incrementality-and-mmm
Incrementality & MMM
Answer the question attribution cannot: would this conversion have happened anyway?
Attribution models tell you which touchpoint preceded a conversion. They cannot tell you which touchpoint caused it, because they only observe the world where the ad ran. The gap between those two things is largest exactly where reported performance looks best — retargeting and branded search routinely show excellent attributed ROAS and much weaker incremental ROAS, because they intercept people who were already going to convert.
The only way to close that gap is to withhold the spend from a comparable group and measure the difference. That's uncomfortable — you're deliberately forgoing revenue to learn something — which is why most teams never do it, and why the ones that do usually reallocate budget afterwards.
The failure mode this skill exists to prevent is the underpowered test. A holdout that finds "no significant difference" when it was never large enough to detect a meaningful effect is not evidence the channel doesn't work. It gets reported as though it were, and budget moves on the strength of it.
When to use
The user is designing or reading a holdout, geo, or lift test — or questioning whether attributed conversions are real. For on-site A/B tests of a page or flow, use ab-test-significance; the statistics overlap but the design question is different. For reporting results upward, use marketing-report.