Incrementality tests: what AdOps needs to get right before launch

A campaign receives credit for 500 sales. The business wants to know how many of those sales would have happened without the campaign. An attribution report cannot answer that question on its own.

Incrementality is the additional outcome caused by an intervention, such as advertising. A study estimates what would have happened without that intervention, then compares it with what happened under the test conditions.

The IAB and IAB Europe guidelines for incremental measurement describe several approaches, including experiments and modeled comparisons. In September 2026, Google also announced broader availability of Meridian GeoX, its library for geographic experiments. More accessible tools make operational discipline more valuable, not optional.

Start with the decision and the outcome

Write down the question before selecting the method. “Should we increase this campaign’s budget?” may require a different test from “Did this channel produce any additional sales?”

Define the outcome precisely: completed purchases, qualified leads, revenue after returns, or another business event. Record when it is measured and how late results will be handled.

Agree on the minimum effect worth acting on. Have the analyst estimate the sample size and duration needed to detect that effect with the chosen method. A small test that cannot distinguish the expected effect from noise may consume time without answering the question.

Protect the assignment

In a randomized holdout test, eligible units are assigned to treatment or control before the outcome is known. The unit might be a person, household, or another supported entity. Geographic studies have their own assignment and analysis requirements.

AdOps needs to know which units can receive the tested campaign and how control exclusions are enforced. Do not replace the assigned groups afterward with “people who saw an ad” and “people who did not.” Actual exposure can depend on behavior and eligibility, making those groups different in ways unrelated to the campaign’s effect.

Document which other campaigns can reach both groups. A control group may continue to receive normal marketing; the study must make clear which additional treatment is being tested.

Understand the arithmetic without overstating it

Consider an illustrative randomized test with 10,000 eligible people in each group:

Group	Purchases	Purchase rate
Treatment	600	6%
Control	500	5%

The observed difference is one percentage point. Relative to the control rate, that is a 20% lift: (6% − 5%) ÷ 5%. Applied to 10,000 treatment units, the point estimate is 100 additional purchases.

Those are descriptive calculations, not a complete study result. The analyst still needs to assess uncertainty, assignment integrity, missing outcomes, and the assumptions of the test. “Twenty percent lift” without an interval and method can sound much more conclusive than the evidence warrants.

If the treatment cost $5,000, dividing by that point estimate gives $50 per estimated incremental purchase. It is not the same as the campaign’s attributed CPA.

Control changes during the study

Maintain a change log for budget, targeting, creative, tracking, promotions, and site incidents. Some changes are necessary, but they should not disappear from the interpretation.

A checkout outage affecting one market, a promotion introduced halfway through, or an exclusion removed by a bulk edit can alter the comparison. Raise the issue with the analyst promptly. Avoid silently restarting the clock or removing inconvenient days.

Agree beforehand on stopping rules and review points. Repeatedly checking for a favorable result and ending the test at that moment can distort the conclusion.

Hand over an interpretable result

The final readout should include the decision being tested, group assignment, outcome definition, study dates, cost basis, effect estimate, uncertainty, operational exceptions, and limitations.

If the result is inconclusive, say what decision remains unresolved. An estimate compatible with both little effect and useful lift does not justify declaring either success or failure.

Keep the test plan beside the campaign launch checklist. A well-run study should leave the next decision clearer, even when the result is that more evidence is needed.