The Paid Media Attribution Models Reference

Most paid-media measurement conversations turn on a question that sounds reasonable but quietly substitutes for the one that matters. The reasonable-sounding question is which touchpoint should get credit for a conversion. The question that actually decides whether a budget is working is whether the spend grew the business. Those are not the same question, and the gap between them is where most attribution practice goes wrong.

This reference walks through the main families of attribution models in use today, what each one is built to do, and where each one stops being measurement. The aim is not to dismiss any single approach but to put each tool in its honest place. Some are useful for steering creative or bids inside a single channel. Some are useful for long-horizon planning. Only one family of methods answers the budget question directly, and that is the one most teams underuse.

Attribution answers the wrong question

Attribution, in everyday paid-media usage, means the practice of assigning credit for a conversion across the touchpoints that preceded it. A user sees a Meta ad, clicks a Google search ad two days later, opens an email a week after that, and converts. Attribution decides how to divide that conversion among the three touches. That division is convenient for dashboards and useful for in-platform optimization, but it is fundamentally a story told after the fact about people who already converted.

The question that decides budgets is different: did this spend grow the business? That is a question about the counterfactual, the version of the world where the campaign never ran. Attribution models do not see the counterfactual; they only see the exposed side. They can tell you how to split credit among the touches you observed, but they cannot tell you which of those conversions would have happened anyway. Treating credit assignment as a budget signal is how teams end up scaling channels that platforms grade generously and cutting channels that produce real lift but show up weakly in the report.

The question that decides whether a media program is working is not which touch gets the conversion but whether the spend grew the business.

— QRY Measurement Reference

Heuristic models: rules pretending to be measurement

Heuristic models split credit by a fixed rule. The familiar names cover most of the field. Last-click gives credit entirely to the final touchpoint before conversion. First-click gives it entirely to the first. Linear divides credit evenly across every touch on the path. Time-decay weights more recent touches more heavily. U-shaped and W-shaped models hand outsized credit to the first and last touches, sometimes with a bump for a mid-funnel event.

These models are useful for narrow diagnostic work. If you want a fast, transparent split that everyone on the team can audit, a heuristic rule will do that. Last-click in particular is still the right default for in-channel bid optimization, because the platform needs a single deterministic signal to learn against. Inside one channel, with one team owning the levers, heuristic credit is good enough to drive bid and creative decisions.

The failure mode is using heuristic credit to make cross-channel budget decisions. The rule itself encodes a bias that does not match how media actually works. Last-click systematically overweights bottom-funnel channels that catch already-decided buyers, because those are the touches sitting closest to the conversion. First-click overweights brand and prospecting at the expense of channels that actually closed. None of the rules can see whether a touch contributed to the conversion or was merely present on the path. They are arithmetic, not measurement.

What heuristic models miss is the counterfactual. They cannot tell you whether the conversion would have happened without that touch, because they have no view of the version of the user journey where the touch never occurred. That is not a flaw to be fixed with a better rule; it is a feature of the entire family.

Platform and vendor statistical models: better math, same blind spot

Statistical attribution models replace the fixed rule with a model fitted to observed conversion paths. Google data-driven attribution (DDA) is the best-known example inside an ad platform. Meta attribution and the various flavors of multi-touch attribution (MTA) sold by independent vendors sit in the same family. The mechanic is to compare paths that converted with paths that did not, then estimate how much each touchpoint contributed to the probability of conversion. The output is a per-touch credit split that is statistically derived rather than rule-derived.

This is genuinely better math than a heuristic rule, and it produces more defensible numbers inside a single platform. Google DDA, used inside Google Ads on Google inventory, is a reasonable optimization layer. Meta attribution, used inside Meta on Meta inventory, is the same. View-through credit, the credit given to an ad that was rendered but not clicked, can be modeled rather than asserted by a fixed rule. For within-channel optimization, statistical attribution is the strongest tool in this family.

The blind spot is the same one the heuristic models have. Platform statistical models only see their own walls. Google DDA cannot model the Meta touches in a user journey because Google does not see them, and Meta cannot model the Google touches for the same reason. Both increasingly rely on modeled conversions as user-level signals disappear under privacy changes, which means the input to the model is itself a modeled estimate. Vendor MTA stitches across platforms but inherits the visibility gaps of every platform it integrates with, plus an additional layer of identity stitching that is rarely as clean as the vendor implies.

What platform and vendor statistical models miss is, again, the counterfactual. The math got better; the question did not change. A more sophisticated credit split is still a credit split. It tells you which observed touches correlate with conversion, not which spend caused the conversion to happen at all. Treat these models as the right tool for within-channel optimization and as the wrong tool for cross-channel budget allocation.

Marketing mix modeling: the right tool for the wrong question?

Marketing mix modeling (MMM) sits at a different level of abstraction. Rather than tracking individual user paths, MMM fits an econometric model to aggregate weekly or daily data. Total spend by channel, total revenue, price, promotional activity, seasonality, and macro factors all go in. The model estimates a response curve for each channel: how revenue moves as spend on that channel changes, controlling for everything else. The output is a contribution number per channel and, in the better implementations, a saturation curve that suggests where additional spend stops paying back.

