The Paid Media Anti-Pattern Glossary

Most paid-media anti-patterns are not exotic. They are familiar habits that survive because they make the dashboard look better than the business. When we audit a program and find reported performance that does not square with what the P&L is doing, the gap almost always reduces to one of two failure modes. Either the program is counting demand that would have arrived without it, or it is hiding the side effects of one channel on another. Every anti-pattern below is a specific instance of one of those two mistakes.

The pattern is structural, not moral. Vendors, platforms, and in-house teams are all incentivized to report numbers that flatter the work. The credit-assignment rules they default to are not neutral; they are built to make the spend look productive. A senior measurement lead needs a working vocabulary for these patterns so they can be named, detected, and corrected before they harden into the operating model.

Use this glossary as a diagnostic. Audit your current program against each anti-pattern and flag any pattern that explains a material share of reported results. Design tests (geo holdouts, a control group withheld from exposure so true lift can be measured; audience holdouts; creative holdouts) to measure the incremental contribution of the suspect spend. Incrementality is the share of conversions that would not have happened without the spend, and it is the only question that separates growing a business from crediting demand the business already owned. The remedies below mostly point to the same destination: design measurement before designing media, and rely on triangulated reads across marketing mix modeling (MMM), multi-touch attribution (MTA), and incrementality rather than any single tool's verdict.

Attribution and measurement inflation

Last-click misattribution. Crediting the final touch before conversion with the full value of the sale. The model is wrong on its face because most purchases involve more than one exposure, and the final touch is disproportionately the channel best positioned to harvest already-decided buyers: branded search, retargeting, direct. Those channels intercept intent the brand created elsewhere; the last-click model calls that intent a paid-media win. Detect it by checking what share of reported conversions are credited to channels whose CPMs and click-through rates suggest they are reaching the same users repeatedly within a 24-hour window. The remedy is to triangulate channel decisions against geo-lift or holdout reads, not against last-click reports.

View-through credit at face value. View-through credit is credit assigned to an ad that was rendered but not clicked. Counting an impression-only exposure as a contributing touch and assigning revenue credit on that basis conflates delivery with influence. A view-through means the ad rendered in a viewport; it does not mean the user noticed it, and the platform that served the impression is the same party grading whether it worked. Detect it by reporting paid performance with view-throughs included and excluded side by side: if the channel's reported ROAS halves when view-throughs are stripped out, the channel is being graded on impressions, not influence. The remedy is to set a view-through window short enough to be defensible (24 to 48 hours), require a click for any cross-channel credit, and treat view-through volume as a media-quality signal rather than a revenue claim.

Walled gardens grading their own homework. A walled garden is a platform that holds its own data and grades its own measurement. Letting Meta, Google, TikTok, or Amazon report their own incremental contribution and treating that report as the source of truth places the judge and the defendant in the same seat. Each platform uses a different credit model, so the sum of platform-reported revenue routinely exceeds total actual revenue. Detect it by summing platform-reported conversions across all paid channels for a given week and comparing the total to backend orders: an overlap of more than 15 to 20 percent is the floor of normal double-counting. The remedy is to run independent incrementality tests on the largest channels and to use platform reporting only for tactical decisions within a channel, never for cross-channel allocation.

Modeled conversions as ground truth. Platform-modeled conversions estimate what tracking lost to iOS privacy changes and cookie deprecation would have shown. Data-driven attribution (DDA), which weighs observed touches by their statistical association with conversion, faces the same limitation: the model is calibrated against the platform's own data and is not an independent read. Detect it by asking what fraction of reported conversions are modeled versus observed and whether that ratio has moved materially in the last six months without a corresponding change in real performance. The remedy is to grade campaigns on observed conversions for in-channel optimization and reserve modeled volumes for context, not for budget decisions.

Demand harvesting disguised as performance

Retargeting harvesting. Running retargeting against site visitors and counting the resulting conversions as paid-media performance. Most of those users arrived through other channels and would have returned on their own; the retargeting ad is collecting credit for a purchase the brand already won. The pattern is especially common when retargeting pools are large relative to the brand's prospecting budget, because the cheap harvested conversions inflate the program's blended efficiency. Detect it by holding back retargeting in a randomly chosen 20 percent of the audience for four weeks and measuring whether total conversions in the holdout drop or stay flat. The remedy is to size retargeting as a small line that captures genuine attention loss, not as a primary acquisition channel, and to exclude retargeting revenue from acquisition ROAS.

