with the same meaning.
\nCohort: A group of users who installed in the same period or from the same source, tracked together so you can compare their behavior over time.
\nCommission-based fee: Compensation calculated as a share of media spend, which can reward higher spend whether or not results improve.
\nCompanion unit: A secondary display element, such as a small banner or card, that appears alongside the video while it plays.
\nCompletion rate: The share of started plays that reach the end of the ad, which is heavily shaped by whether the format allows skipping.
\nConcentration risk: The exposure created when a large share of spend, revenue or audience depends on a single platform or partner.
\nConcept: The central idea or promise an ad is built on, independent of its wording, format, or visuals.
\nConquesting: Bidding on a competitor's name to show your ad to people searching for that company.
\nConsent ledger: An append-only record of every consent event, including time, wording version, categories and source.
\nContent recommendation widget: A module under or beside an article that lists links to other content, some of which are paid placements bought through a network.
\nContextual targeting: Placing an ad based on the content of the page or program the person is viewing, not on a profile of the person.
\nContingency reserve: A set-aside portion of budget held for unscheduled events and released only under written rules.
\nContribution margin: Revenue from a sale minus its variable costs, such as product cost, shipping, payment fees and returns. It is the money available to pay for advertising and overhead.
\nControl: The current best-performing ad, which new variations must beat to replace it.
\nCookie stuffing: Setting a partner tracking cookie in a shopper's browser without that shopper actually clicking the partner's link.
\nCore web vitals: A set of page experience measures covering load time, responsiveness, and visual stability, used to judge how a page feels to a visitor.
\nCounterfactual: The outcome that would have occurred if the advertising had not run, which no platform can observe directly.
\nCrawl budget: The amount of attention a search engine is willing to spend fetching pages from a given site, which matters most on large sites with many near-duplicate pages.
\nCreative fatigue: The decline in results that occurs when an audience has seen an ad often enough to stop responding to it.
\nCuration: The packaging of inventory by a third party into a ready-made deal, often by audience or theme, usually for an added fee.
\nData clean room: A controlled environment where two parties match data and receive only aggregate answers, without exposing raw records to each other.
\nData freshness: The time elapsed since a dataset was last successfully updated, used as an early warning that a pipeline has stalled.
\nData layer: A structured object on a page that holds the facts about a visit and its events, read by the tag manager rather than scraped from the visible page.
\nData map: An inventory of the systems holding personal data, with purpose, owner, vendors and retention for each.
\nData minimization: The practice of collecting only the information needed for a stated purpose, and nothing more.
\nData processing addendum: A contract add-on that sets what a vendor may do with personal data on your behalf.
\nDayparting: Limiting or raising bids by hour or day of week to match when your audience is most likely to act.
\nDeal ID: A code passed in a bid request that tells the seller's system which agreed terms, such as price and inventory pool, apply to the bid.
\nDecision gate: A scheduled checkpoint where the team compares evidence to written criteria and decides to pass, adjust or stop.
\nDeduplicated reach: The count of unique people or households exposed to a campaign across all channels, with repeat counts removed.
\nDeduplication key: A unique identifier, such as an order number, that lets systems recognize and discard a repeated report of the same event.
\nDeferred deep link: A link that remembers the intended screen and opens it on the first launch after install, so the new user does not land on a generic home screen.
\nDemand capture: Advertising that collects people who already want something and are actively looking for it.
\nDemand creation: Advertising that puts a need, problem or brand name into the mind of a person who did not hold it before.
\nDemand-side platform (DSP): Software an advertiser or agency uses to bid on impressions across many exchanges against targeting rules and a budget.
\nDeterministic matching: Linking records to one person using an identifier the person provided and confirmed, such as a login or email address.
\nDiminishing returns: The pattern in which each additional dollar in a channel buys fewer results than the one before.
\nDirect path: A route in which the seller holds an account the publisher itself controls, shown as DIRECT in the publisher's ads.txt.