MMM is the right tool for cross-channel allocation and long-horizon planning. It is privacy-robust because it works on aggregate data rather than user-level signals. It captures channels that user-level attribution cannot see, including offline media, brand campaigns, and ambient channels with delayed effects. For the questions of how much to spend in total and how to divide that total across channels over a planning horizon, MMM is the most appropriate measurement tool in common use.

Where MMM weakens is the gap between modeling a counterfactual and constructing one. The model estimates what would have happened without each channel by reading variance in the historical data, which means the answer is only as good as the variance the data contains. If a channel ran at a flat spend level for the full window, the model has very little signal to work with on that channel. If two channels moved together, the model cannot reliably separate them. The output is an estimate, and like every estimate it is sensitive to the modeling choices and assumptions behind it.

What MMM misses, on its own, is a ground truth to anchor it. The same dataset will yield different channel contributions under different modeling specifications. Without an external anchor, MMM becomes a sophisticated form of storytelling: internally consistent, defensible to a degree, but not measurement in the strict sense. The fix is to calibrate MMM against real lift tests on at least one channel and then re-fit. The full taxonomy of how MMM compares to MTA and to designed-in incrementality lives in MMM vs MTA vs incrementality.

Designed-in incrementality: what actually works

Designed-in incrementality is the family of methods that answers the budget question directly rather than modeling it after the fact. Incrementality, in the technical sense, is the share of conversions that would not have happened without the spend. The only way to measure incrementality honestly is to construct a counterfactual, the version of the world where the campaign did not run, and then compare it to what actually happened. That construction has to be planned into the campaign before launch. It cannot be reverse-engineered from observed conversions.

Geo holdouts. A geo holdout is a control group of regions where the channel is withheld so its true effect can be measured. The campaign runs at full weight in a set of treatment markets and is pulled entirely from a matched set of controls for the test window. The gap between the two groups, scaled across the treatment population, is the lift. Because spend is the only thing that systematically differs between the groups, the gap is causal rather than correlational. The design details, including matched-market selection, power, and contamination risks, are covered in geo-lift testing explained.

Ghost ads and public service announcement (PSA) tests. A ghost ad or PSA test is a placebo-ad design that randomizes exposure within a platform. The platform randomly assigns part of the target audience to see the campaign and the rest to see a neutral placebo ad or no ad at all. Because assignment is randomized at the user level, the comparison is tight and the test reads out quickly. The trade-off is independence: the platform runs the experiment and grades the result, which is why ghost ads are best for rapid creative or audience iteration rather than for cross-channel budget validation.

Calibrated MMM. MMM becomes measurement rather than storytelling when at least one channel in the model is anchored by a real lift test. A geo-lift result on the largest paid channel constrains the MMM during fitting, which forces the rest of the model to reconcile against a known truth. The hybrid is stronger than either tool alone: MMM provides the cross-channel view, and the lift test provides the ground truth that makes the contributions credible.

Post-purchase surveys. Asking buyers how they heard about the brand adds a stated lens that platform attribution cannot capture. Surveys overweight memorable channels and underweight ambient ones, so they are not a standalone source of truth, but they are a useful counter-bias to platform-reported performance. Triangulating survey results against MMM and lift tests is cheap and surfaces channels that the other tools systematically under- or over-credit.

These methods share a structural property that none of the other model families have. They build the counterfactual into the test rather than estimating it after the fact. A geo holdout sees the world without the spend because the spend was withheld. A ghost ad sees it because exposure was randomized. A calibrated MMM inherits a counterfactual from the test that anchors it. The result is causal in the strict sense: the gap between the exposed and the unexposed groups is the effect of the spend, not a story about credit on the people who happened to convert.

The output of a clean incrementality test is two numbers worth more than every dashboard read against them. Lift, the percentage gap between treatment and control. And incremental ROAS (iROAS), the ROAS calculated only on the revenue the campaign actually caused. The math, the sample-size formulas, and the minimum detectable effect calculations are in the incrementality formulas reference.

Put each model in its honest place

Every model in the field has a job it does well and a job it does badly. Heuristic and platform attribution are the right tools for within-channel optimization and the wrong tools for cross-channel budgets. MMM is the right tool for cross-channel allocation and long-horizon planning and the wrong tool to use uncalibrated. Designed-in incrementality is the only family that answers the budget question directly, and it is structurally the most expensive to set up because it requires real money withheld from a real control. The cost is what makes the answer worth more than the cheaper ones.

The error most teams make is not picking the wrong model. It is using a single model to answer every question, usually the model that came free with the ad platform. Platform-attributed ROAS is a fine optimization signal and a poor budget signal. Treating it as both is how budgets end up allocated by the vendors selling the media. The companion glossaries cover the recurring patterns that follow from that confusion, including the specific failure modes that platform-attributed reporting tends to produce.

The vocabulary used here is collected in the paid media measurement glossary, and the recurring anti-patterns are catalogued in the paid media anti-pattern glossary.

The practical move is to reserve holdouts, plan a geo-lift on the largest paid channel, wire a brief survey into checkout, and calibrate MMM before the next campaign launches. None of those steps are difficult; they are operational rather than technical. The reason most programs do not do them is that platform attribution already produces a number every Monday morning, and a number that arrives on time tends to win against a number that arrives correctly. The fix is to stop treating the on-time number as truth.