Incrementality
Incremental lift = conversions in the exposed group − conversions in the holdout, scaled for group size
| Exposed group | The audience or regions where the ads run as normal. |
| Holdout | A matched audience or set of regions where the ads are withheld. This supplies the baseline: what happens without the spend. A geography split is a geo-lift test; an audience split is a holdout test. |
| Incremental ROAS | Incremental revenue ÷ ad spend: the causal version of ROAS. |
Worked example
Example numbers. Half of what the platform claimed would have happened anyway; the experiment, not the dashboard, is what showed it.
What is a good incrementality result?
There is no universal good lift percentage. What a result means depends on the channel's role (demand-capture placements like branded search and retargeting repeatedly test far less incremental than prospecting, because they intercept buyers already on the way), on spend level (incrementality falls as spend scales into colder audiences), and above all on your margin structure.
The honest gate is profit: incremental revenue is only worth buying when it clears your margin-based threshold, which is the same 1 ÷ contribution margin logic as break-even ROAS, applied to measured lift instead of claimed revenue. A 2.0x incremental ROAS is strong on a 70% contribution margin and underwater on a 40% one.
Incrementality vs related methods
| Method | What it is | How it differs |
|---|---|---|
| Attribution model | A credit rule over observed touchpoints | Correlational; incrementality is an experiment that measures causation. |
| Geo-lift test | Ads withheld in matched regions | A method of measuring incrementality; privacy-proof, no user tracking. |
| Holdout test | Ads withheld from a matched audience slice | The other main method; needs platform support for clean splits. |
| Marketing mix modeling (MMM) | Statistical estimation from aggregate history | Always-on causal estimate; well-run teams calibrate it against incrementality tests. |
Common mistakes
- Reading platform-attributed conversions as incremental. A pixel touch is not causation; much of that claimed revenue arrives anyway.
- Running underpowered tests. A holdout too small, or a window too short, cannot separate lift from noise, and a null result gets misread as zero incrementality.
- Contaminating the control. Ads leaking into holdout regions, or audience overlap between groups, shrinks the measured gap and understates lift.
- Judging lift on revenue. Incremental revenue at a thin contribution margin can still lose money per incremental dollar.
- Extrapolating one test forever. Incrementality moves with spend level, season and creative; retest when any of them changes materially.
Frequently asked questions
Inside Blufire, S5 Acquisition & Attribution puts attribution models side by side so reported numbers are labelled as reported, the starting point for judging what each channel really causes.
Updated July 2026