\nDiscrepancy window: The period during which a buyer may dispute the seller's delivery or billing figures.
\nDispute ratio: The share of card transactions that customers contest, which payment processors watch when deciding whether a program is acceptable.
\nDouble opt-in: A two-step confirmation in which the person must verify their contact detail, usually by clicking a link or entering a code, before entering your follow-up system.
\nDynamic ad insertion: Stitching an ad into an audio file at the moment of play or download, so messages can change after publication.
\nEngaged session: A visit that meets a threshold you define in advance, such as a minimum time on page plus a second page view or a meaningful scroll.
\nEngagement pod: A group of accounts that agree to like and comment on one another's posts to inflate engagement figures.
\nEstimated action rate: The platform's prediction of how likely a specific viewer is to complete the optimization event after seeing a specific ad.
\nEvent schema: The agreed list of event names and attached properties that keeps recorded behavior consistent across tools and teams.
\nFalse positive: A result that looks like a real effect but is produced by chance. The significance threshold caps how often this happens.
\nFAST channel: A free ad-supported streaming television channel, usually a continuous lineup of programming with commercial breaks, delivered through an app or a device home screen.
\nFirst-party data: Information a business collects directly from its own customers and visitors through its own properties and relationships.
\nFirst-price auction: An auction in which the winning bidder pays exactly the amount it bid, which makes bid shading a rational strategy.
\nFlight: A defined period with a start date, an end date and a budget for a campaign or set of campaigns.
\nFlighting: Scheduling ads to run in bursts with gaps in between, instead of at a steady level, so spend matches demand or events.
\nFloor: The minimum budget a channel keeps regardless of short-term results, set to protect defense and learning.
\nFloor price: The minimum bid a seller will accept for an impression, placement or deal, below which no bid can clear.
\nFrequency: The average number of times a person sees your ad in a given period, which climbs quickly in small audiences and signals fatigue.
\nFrequency cap: A setting that limits how many times one person can see an ad within a time window.
\nFrequency distribution: A breakdown of how many reached people received one exposure, two, three, and so on, which shows the spread that an average hides.
\nGeneral invalid traffic (GIVT): Non-human or non-qualifying traffic that routine filtering, such as bot lists and data center lists, can identify.
\nGeo holdout: A test in which ads are paused in selected regions while matched regions continue, so total results can be compared.
\nGeographic holdout: A test that withholds a campaign from matched places to estimate what the campaign added.
\nGlobal Privacy Control: A browser signal that tells sites a person refuses the sale or sharing of their data for advertising.
\nGrain: The level of detail at which a figure is calculated, such as per order, per campaign or per day.
\nGTIN: Global trade item number, the barcode identifier a manufacturer assigns to a product, used to match your item to known products.
\nGuardrail metric: A secondary number with a pre-agreed limit that must not deteriorate while the main objective is pursued.
\nHallucination: A confident, fluent model output that is false or unsupported by any source.
\nHashing: A one-way transformation that turns an identifier into a fixed string, useful for matching but not a guarantee of anonymity.
\nHeader bidding: A publisher technique that lets several buyers bid on a slot at the same time before the page's ad server makes its decision.
\nHoldout: A deliberately unexposed group, whether a set of markets or a slice of an audience, used as the comparison for measuring incrementality.
\nHoldout group: A randomly selected slice of an audience that is deliberately not shown ads, used as the baseline to measure what the ads caused.
\nHoldout test: A measurement method that pauses a campaign in matched regions or periods and compares results against where it kept running, to estimate how much of its output was truly caused by the ads.
\nHook: The opening beat of an ad, usually the first one to three seconds, designed to stop the scroll and signal relevance.
\nHop count: The number of companies that handle an impression between the publisher and the buyer, read from the nodes in the SupplyChain object.
\nHost-read: An ad spoken by the show's own host in a personal voice, usually from a short brief.
\nHuman-in-the-loop: A process in which a named person must approve model output before it is used.
