Why incrementality answers a question attribution cannot

A purchase after a click is a sequence

Attribution links a recorded outcome to a touchpoint using a rule or model. That is useful for describing journeys and operating marketing reports. It does not, by itself, establish that the touchpoint caused the purchase. People with strong existing interest may be especially likely both to click and to buy.

Incrementality asks a counterfactual question: how would outcomes differ without the intervention? You cannot observe both possibilities for the same person at the same moment. A well-designed randomized comparison estimates that difference across groups rather than assigning all post-click purchases to a campaign.

Read the denominator before the result

Consider a hypothetical experiment with 1,000 eligible people assigned to each group. Suppose 60 in the treatment group and 50 in the holdout purchase within the same predefined window. The observed difference is one percentage point, or ten purchases in this example. It is not sixty incremental purchases.

This arithmetic is not a significance test. Sampling uncertainty, assignment quality, missing outcomes and interference still matter. Keep all eligible randomized units in their assigned groups when using an intention-to-treat analysis, including those who never saw the intervention. Do not silently replace assignment with exposure after the fact.

Design the comparison before launching

Define eligibility, assignment unit, primary outcome, observation window and stopping rule before starting. Check whether visitors can cross between groups or whether campaigns overlap. Google’s published geo-experiment method illustrates that randomized comparisons can also use geographic units; that does not make a geographic experiment appropriate for every small site.

Report costs separately from revenue and refunds. More incremental orders may still produce negative economic contribution if incentives are expensive. If the sample is too small or instrumentation is unreliable, the honest result is uncertainty, not a winning percentage.

Keep both tools in their place

Attribution remains useful for journey description and operational reporting. Incrementality is the more relevant question when deciding whether an intervention created value. Neither excuses poor data quality. Sellwise preserves this distinction in its product direction; current public illustrations and internal tests are not evidence of real merchant uplift or available production experimentation.