Name the question before naming the metric
Which observed interaction receives credit for this order is an allocation question. How many extra orders did this campaign create is a causal question. A report can answer the first without establishing the second.
That distinction matters when evaluating branded search, remarketing or a message to an already engaged customer. A purchase can be attributed to an interaction even when some customers would have purchased without it. The size of that overlap needs evidence rather than an assumption.
Design a comparison that can support the claim
A suitable experiment compares an intervention with a credible control. Google describes Conversion Lift as an incrementality experiment and notes that access is not available for every account. Tool availability is only one part of designing a useful test.
Agree the outcome, eligible population, duration and practical constraints before launch. In fashion, consider promotions, stock availability and return timing. Do not switch the primary outcome after seeing the result just because another metric looks better.
Read the uncertainty
An experiment may produce an inconclusive result. That is different from proving that the intervention had no effect. The quality of the design, available volume and uncertainty around the estimate all matter to the interpretation.
For an observed before-and-after change, state what changed alongside the intervention. Seasonality, product availability and other campaigns may offer alternative explanations. Use careful wording until the evidence supports a stronger causal claim.
A lifecycle campaign with a holdout
A team wants to understand the additional purchases from a post-purchase message. It defines an eligible cohort and a suitable holdout before sending, then evaluates a consistent outcome over an agreed period. The attributed orders in the email tool remain a separate view.
Turn the guide into a useful review
- Write the causal question in plain language.
- Choose an appropriate control and outcome.
- Account for business timing and operational constraints.
- Report uncertainty and avoid adding overlapping agent effects.
Reference material
Platform guidance checked 3 October 2026. Examples and working checklists are Faccelerate editorial illustrations.
Bring the question.
Keep the context.
Map the relevant sources and specialist workflow for your team.
Explore your stack