\nHurdle rate: The minimum return a marginal dollar must earn, usually set in gross profit terms, to justify spending it.
\nIAB Content Taxonomy: A hierarchical list of content categories maintained by the IAB Tech Lab, used so buyers, sellers and vendors describe pages in the same terms.
\nIdentity resolution: The process of linking devices, accounts or households so that exposures can be attributed to the same person or home.
\nImpression: One instance of an ad being counted as served to a page or app, whether or not anyone saw it.
\nIncremental sales: Sales that happened because of the ad and would not have happened without it, estimated through a holdout or similar test.
\nIncrementality: The portion of an outcome that exists because of the advertising and would not have occurred otherwise.
\nInformation scent: The set of cues that tells a person they are getting closer to what they wanted, so they keep going instead of leaving.
\nInsertion order: A written purchase document that lists the campaign, the dates, the rates, the volumes, and the terms of sale between a buyer and a media seller.
\nIntent tier: A group of search queries that share the same reason for searching and therefore the same job, budget logic and measurement method in the account.
\nInvalid traffic: Ad requests or views produced by automated scripts, hijacked devices, or other non-human sources .
\nLargest Contentful Paint: A speed measurement of how long the biggest visible element on the first screen takes to appear.
\nLast-click credit: A rule that gives the full commission to the final partner link clicked before purchase, regardless of who introduced the shopper.
\nLead form ad: A paid social format that opens a form inside the app, often prefilled from the person's profile, so they can submit without visiting a website.
\nLearning phase: The early exploration period when a delivery system gathers data and results are volatile, usually restarted by significant edits.
\nLeast-privilege access: The practice of giving a person only the system permissions needed for the assigned task.
\nLong tail: The mass of specific, low volume searches that together bring meaningful traffic at typically lower competition.
\nLookback window: The period before a conversion during which a touch is eligible to receive credit.
\nMade-for-advertising: Inventory built primarily to carry ads and collect ad revenue, with little genuine audience or content value behind it.
\nMade-for-advertising inventory: Sites or apps built mainly to host ad slots, often with thin content and heavy ad density, regardless of whether visitors are human.
\nMake-good: A contract remedy in which the seller provides credit or replacement inventory when promised delivery is not met.
\nMakegood: A remedy offered when delivery falls short, usually additional placements or impressions of equal value, or a credit.
\nMarginal cost: The extra cost of the next unit of outcome gained by adding spend, which is usually higher than the average cost.
\nMarginal return: The extra result earned by the next block of spend, as opposed to the average result of all spend so far.
\nMarketing mix modeling: A statistical method that relates total sales over time to spend by channel and outside factors, without tracking individuals.
\nMaterial connection: Any relationship between an endorser and a brand that could affect how viewers weigh the endorsement, such as payment, free product, commissions or personal ties.
\nMessage match: The degree to which an ad and its landing page communicate the same promise, offer, and tone.
\nMetric dictionary: A versioned document that defines each business measure by name, formula, grain, owner and exclusions.
\nMinimum detectable effect: The smallest lift a test can reliably tell apart from random noise.
\nNative ad unit: An ad assembled from parts such as headline, image and description, then displayed in a frame styled by the publisher's page template.
\nNegative keyword: A word or phrase that blocks your ad from showing on any query containing it, used to route traffic between campaigns and to remove waste.
\nNovelty effect: A short-lived lift in response caused by a version being new, which fades as visitors get used to it.
\nOffline conversion upload: A method of sending outcomes that happen after the click, such as booked appointments or sales, back to an ad platform so it can optimize toward them.
\nOpen-weight model: A model whose trained parameters are published for download, though the license may still restrict how you use it.
\nOptimization event: The specific conversion action an ad platform is told to maximize when it decides whom to show ads to.
\nOpt-in: A clear, voluntary action by a person agreeing to receive messages from a brand.
\nOrganic to paid loop: A repeating cycle in which unpaid posts are tested, the winners are promoted with paid media, and the people they attract produce the next round of content.
