Run a randomized recovery experiment

Choose one question before starting

1. Create a conditional flow and open its experiment settings. Choose copy, timing or offer as the single changing dimension.
2. Set the enrollment start and end, a purchase observation window of 1–14 days and a minimum completed sample per variant of at least 30. The minimum is a guardrail, not a statistical power calculation.
3. Edit variant B while keeping the other dimensions unchanged. The API rejects multiple-dimension changes and identical variants.
4. Save and enable deliberately. Eligible shoppers receive an independent 50/50 assignment that stays fixed for their email within this store and experiment.
5. Wait until enrollment has ended and the full outcome window has elapsed. Review uncertainty and sample size before starting a separate follow-up experiment.

Use purchases, not opens, as the outcome

The primary outcome is any recorded paid order by the assigned email within its observation window, including purchases without a recovery click. Actual payment timestamps are used when supplied by connector 0.8; a delayed or repeated hook is not treated as a new purchase. Each shopper contributes at most one positive outcome. The purchase-rate metric is not reduced by a later refund; refunds are separately visible in financial reports.

Only fully observed shoppers enter purchase-rate estimates. The report shows assigned and mature counts, 95% Wilson intervals for each rate and a Newcombe interval for B minus A. It never declares an early winner or promotes a variant automatically. Insufficient samples and overlapping uncertainty produce an inconclusive result.

This compares A with B. It does not measure incremental lift against sending no email. Enrolled experiment designs are frozen; use a new experiment ID for a new question. Assignment records use a store-scoped pseudonymous key and retain the recorded purchase outcome even after the short-lived order ledger expires.