Branded-search cannibalization. Buying paid search on your own brand terms and counting the clicks as paid-media success. The same users would have clicked the organic listing for free, so the spend is harvesting traffic the brand already owned. The effect is invisible in the channel report but shows up in a geo test: toggle the branded-search line off in a holdout geo for two weeks and watch whether total branded clicks drop or whether organic absorbs the volume. The remedy is to bid on the brand defensively when competitors are actively poaching, but exclude branded-search revenue from paid-media ROAS calculations.

Audience overlap with organic and CRM. Targeting paid social or display at audiences that already receive your email, follow your social accounts, or visit your site weekly, and counting their conversions as paid acquisition. The paid impression rides on top of a relationship the brand already had; the customer acquisition cost (CAC) credited to the paid channel is actually the cost of touching someone who was already in the pipeline. Detect it by suppressing existing customers and engaged subscribers from prospecting audiences for one campaign cycle and watching whether reported acquisition volume holds. The remedy is to require strict exclusion lists on every prospecting campaign and to report prospecting performance separately from retention.

Lookalike incest. A lookalike audience is an audience modeled on the traits of existing customers or converters. Building lookalike audiences off recent purchasers, then building the next lookalike off the buyers acquired through that lookalike, and so on, is how a brand slowly turns a prospecting tool into a demand-harvesting loop. The seed audience narrows on each generation and the platform converges on a tighter slice of the brand's existing demand pool rather than expanding into new buyers. Detect it by measuring the share of net-new customers (no prior brand interaction in the last 12 months) inside each lookalike cohort and watching whether that share trends down over successive iterations. The remedy is to rebuild seed audiences against high-value or hard-to-reach customer segments rather than recent converters, and to grade lookalikes on new-customer share, not on ROAS.

Promo-stacked conversions. Running paid media into a sitewide discount or free-shipping promotion and crediting the channel with the resulting conversion volume. The promotion is doing the work; the paid impression is the delivery mechanism. Stacking both in the same reporting window produces an efficiency number that belongs to the discount, not the media. Detect it by comparing channel ROAS during promo and non-promo weeks and looking for a step-change that is larger than the underlying spend or creative change can explain. The remedy is to report channel performance on a promo-adjusted basis and to insist that incrementality tests run outside of promotion windows.

Vanity metrics that flatter the dashboard

Vanity ROAS. Reporting blended or channel-level ROAS without distinguishing new from returning customers, branded from non-branded, or promo-stacked from clean weeks. The headline number aggregates harvested demand with incremental demand and produces a figure that moves with promo calendar and customer-mix shifts rather than with media quality. A high ROAS on a retargeting-heavy program tells you the harvest rate is good; it does not tell you the program is growing the business. Detect it by recomputing ROAS on net-new customers only and watching how much of the original number was returning-buyer revenue. The remedy is to report incremental ROAS (iROAS), the ROAS calculated only on the revenue the campaign actually caused, on a new-customer basis as the primary efficiency metric and treat blended ROAS as context.

Channel-level efficiency without MER. Optimizing each channel against its own ROAS target while ignoring marketing efficiency ratio (MER), which is total revenue divided by total marketing spend, at the program level. Channels can each hit their individual target while the program as a whole degrades, because the channel-level numbers are double-counting overlapping demand. MER is the only metric that cannot be gamed by shuffling credit between channels. Detect it by tracking MER weekly alongside channel ROAS and watching whether the two diverge over a quarter. The remedy is to set MER as the controlling efficiency metric for the program and treat channel ROAS as a contributor, not as a goal.

Last-month CAC on cohorts that have not closed. Reporting customer acquisition cost (CAC) based on conversions that landed in the reporting month, against spend that landed in the same month, when the actual sales cycle or refund window is longer than that. The numerator and the denominator are not measuring the same cohort, so the CAC is structurally biased low during growth and structurally biased high during pullback. Detect it by recomputing CAC on a cohort basis, spend in week N with conversions from that spend tracked over the full sales cycle, and comparing it to the month-over-month number. The remedy is to standardize CAC on a cohort-completion basis for any business with a sales cycle longer than 30 days and to report a leading indicator separately.

Org and operating-model failures

Siloed channel teams grading themselves. Channel leads for paid social, paid search, and connected TV (CTV) own both the spend and the report on their channel's performance, with no independent measurement function reconciling the numbers. The incentive is to produce a flattering channel narrative every cycle, and the numbers comply. This is not bad faith; it is the predictable output of a structure where the grader's budget depends on the grade. Detect it by checking whether the person who reports channel performance to leadership is the same person whose budget depends on that performance, and whether any holdout or causal test has been run by a party other than that channel lead in the last 12 months. The remedy is to assign incrementality and cross-channel reporting to a measurement function that does not own channel budgets.