\nOwned audience: The people you can contact directly, with their permission, without paying a platform for access.
\nPacing: The method a buying system uses to spread a budget across time, which affects which auctions you enter and what you pay.
\nPacing ratio: Actual cumulative spend divided by planned cumulative spend to date, where a value near 1.00 means on plan.
\nPanel: A sample of consenting people or households whose viewing is measured and then scaled to represent a larger population.
\nPayback period: The time it takes for contribution from an acquired customer to repay the cost of acquiring them.
\nPermissive license: A license such as Apache 2.0 or MIT that allows commercial use with minimal conditions.
\nPod: A small group that owns an outcome from start to finish, such as paid acquisition or creative production.
\nPosterior: The range of plausible answers a Bayesian model reports after combining its prior with your data.
\nPrecision and recall: Precision is the share of pages labeled relevant that really are; recall is the share of truly relevant pages the system managed to find.
\nPre-registration: Writing and dating the primary outcome, window and decision rule before launch so results cannot be reinterpreted afterward.
\nPrimary category: The single business type a listing is mainly classified under, which strongly influences the searches it can appear in.
\nPrincipal-based buying: A practice in which an agency purchases media in its own name and resells it to the client, often at a markup that should be disclosed.
\nPrincipal-based trading: A model in which a vendor buys media in its own name and resells it to the client at a price the vendor sets, keeping the difference.
\nPrior: The belief a Bayesian model holds before it sees your data, which can be set from experiments or outside knowledge.
\nPrior express written consent: A documented agreement, signed or given electronically, in which a person clearly authorizes marketing texts to a specific number from a specific sender.
\nPrivate marketplace: An invitation-only auction in which a publisher selects the buyers who may bid, usually offering better inventory and clearer terms.
\nPrivate marketplace (PMP): An invitation-only auction in which a publisher offers selected inventory to chosen buyers, often with fewer intermediaries.
\nProduct feed: A structured file listing every product with its identifier, title, price, availability, and attributes.
\nProgrammatic guaranteed: A deal that reserves a set volume of inventory at a fixed price and executes it through ad technology.
\nPublisher share: The fraction of an advertiser's spend that ultimately reaches the publisher as revenue after every fee and margin is taken.
\nQA gate: A checkpoint at which work stops until a second person approves it.
\nQualified lead: A lead that meets a written definition of fit and intent, agreed by marketing and sales, such as a verified contact in your service area who wants service soon.
\nQualifying copy: Ad language that discourages clicks from people the business cannot serve, lowering wasted spend.
\nQuality score: A platform rating of ad relevance and expected performance that influences price and position.
\nQuery drift: The gradual movement of triggered searches away from the intent you planned for, common with broad and automated matching.
\nQuiet hours: The time window, based on the recipient's local time, outside of which marketing messages should not be sent.
\nRACI: A responsibility chart that names who is Responsible, Accountable, Consulted, and Informed for each task.
\nReach: The count of unique people who saw an ad at least once in a given period.
\nRecency window: The number of days after an interaction during which a person remains eligible to see retargeting ads.
\nReconciliation: The routine of tying reported advertising figures back to the general ledger and documenting any remaining gap.
\nReconciliation ratio: Platform-reported conversions divided by orders in your own records, tracked over time to expose tracking faults.
\nReseller: A seller that passes along inventory it does not own, adding a hop and usually a margin between buyer and publisher.
\nResponse curve: A line showing how outcomes change as spend changes, usually rising quickly at first and flattening as spend grows.
\nRetail media: Advertising sold inside a retailer's own sites and apps, and sometimes on outside sites, using the retailer's shopper data.
\nRetail media network: An advertising business run by a retailer that sells placements on its own site, app and sometimes off-property channels to brands that stock its products.
\nRetainer: A fixed monthly fee for an agreed scope of work, regardless of how much media is bought.
