Secondary metrics that must not regress (e.g. return rate, refunds, page load).

Check the result holds within sub-groups (e.g. Mobile vs Desktop). Enter control & variation per segment.

Removes pre-experiment noise to tighten the confidence interval, and can reveal a real effect a raw test misses. Runs on Control vs the best variation.

Checks whether the win reverses when two segments are combined, a classic trap from imbalanced traffic.