Lift

The incremental improvement a treatment cohort produces over control, the number an experiment actually exists to measure.

Lift is the incremental improvement a treatment cohort produces over a control cohort. It's a different number from attribution ("how many people who saw variant B converted") because some of those conversions would have happened anyway. Lift is what's left after you subtract that baseline out.

Measuring lift honestly requires a real holdout: a group that saw nothing, so there's something to subtract. Without one, every number a campaign reports is attribution dressed up as lift.

A reported lift number is also only as trustworthy as the randomization behind it. Before reading lift off a dashboard, it's worth checking for sample ratio mismatch first: a broken split between cohorts can produce a lift figure that looks real but is actually measuring a plumbing defect, not a treatment effect.