Geo Lift Tests for Ecommerce
A geo lift test pauses ad spend in a set of matched geographic markets while running normally in comparable markets elsewhere, then compares sales between the two to measure incremental lift. It's the practical alternative to a shopper-level holdout when audience-level randomisation isn't available or reliable, and it works across both DSP and sponsored ads simultaneously.
What this looks like in a real account
Why geo lift instead of a holdout
A shopper-level holdout requires the platform to support genuine randomised exclusion of individuals from an audience — something DSP can do reasonably well, but sponsored ads and cross-channel campaigns generally can't, since search ads serve against a query, not a pre-defined audience segment. Geo lift sidesteps that limitation entirely: instead of randomising people, you randomise markets, pausing all advertising activity — DSP, sponsored ads, everything — in a set of test markets while running normally in matched control markets. That makes it the more practical choice whenever you want to measure a channel or a whole account's combined incremental effect, not just one audience segment.
Choosing matched markets
Matching quality determines whether the test means anything. Pick control markets that resemble the test markets on the dimensions that actually drive your sales — historical revenue trend, seasonality pattern, population size, category penetration — not simply markets of similar total population. A common approach is to match on 8-12 weeks of pre-test sales trend between candidate market pairs before the test begins, selecting pairs whose trends track closely together historically; markets that diverge even before the test starts make a poor comparison once it's running.
A worked example
Take 4 test markets and 4 matched control markets, historically tracking within 3% of each other in weekly sales. Pause all Amazon advertising in the 4 test markets for 4 weeks while running normally in the 4 control markets. Test markets generate $210,000 in sales during the pause; based on the pre-test trend ratio, the control markets' $240,000 in sales during the same period implies the test markets would have generated roughly $232,000 absent the pause (scaling for the markets' historical 3% gap). The shortfall — $232,000 expected minus $210,000 actual — is $22,000, or roughly 9.5% of what those markets would have made with advertising running. That 9.5% is your estimate of incremental lift from the paused spend.
What geo lift can and can't isolate
A pause-based geo lift measures the combined incremental effect of everything paused — if you pause both DSP and sponsored ads together in the test markets, the result tells you the combined lift of both channels together, not which one contributed more. To isolate a single channel's effect, pause only that channel in the test markets while leaving the others running normally in both test and control — a cleaner but more constrained design that requires you to be confident the channel you're leaving running behaves consistently across both market sets.
The common mistake, including ours
The mistake is running a geo lift test during or immediately around a tentpole promotional event without adjusting for it. Our own data shows Prime week roughly doubling DSP click costs agency-wide — a 116% measured increase — and while that's a cost effect specifically, the broader demand and pricing distortions around major sale events move organic and paid sales in test and control markets alike, in ways that can swamp a real but smaller underlying lift signal. We've had a geo lift result come back looking artificially large once we traced the anomaly back to a promotional week neither market set's historical baseline had accounted for, and had to extend the test to get a clean read.
Combining geo lift with your standard attribution reporting
A geo lift result is worth comparing directly against whatever your standard attributed ROAS was reporting for the same markets and period — the gap between the two is informative on its own. A large gap, where attributed ROAS looked strong but geo lift showed modest real incrementality, is a sign the attributed number was substantially capturing demand that would have converted anyway. A small gap, where the two broadly agree, is reassuring but shouldn't be treated as proof the attribution model is now validated for every other market or period — a single geo lift result describes the markets and window it tested, not a permanent calibration.
When the lift comes back small or negative
A small measured lift doesn't automatically mean the spend isn't working — check the matching quality first, since a control market pair that diverged even slightly before the test can produce a misleading gap once real-world noise is added on top. If the matching was solid and the market pair tracked closely pre-test, a genuinely small or flat lift is real information: at that spend level, in those markets, the advertising isn't producing sales beyond what would have happened anyway, and that's worth acting on rather than dismissing because it contradicts a healthier-looking attributed ROAS number from the same period.
| Design decision | What it affects | Practical guidance |
|---|---|---|
| Market matching window | Test validity | Match on 8-12 weeks of pre-test sales trend, not just population |
| Test duration | Statistical reliability | Long enough to smooth normal week-to-week variance |
| What's paused (single channel vs everything) | What the result tells you | Pause one channel only if you need to isolate its individual effect |
| Tentpole event timing | Read accuracy | Avoid or explicitly account for major promotional weeks |
Which one you should actually pick
Any brand with revenue spread across enough distinguishable markets can design and run a geo lift test themselves — it requires careful matching and patience, not proprietary tooling or a paid platform. Where reMKTR adds value is in matching-quality discipline and tentpole-aware test timing, learned in part from tests we've had to extend after an unaccounted-for promotional week distorted an early read, as part of the same practice behind Full Circle's $500M+ in managed Amazon spend across 100+ brands.
Shortlist on the job, not the feature grid. Pull your search-term report for the last 90 days and total the spend against terms that produced no orders — 33.6% on the account above. Then ask each vendor on your list what they would do about it in week one, and see who answers with a process rather than a screenshot.
Common questions
How many markets do I need for a geo lift test?
There's no fixed minimum, but more market pairs generally improve statistical reliability — a handful of well-matched pairs can work for a directional read, while a more rigorous test benefits from a larger, more varied set.
Can a geo lift test measure DSP and sponsored ads separately?
Only if you pause them separately — pausing both channels together in the test markets measures their combined effect, not each one's individual contribution.
How do I know if my control markets are well matched?
Check that their pre-test sales trend tracked closely with the test markets over several weeks before the test began — a pair that already diverges before the pause starts will produce an unreliable comparison.
Should I run a geo lift test around a promotional event?
Avoid it where possible — major sale events distort demand and pricing across both test and control markets in ways that can swamp the underlying lift signal you're trying to measure. If unavoidable, extend the test window so the event is a small share of the total measured period.
We show the method before the number.
Claim the free auditRead next
- Acorn Cost: What Decides the Invoice, and What to AskPricing · acorn cost
- Criteo Alternatives: Sorted by Why You Are LeavingAlternative · criteo alternative
- Pacvue Pricing: No Public Number — What to AskPricing · pacvue pricing
- Skai vs Pacvue: Contracts, Not Feature GridsHead to head · skai vs pacvue