Key takeaways
- Attribution allocates credit; incrementality estimates outcomes caused by the intervention relative to a control.
- Write the hypothesis, primary outcome, eligible population, assignment, exclusions, duration, and decision rule before launch.
- Choose user, geography, or another design only when randomization, contamination, power, and operational feasibility are credible.
- An underpowered or inconclusive study is not proof of zero effect—and a positive platform result is not automatically portable.
01
Ask the question ROAS cannot answer by itself
Attributed conversions are outcomes assigned to an ad interaction under a platform or analytics rule. Incremental conversions are the difference between what occurred with an intervention and a credible estimate of what would have occurred without it. Loyal buyers, branded demand, seasonality, and other channels can all appear inside attributed ROAS.
Write one decision first: maintain, expand, reduce, or stop a defined intervention for an eligible population. Name the spend, campaign, audience, market, offer, and period. “Does paid social work?” is too broad to test and too vague to act on.
02
Pre-register the operating definition
Before results are visible, record the hypothesis, primary metric, guardrails, unit of assignment, eligibility, exclusions, sample-size or feasibility assessment, duration, conversion lag, analysis method, and decision threshold. This limits the temptation to choose whichever segment or window looks most persuasive afterward.
Google describes lift as treatment outcomes minus control outcomes and offers user-based and geography-based studies subject to eligibility. Platform tooling can make assignment and measurement practical, but the team still owns the business outcome and the assumptions around spillover, interference, and data quality.
Swipe to compare every column
| Design choice | Useful when | Threat to validity |
|---|---|---|
| User holdout | Eligible users can be randomized | Identity gaps or cross-device contamination |
| Geo holdout | Markets are comparable and separable | Spillover, local shocks, weak market count |
| Platform experiment | Native assignment is available | Result applies only to tested setup and population |
| Before/after | Exploratory monitoring only | Seasonality and concurrent changes confound the estimate |
03
Let the test cover the real conversion cycle
A study that ends before delayed conversions arrive measures a different outcome from the one the business cares about. Google’s guidance notes that duration should reflect conversion lag and that many lift studies run longer than fourteen days. Promotions, inventory changes, outages, pricing, and other campaigns should be logged before and during the test.
Do not peek each morning and stop at the first attractive result. Follow the planned duration and analysis unless a safety, budget, or data-quality condition requires stopping. Record that reason. Frequent unplanned checking raises the chance of mistaking noise for evidence.
04
Read the interval, the economics, and the limitations
Report the estimated lift with uncertainty, treatment and control outcomes, cost, population, period, exclusions, and deviations. An estimate that spans commercially negative and positive effects is inconclusive for that decision. It should not be polished into either “the ads did nothing” or “the test proved success.”
Translate the range into unit economics and capacity. Then decide whether to repeat with more power, change the intervention, narrow the question, or stop. Store the study beside creative, audience, landing-page, and offer context so the next team does not treat a local result as a universal law.
Primary sources and further reading
Use the source material to validate details against your own context and current platform configuration.
- Google Ads Help: About conversion lift
- Google Ads Help: Lift studies
- Google Ads Help: Set up a user-based conversion lift study
- Systematic review of validity trade-offs in entrepreneurship experiments
This field note follows the XenGrowth editorial policy: primary sources where available, visible limitations, material review dates, and no invented first-hand experience.
Stay with the problem



