A customer sees an ad, later searches for the company, and buys. An attribution model can assign credit to one or more observed interactions. It cannot directly reveal whether that customer would have bought without the advertising.
That counterfactual is the core of incrementality. Keeping the two questions separate helps teams use everyday reports without turning them into stronger causal claims than their design supports.
Ask the question before selecting the report
“Which campaigns received conversion credit?” is a reporting question. “What happened after we changed the budget?” is an operational observation. “How many additional sales did the advertising produce?” is a causal question.
These questions can inform one another, but they require different evidence. A campaign report can be useful for detecting delivery changes, reviewing creative response, or locating an unexpected concentration of attributed value. It does not become a controlled experiment because the dashboard uses a sophisticated attribution model.
Write the decision in one sentence before choosing the measurement method. That prevents the team from selecting whichever report produces the most convenient answer.
Understand why selection matters
People exposed to advertising may differ from those not exposed. They may already have stronger purchase intent, know the brand, or be selected because their behavior predicts conversion. A comparison between exposed and unexposed customers can therefore mix advertising effects with pre-existing differences.
The 2019 Marketing Science study comparing advertising measurement approaches found that observational methods often differed from randomized experiment results in the Facebook campaigns studied. That is evidence about the measurement challenge in those settings, not proof that every observational analysis is useless or that every current campaign has the same bias.
The practical lesson is to examine how a method addresses selection before calling its estimate causal.
Use an evidence map
| Evidence | Useful for | Main limit to keep visible |
|---|---|---|
| Platform attribution | Campaign operations and assigned conversion credit | Credit is not automatically causal lift |
| Business time series | Overall revenue and spend context | Many things can change together |
| Randomized experiment | A defined treatment comparison | Scope, power, compliance, and external validity |
| Geographic experiment | Regional media changes and outcomes | Region comparability, spillover, and design quality |
| Marketing mix model | Broader allocation analysis under model assumptions | Identification, data quality, priors, and uncertainty |
No row is a universal winner for every decision. The right method depends on feasibility, consequence, and the specific effect the business wants to estimate.
Interpret a simple example carefully
Suppose an illustrative campaign receives credit for 100 purchases. A suitable experiment estimates that the campaign produced 20 additional purchases during its test conditions, with an uncertainty interval around that estimate.
The two figures are not necessarily contradictory. They describe different quantities. The attributed count reflects the reporting rules; the experiment estimates a counterfactual difference for the tested population and period.
Do not automatically apply the ratio between them to every future campaign. A different audience, spend level, season, or offer may have a different incremental effect. The experiment is most useful when its scope remains attached to the result.
Treat branded demand as a question to test
A brand search often indicates prior awareness, but that does not by itself establish whether an ad adds value. The Performance Max brand audit helps describe coverage and attribution before a causal test is considered.
Historical research such as the eBay paid-search field experiments demonstrates why existing intent can matter to measured returns. Those experiments concerned a particular business and period. They should motivate careful evaluation, not a blanket rule that all brand advertising is ineffective.
A business may still choose coverage for operational reasons while acknowledging that its incremental effect remains uncertain.
Plan experiments around meaningful decisions
Define the treatment, eligible population, primary outcome, observation window, and smallest effect that would change the budget decision. Check whether the available volume can plausibly resolve that effect.
Use the geographic holdout planning guide when regional variation is the practical design. For other experiments, use a supported setup that matches the campaign and question.
Predefine how the team will handle implementation failures and early operational problems. A test in which the intended treatment was not delivered needs a different interpretation from a clean test with an uncertain outcome.
Use models with their assumptions attached
Marketing mix modeling can connect media and business outcomes over time, but a fitted curve is not self-validating. The MMM readiness guide covers the data and identification questions to resolve before using recommendations for allocation.
Experiments can help inform models, but the relationship is not a mechanical conversion. Differences in time, geography, spend, and the effect being estimated need to be considered.
Keep uncertainty visible in budget scenarios. If a recommendation changes dramatically across plausible assumptions, the business has learned where additional evidence would be valuable.
Make the next budget decision proportionate
Teams often need to act before perfect evidence exists. Use the best available information, state the uncertainty, and keep the exposure appropriate to the confidence and business constraints.
For routine operations, attribution and business context may support a bounded adjustment. For a major channel reallocation, a more deliberate experiment or model review may be worth the effort.
The reporting language should match the method: credited, observed, estimated, or causally supported under specified conditions. That precision makes measurement more useful because everyone can see what the number can and cannot justify.
Work through the lift arithmetic
The incremental ROAS calculator scales a control group to treatment size before subtracting revenue and media cost. It is a simple worksheet for compatible experimental groups; it does not correct selection bias or replace a study-specific uncertainty model.