Geo-lift test
A geo-lift test measures the incremental sales a marketing channel causes by turning its spend up or down in a randomly chosen set of geographic regions while holding a comparable set unchanged, then comparing the two. Ecommerce operators use it to prove whether a channel actually drives sales rather than just co-occurring with them.
| Term | What it covers |
|---|---|
| Test regions | Markets where spend was deliberately changed - turned off, or scaled up. |
| Control regions | Comparable markets left untouched, used to predict what the test regions would otherwise have done. |
| Counterfactual | The expected test-region sales with no change, estimated from the control group's behavior. |
| Incremental sales | Observed minus counterfactual; divided by the spend change it gives incremental ROAS or CAC. |
Lift % = (observed − expected) ÷ expected. The result is causal because the region split was randomized, not chosen after the fact.
Worked example
Example numbers. The pixel reported roughly 3.0x for the same spend; the test shows the channel is incremental, but at half the platform's claim.
What is a good geo-lift test?
A good geo-lift test is not one with a high lift number - it is one whose result you can act on. Two things decide that. First, the design: test and control regions have to track each other closely before the test and cover enough markets to reach statistical significance, or the "lift" is just noise. Second, the economics: an incremental ROAS is only good if it clears your break-even ROAS. A 1.5x that clears break-even is a keep; a 3.0x that does not is a cut - the threshold depends entirely on your margin, not on the headline multiple. Judge the test on significance first, then on whether the incremental return beats the number where the channel actually pays for itself.
Geo-lift test vs related metrics
| Metric | What it measures | How it differs from a geo-lift test |
|---|---|---|
| Holdout test | Incremental effect measured by withholding from random users. | Geo-lift randomizes by region instead of by user - the tool for broad-reach channels you cannot split by person. |
| Incrementality | The extra sales a channel actually causes. | The concept; a geo-lift test is one experimental way to measure it. |
| Marketing mix modeling (MMM) | Channel contributions modeled from history. | MMM infers lift observationally; geo-lift measures it directly and validates the model. |
| Attribution model | Credit assigned to tracked touchpoints. | Correlational and daily; geo-lift is causal and periodic, and rescales what attribution reports. |
Common mistakes
- Non-comparable regions. If test and control markets differ in seasonality, store density or weather, the gap you measure is that difference, not the channel.
- Too few regions. A handful of markets rarely reaches significance; the smaller the true lift, the more regions you need.
- Contamination. Spillover between neighbouring regions, or a national campaign leaking into the test, quietly erases the effect you are trying to isolate.
- A window that is too short. Cutting the test before lagged effects land understates a channel that works slowly.
- Reading lift on revenue, not margin. A channel can lift revenue while losing money; convert incremental sales to contribution before you call it a win.
Geo-lift test FAQ
Related
Blufire S13 Experiments runs geo-lift tests alongside email and ad-set holdouts, so the channels flagged by S12 Financial Models marketing mix modeling can be checked against a controlled, causal result.
Updated July 2026