\nRetargeting: Showing ads to people who have already interacted with a brand, such as by visiting its site or abandoning a cart.
\nRetention: The share of a cohort that returns to the app on a given later day, such as day seven or day 30.
\nRights fee: The payment that gives a brand permission to use a property's name, marks and access under a written agreement.
\nRunbook: A short written procedure for a recurring task, detailed enough that a new teammate can follow it alone.
\nSafe zone: The part of a vertical frame that stays clear of platform buttons, names, and captions, where key content should sit.
\nSaturation: The point at which added spend in a channel produces smaller and smaller gains, shown as a curve that flattens.
\nSchema: A written definition of each event's required properties, their types, and their allowed values.
\nSearch intent: The underlying goal behind a query, such as learning, comparing or buying, which determines the kind of page that deserves to rank.
\nSearch term report: The platform report listing the actual queries people typed that triggered your ads, as opposed to the keywords you chose to bid on.
\nSeasonal index: A multiplier showing how far a given week or month typically runs above or below the average period.
\nSecond-price auction: An auction in which the winner pays a price set by the next-highest bid, usually plus a small increment, so bidding true value is safe.
\nSeed audience: The source list or set of converters handed to a platform so its model can find similar people.
\nSeeding: Contributing useful posts, answers, or assets to a community so that members value the brand's presence before any promotion.
\nSeller identifier: The account code that names a specific seller inside an advertising system, used to match ads.txt lines against sellers.json entries.
\nsellers.json: A public file on an ad platform's domain listing the sellers it represents and whether each is a publisher or an intermediary.
\nSender reputation: The track record mailbox providers assign to your sending domain and network address based on complaints, bounces, engagement and consistency.
\nSequential testing: A test design that allows planned early looks at the data while keeping the overall false positive risk at the stated level.
\nServer-side collection: A setup in which the browser sends events to an endpoint you control, which then decides what to forward to each platform.
\nService-area business: A business that goes to customers rather than receiving them at a public address.
\nSingle owner: The one named person accountable for a workstream's deliverables, as distinct from the people who contribute to it.
\nSingle source of truth: One agreed location and set of rules from which every report draws its numbers, so reports agree because they read the same records.
\nSkip offset: The delay before a skip control appears in a skippable ad, set in the tag and by the platform.
\nSOP: A standard operating procedure, a short written set of steps for a recurring task, with an owner and a checker.
\nSophisticated invalid traffic (SIVT): Invalid traffic designed to look human, needing multi-point analysis to detect, such as hijacked devices or falsified identity.
\nSpam trap: An address kept by mailbox providers or blocklist operators to catch senders who collect or buy addresses carelessly.
\nSpeed to lead: The time between a person submitting a form and a human or automated first response.
\nSponsored content: An article, video or guide produced by or for an advertiser and published in a labeled section of a publisher's property.
\nSponsored listing: A paid product placement inside search or category results that resembles organic results and is usually billed per click.
\nStage gate: A written list of conditions that must be met before a channel is added or expanded.
\nStatement of work: A written document that defines deliverables, timelines, fees, and acceptance criteria for an engagement.
\nStatistical power: The probability that a test detects a real effect of a stated size. Eighty percent is a common planning target.
\nStopping rule: A pre-written condition, such as a minimum number of conversions or a date, that decides when a test may be read.
\nStructured data: Labeled markup on a web page, such as schema.org, that tells machines what each piece of content means.
\nSupply chain object: A record carried with a bid that lists each company that handled the ad opportunity between publisher and buyer.
\nSupply path: The chain of companies between the buyer and the screen, including platforms, resellers, exchanges, and the publisher or app owner.
\nSupply path optimization: The practice of choosing the shortest, most direct and most transparent routes to publishers, rather than buying through every available intermediary.
\nSupplyChain object: A record inside the bid request listing each company that handled the impression between the publisher and the buyer.
\nSupply-side platform (SSP): Software a publisher uses to offer ad space to many buyers at once, earning a fee on each winning bid.
