Retargeting incrementality: calculate lift beyond attributed ROAS
Work through a retargeting holdout test, normalize group sizes, calculate incremental CPA, and interpret the result using margin and uncertainty.
Define the advertising effect you want to measure
Retargeting reaches people who already know something about the product. Some would purchase without another ad. The experiment should therefore ask a narrow question: does this particular retargeting campaign change purchase probability among the selected eligible audience? The study concerns a campaign, audience window, and period rather than the value of every advertising activity the company operates.
Specify the business event before launch: a paid order, accepted lead, or another meaningful outcome. Decide how returns and repeat purchases will be handled. Comparing all orders in one group with first purchases in the other would make the result uninterpretable. Save the standard attribution report separately so you can explain why its numbers differ from the experiment.
Assign comparable groups before exposure
Where available, use a suitable experimental mechanism to randomly assign eligible users to treatment and holdout. The treatment group is eligible for the tested ads; the control group is withheld from that exposure. Other communications should remain as comparable as possible. If control users receive the same retargeting message through another campaign, the contrast between groups becomes diluted.
Simply comparing people who saw an ad with people who did not is insufficient. Auction delivery and user behavior influence exposure, and the groups may already differ in purchase intent. Google documents a controlled Conversion Lift approach, but it is not available to every account. Confirm current eligibility and the ability to maintain the intended holdout before committing resources.
Normalize outcomes to the group size
Consider an illustrative test with 5,000 people and 150 buyers in treatment, compared with 5,000 people and 125 buyers in control. Conversion rates are 3% and 2.5%. The absolute difference is 0.5 percentage points, while relative lift is 20%. Estimated incremental buyers within the treatment group are 25, assuming the comparison design is valid.
If group sizes differ, compare rates first and scale the control rate to the treatment population. A control group of 10,000 people with 250 buyers still converts at 2.5%. Subtracting 250 from 150 would compare incompatible population sizes. Expected treatment purchases without the tested exposure remain 5,000 multiplied by 2.5%, or 125.
| Metric | Illustrative calculation | Result |
|---|---|---|
| Treatment conversion rate | 150 / 5,000 | 3% |
| Control conversion rate | 125 / 5,000 | 2.5% |
| Absolute lift | 3% − 2.5% | 0.5 percentage points |
| Relative lift | 0.5 / 2.5 | 20% |
| Incremental buyers | 5,000 × (3% − 2.5%) | 25 |
Translate incremental outcomes into economics
Continue the illustrative example with $5,000 USD spent on the tested ads. Dividing that spend by 25 estimated incremental buyers gives an incremental CPA of $200 USD. This may be much higher than the platform’s attributed CPA because some attributed purchases would have happened without those ads. The difference does not automatically mean that either reporting system is broken.
Suppose each additional order contributes $150 USD after variable operating costs but before advertising. Estimated incremental contribution is $3,750 USD; after $5,000 USD of ad spend, the net result is negative $1,250 USD. This assumes equal contribution per order. If purchase mix differs, use the relevant values and margins. Add future repeat-purchase value only with explicit, supported assumptions.
Evaluate uncertainty before deciding
The example demonstrates arithmetic, not statistical significance. Before launch, define the smallest effect that would change the budget decision and the sample required to detect it. Allow enough time for the conversion cycle. If purchases often complete two weeks after the first visit, yesterday’s audience has not had the same opportunity to convert as an older cohort.
Review the confidence intervals or other uncertainty measures provided by the study method. Do not end the test on the first favorable day. When estimated lift is negative or close to zero, dividing spend by that estimate may produce an unhelpful or unstable CPA. Explain the absence of a reliable positive effect instead of presenting a dramatic number as a precise finding.
Keep the decision within the experiment’s scope
The result applies to the tested audience, offer, delivery conditions, and period. If incremental contribution does not cover spend, a shorter retargeting window, different message, or narrower audience may be worth a separate experiment. A positive result also does not justify unlimited scaling: the people reached by the next budget increase may behave differently.
The site’s unit-economics calculator can help examine order contribution using supplied inputs. It does not create a holdout group or establish lift significance. Common errors include comparing unequal groups using raw totals, ignoring conversion delay, and describing attributed ROAS as a causal effect. Preserve the design, exclusions, and uncertainty alongside the decision so future teams can use the finding responsibly.
Put it into practice
Frequently asked questions
Can we measure lift by switching retargeting off for one week?
A before-and-after comparison is vulnerable to seasonality, promotions, and demand changes. It may provide directional information, but it is not equivalent to a randomized holdout. Geographic or time-based designs need their own comparability checks.
What if the account has too few conversions?
Check the method’s requirements, feasible duration, and smallest commercially meaningful effect. Do not promise precision the data cannot support. Improving event quality and margin estimates may be more useful than launching an underpowered study.