Agency-graded performance without a holdout. Letting the agency that buys the media also produce the performance report, with no client-side holdout test against which the agency's claim can be checked. The agency's commercial interest is in defending the program; their reporting reflects that interest, often without anyone acting in bad faith. Detect it by asking when the last client-commissioned incrementality test on the agency's largest channel was run and who ran it. The remedy is to commission an independent holdout on the largest one or two channels at least once a year and to make that read the reference point for the quarterly business review.

Creative volume without creative testing. Producing 30 to 50 ad variants per cycle, shipping them all, and treating the spend distribution that emerges from auction dynamics as the test result. The platform optimizes for short-term in-platform conversions, which means the surviving creative is the one that harvests existing demand most efficiently, not the one that builds new demand. The creative learning compounds the measurement problem: the team learns what wins at harvesting and calls it insight. Detect it by checking whether any creative test in the last quarter held variables constant long enough to read a winner on incremental, not platform-graded, performance. The remedy is to run structured creative tests against an off-platform incrementality signal and to treat platform-optimized creative as a tactical tool, not as a strategic insight.

Test design and statistical anti-patterns

Peeking at running tests and calling early winners. Checking a holdout or A/B test daily and declaring a winner the moment the difference crosses a significance threshold, before the pre-registered duration is up. Repeated peeking inflates the false-positive rate because any random run of favorable days will eventually cross the line. A test that looks like a winner at week two often regresses toward the mean by week six. Detect it by asking whether the test had a pre-registered end date and whether the call to stop or scale was made before that date. The remedy is to pre-register test duration and decision rules, and to use geo-lift designs that explicitly account for sequential testing.

Underpowered tests called as conclusive. Running a holdout on a channel that does not have enough conversion volume to detect the lift size the business cares about, and then reporting the null result as evidence the channel does not work, or the positive result as evidence it does. With insufficient power, the test simply cannot distinguish real lift from noise. The minimum detectable effect (MDE) is the smallest lift the design can pick out of the noise, and if the lift you need to detect sits below the MDE, the test would need an unrealistic effect size or runtime to read out. Detect it by computing the MDE for the test as designed and comparing it to the lift size that would actually change the decision. The remedy is to run a power calculation before the test, extend duration or pool geos if power is insufficient, and refuse to draw conclusions from underpowered runs.

MMM without external calibration. Running a marketing mix model and treating the output as truth without checking the channel coefficients against incrementality reads from the same period. MMM is a structural tool with known identification problems, and an uncalibrated model can produce confidently wrong allocation recommendations that survive because no one stress-tests them. Calibration means using a geo-lift result to lock one channel's contribution in place so the rest of the model has to fit around it. Detect it by asking whether the most recent MMM run was reconciled with at least one independent incrementality result on a major channel. The remedy is to calibrate MMM coefficients against documented incrementality formulas and to discard model outputs that contradict experimental reads without an explanation.

Geo-lift on unmatched markets. Running a geo-holdout where the treatment and control markets differ on demand drivers, including population, income, seasonal pattern, and competitive presence, that are not balanced out by the design. The lift estimate then conflates the media effect with the market difference. A synthetic control is a weighted blend of similar geos that approximates one treatment market's pre-test behavior, and it can help when no natural pair exists, but done badly it hides a power problem behind statistical complexity. Detect it by comparing pre-period trends in the treatment and control geos: if they do not move together cleanly for at least eight weeks before the test starts, the markets are not matched. The remedy is to use synthetic-control or matched-market designs that explicitly construct a control from a weighted combination of similar geos rather than picking a single comparable market on intuition.

Anti-patterns compound. A program running three of these at once is not measuring growth; it is measuring its own credit-assignment rules. Last-click misattribution feeds retargeting harvesting feeds vanity ROAS, and the channel team responsible for grading all three reports a number that confirms every decision already made. The dashboard glows. The P&L does not move.

The shift is to design measurement before designing media, and to treat the question as primary, not the tool. Cross-channel allocation belongs to MMM; tactical optimization belongs to MTA; the causal question of whether a channel does real work belongs to incrementality. The right deployment is all three in their lanes, with geo-lift tests sized for the decisions they are meant to inform. The anti-patterns in this glossary survive in the absence of that discipline. Naming them is the first step to designing them out.