\nSuppression audience: A list of people deliberately excluded from seeing a given campaign, such as existing customers.
\nSuppression list: A list of people excluded from an audience, such as recent buyers or people who withdrew consent.
\nSuppression record: A stored instruction not to contact or target a person, kept even after other data is deleted.
\nTag manager: A tool that stores and controls which tracking snippets run, under which conditions, with version history and preview.
\nTest velocity: The number of tests a team completes and documents in a set period, usually a quarter.
\nTotal value: The combined score an ad platform uses to rank competing ads, built from the bid, the estimated action rate and an ad quality measure.
\nTracking: The recording of ad events and user actions, such as impressions, clicks and purchases, without any claim about what caused them.
\nUsage rights: The brand's licensed permission to reuse creator content, defined by channel, duration and territory, usually priced separately from the creative fee.
\nUser-generated content (UGC): Photos, videos, reviews, or posts created by customers or members rather than by the brand or its agencies.
\nVanity link: A short, memorable web address spoken in an ad that redirects to a tracked page.
\nVAST: The IAB Tech Lab standard for describing a video ad to a player, including the media file, tracking and click behavior.
\nViewable impression: An impression that met a defined standard for how much of the player was on screen and for how long.
\nView-through conversion: A conversion credited to an ad because the household saw it and later acted, without proof that the ad caused the action.
\nWarm-up: The practice of raising sending volume gradually on a new domain or address so mailbox providers learn that your mail is wanted.
\nWhitelisting: A permission setup in which a creator lets a brand run paid ads through the creator's handle, with the brand controlling budget and targeting. Also called partnership ads or creator licensing.
\nWhitelisting rights: Permission for a brand to run an ad using a creator's or customer's account handle and content, usually limited by time and channel.
\nWorking media: The portion of a budget that pays for the placement of the ad itself, as distinct from fees. Definitions vary by vendor, so always ask for the one in use.
\nWrapper: A VAST tag that does not contain a creative itself but redirects the player to another tag, often through a vendor chain.
\nZero-click search: A search where the person gets what they need from the results page itself and never visits a site.
\nDan Kost is the Founder and Chairman of Dakdan Worldwide and the CEO of Sports Media Inc. He is a U.S. Army Veteran based in Colorado. Sports Media Inc. operates platforms including SportsMedia.net, MySportsMedia.com, OOHSports.com, and RAMMww.com. The Brand Manager's Guide Series is published by Dakdan LLC Media and Publishing.
\nContact: info@dakdan.com
\nA2P (application-to-person): Chapter 39
\nA/A test: Chapter 50
\nActivation: Chapter 40
\nAd pod: Chapter 24
\nAdstock: Chapter 46
\nads.txt: Chapter 28
\nAffiliate network: Chapter 36
\nAnswer layer: Chapter 60
\nAssets (extensions): Chapter 9
\nAssisted conversion: Chapter 31
\nAttributed sales: Chapter 37
