The best experiment review starts before the dashboard opens. Everyone should know the original hypothesis, decision options and quality checks. This prevents a polished chart from replacing disciplined judgment.

Lead with validity

Confirm that randomisation, exposure and metric collection behaved as expected. Check sample-ratio mismatch, unusual missingness, concurrent releases and major calendar events. If validity is compromised, say so before discussing uplift.

Separate evidence from recommendation

Present the estimated effect and uncertainty first. Then explain the recommendation using business value, implementation cost, customer risk and reversibility. A statistically uncertain result is not automatically a failure; it may still narrow the plausible outcomes enough to guide a low-risk decision.

Record the decision

Close with one of four outcomes: roll out, iterate and retest, stop, or gather more evidence. Assign an owner and date. Store the hypothesis, data-quality notes, result and decision together so future teams can reuse the learning.

Review question: what will we do differently because of this evidence?