Choosing a winner

Before you call it#

• You reached the sample size you planned (see Reading results).
• The test ran for at least one full week.
• The visitor split between variations roughly matches your weights. A big mismatch usually means a bug in your snippet.
• Results for Production only (GET /api/experiments/{id}/results?environment=prod) agree with the dashboard, which combines all environments.

Complete the experiment#

Choose Complete on the experiment page. Completing is final: no new visitors are assigned, the SDK returns assigned: false to everyone, and the experiment becomes read only. Results stay available.

Roll out the winner#

1. 1

   Make the change permanent

   Update your page so the winning version is the default content.

2. 2

   Remove the experiment code

   Delete the ABTest.run callback for this experiment. Keep sdk.js and init if other experiments use them.

3. 3

   Record what you learned

   Write the outcome in the experiment description, so the next test builds on it.

A gradual rollout today#

There are no feature flags yet. If you want to keep showing the winner through the SDK for a while, leave the experiment running and keep the winning code in your callback, but remember the control share still sees the old version. Weights cannot be changed after creation, so you cannot move a running test to 100%.

When the control wins#

Complete the experiment, remove the variant code and keep your page as it is. A losing variant is a result too: you avoided shipping a change that would have hurt.