\nAttribution window: Chapter 1
\nBaseline: Chapter 61
\nBid shading: Chapter 3
\nBlackout: Chapter 25
\nBrand defense: Chapter 10
\nBrand safety: Chapter 34
\nBrand suitability: Chapter 32
\nBranded query: Chapter 12
\nBranded search: Chapter 5
\nBreak-even return on ad spend: Chapter 37
\nB-roll: Chapter 16
\nBurn-in captions: Chapter 16
\nCalibration: Chapter 46
\nCannibalization: Chapter 7
\nCanonical tag: Chapter 12
\nCareer ladder: Chapter 54
\nCeiling: Chapter 56
\nChannel mix: Chapter 57
\nClassification: Chapter 59
\nClawback: Chapter 36
\nClean room: Chapter 15
\nClick injection: Chapter 38
\nClose variant: Chapter 8
\nCohort: Chapter 38
\nCommission-based fee: Chapter 52
\nCompanion unit: Chapter 21
\nCompletion rate: Chapter 21
\nConcentration risk: Chapter 60
\nConcept: Chapter 49
\nConquesting: Chapter 10
\nConsent ledger: Chapter 42
\nContent recommendation widget: Chapter 31
\nContextual targeting: Chapter 15
\nContingency reserve: Chapter 58
\nContribution margin: Chapter 6
\nControl: Chapter 49
\nCookie stuffing: Chapter 36
\nCore web vitals: Chapter 9
\nCounterfactual: Chapter 1
\nCrawl budget: Chapter 12
\nCreative fatigue: Chapter 16
\nCuration: Chapter 23
\nData clean room: Chapter 27
\nData freshness: Chapter 47
\nData layer: Chapter 43
\nData map: Chapter 42
\nData minimization: Chapter 41
\nData processing addendum: Chapter 42
\nDayparting: Chapter 58
\nDeal ID: Chapter 23
\nDecision gate: Chapter 61
\nDeduplicated reach: Chapter 27
\nDeduplication key: Chapter 43
\nDeferred deep link: Chapter 38
\nDemand capture: Chapter 5
\nDemand creation: Chapter 5
\nDemand-side platform (DSP): Chapter 4
\nDeterministic matching: Chapter 41
\nDiminishing returns: Chapter 57
\nDirect path: Chapter 29
\nDiscrepancy window: Chapter 53
\nDispute ratio: Chapter 55
\nDouble opt-in: Chapter 19
\nDynamic ad insertion: Chapter 24
\nEngaged session: Chapter 31
\nEngagement pod: Chapter 17
\nEstimated action rate: Chapter 14
\nEvent schema: Chapter 41
\nFalse positive: Chapter 51
\nFAST channel: Chapter 22
\nFirst-party data: Chapter 15
\nFirst-price auction: Chapter 3
\nFlight: Chapter 48
\nFlighting: Chapter 58
\nFloor: Chapter 56
\nFloor price: Chapter 3
\nFrequency: Chapter 14
\nFrequency cap: Chapter 20
\nFrequency distribution: Chapter 27
\nGeneral invalid traffic (GIVT): Chapter 34
\nGeo holdout: Chapter 13
\nGeographic holdout: Chapter 24
\nGlobal Privacy Control: Chapter 42
\nGrain: Chapter 47
\nGTIN: Chapter 11
\nGuardrail metric: Chapter 6
\nHallucination: Chapter 59
\nHashing: Chapter 15
\nHeader bidding: Chapter 28
\nHoldout: Chapter 1
\nHoldout group: Chapter 33
\nHoldout test: Chapter 7
\nHook: Chapter 16
\nHop count: Chapter 29
\nHost-read: Chapter 24
\nHuman-in-the-loop: Chapter 59
\nHurdle rate: Chapter 56
\nIAB Content Taxonomy: Chapter 32
\nIdentity resolution: Chapter 27
\nImpression: Chapter 28
\nIncremental sales: Chapter 37
\nIncrementality: Chapter 1
\nInformation scent: Chapter 9
\nInsertion order: Chapter 53
\nIntent tier: Chapter 7
\nInvalid traffic: Chapter 4
\nLargest Contentful Paint: Chapter 50
\nLast-click credit: Chapter 36
\nLead form ad: Chapter 19
\nLearning phase: Chapter 14
\nLeast-privilege access: Chapter 54
\nLong tail: Chapter 8
\nLookback window: Chapter 44
\nMade-for-advertising: Chapter 22
\nMade-for-advertising inventory: Chapter 34
\nMake-good: Chapter 25
\nMakegood: Chapter 53
\nMarginal cost: Chapter 48
\nMarginal return: Chapter 56
\nMarketing mix modeling: Chapter 44
\nMaterial connection: Chapter 17
\nMessage match: Chapter 9
\nMetric dictionary: Chapter 47
\nMinimum detectable effect: Chapter 45
\nNative ad unit: Chapter 31
\nNegative keyword: Chapter 7
\nNovelty effect: Chapter 50
\nOffline conversion upload: Chapter 19
\nOpen-weight model: Chapter 59
\nOptimization event: Chapter 6
\nOpt-in: Chapter 40
\nOrganic to paid loop: Chapter 18
\nOwned audience: Chapter 60
\nPacing: Chapter 3
\nPacing ratio: Chapter 48
\nPanel: Chapter 27
\nPayback period: Chapter 6
\nPermissive license: Chapter 59
\nPod: Chapter 54
\nPosterior: Chapter 46
\nPrecision and recall: Chapter 32
\nPre-registration: Chapter 45
\nPrimary category: Chapter 11
\nPrincipal-based buying: Chapter 52
\nPrincipal-based trading: Chapter 53
\nPrior: Chapter 46
\nPrior express written consent: Chapter 39
\nPrivate marketplace: Chapter 2
\nPrivate marketplace (PMP): Chapter 4
\nProduct feed: Chapter 11
\nProgrammatic guaranteed: Chapter 25
\nPublisher share: Chapter 2
\nQA gate: Chapter 54
\nQualified lead: Chapter 19
\nQualifying copy: Chapter 9
\nQuality score: Chapter 10
\nQuery drift: Chapter 8
\nQuiet hours: Chapter 39
\nRACI: Chapter 52
\nReach: Chapter 20
\nRecency window: Chapter 33
\nReconciliation: Chapter 47
\nReconciliation ratio: Chapter 13
\nReseller: Chapter 2
\nResponse curve: Chapter 48
\nRetail media: Chapter 60
\nRetail media network: Chapter 37
\nRetainer: Chapter 52
\nRetargeting: Chapter 33
\nRetention: Chapter 38
\nRights fee: Chapter 40
\nRunbook: Chapter 62
\nSafe zone: Chapter 16
\nSaturation: Chapter 46
\nSchema: Chapter 43
\nSearch intent: Chapter 12
\nSearch term report: Chapter 7
\nSeasonal index: Chapter 48
\nSecond-price auction: Chapter 3
\nSeed audience: Chapter 15
\nSeeding: Chapter 18
\nSeller identifier: Chapter 29
\nsellers.json: Chapter 23
\nSender reputation: Chapter 35
\nSequential testing: Chapter 51
\nServer-side collection: Chapter 43
\nService-area business: Chapter 11
\nSingle owner: Chapter 61
\nSingle source of truth: Chapter 47
\nSkip offset: Chapter 21
\nSOP: Chapter 54
\nSophisticated invalid traffic (SIVT): Chapter 34
\nSpam trap: Chapter 35
\nSpeed to lead: Chapter 19
\nSponsored content: Chapter 31
\nSponsored listing: Chapter 37
\nStage gate: Chapter 57
\nStatement of work: Chapter 52
\nStatistical power: Chapter 51
\nStopping rule: Chapter 61
\nStructured data: Chapter 11
\nSupply chain object: Chapter 25
\nSupply path: Chapter 22
\nSupply path optimization: Chapter 2
\nSupplyChain object: Chapter 23
\nSupply-side platform (SSP): Chapter 4
\nSuppression audience: Chapter 20
\nSuppression list: Chapter 33
\nSuppression record: Chapter 42
\nTag manager: Chapter 43
\nTest velocity: Chapter 62
\nTotal value: Chapter 14
\nTracking: Chapter 1
\nUsage rights: Chapter 17
\nUser-generated content (UGC): Chapter 18
\nVanity link: Chapter 24
\nVAST: Chapter 26
\nViewable impression: Chapter 21
\nView-through conversion: Chapter 22
\nWarm-up: Chapter 35
\nWhitelisting: Chapter 17
\nWhitelisting rights: Chapter 18
\nWorking media: Chapter 2
\nWrapper: Chapter 21
\nZero-click search: Chapter 12
"}]}]}