What Is Geo Experiment (Geo Lift)?
A geo experiment is an incrementality test that turns ads on in some geographic regions (treatment) and off in comparable regions (control), then measures the difference in total revenue between the two groups. Because it runs at the market level rather than the user level, it is immune to iOS 14, consent blocking, and cookie loss, making it the most trusted measurement technique in a cookieless 2026.
Formula
The scaling factor normalizes the two groups by pre-test baseline (population, historical sales, or synthetic-control weighting). Divide incremental revenue by treatment ad spend to get true incremental ROAS (iROAS). Statistical significance is usually assessed with a synthetic control method (Google's `CausalImpact` R package or Meta's `GeoLift` open-source library).
Meta ASC geo holdout on a 4-city BD apparel account
Ads live in Dhaka + Chittagong (treatment) and paused in Sylhet + Rajshahi (control) for 4 weeks. Baseline weighting matched control to 82 percent of treatment revenue. Treatment cities generated BDT 62 lakh, expected baseline (from control) was BDT 51 lakh, so incremental revenue was BDT 11 lakh on BDT 8 lakh of ad spend, or an iROAS of 1.38x, versus a reported Meta ROAS of 4.2x. The brand cut ASC budget 40 percent and reallocated to Google Search brand-defence, which tested at iROAS 3.1x.
Benchmarks
- Minimum treatment vs control cities: 3 vs 3 for a directional read, 5 vs 5 for 90 percent confidence.
- Minimum window: 3 weeks; 4 to 6 weeks is standard for DTC.
- Detectable lift with USD 15k spend and 5 vs 5 markets: roughly 10 to 15 percent.
- Recommended cadence: one geo experiment per major channel per quarter.
Why it matters
Every user-level measurement method (Pixel, CAPI, MTA) has been degraded by ATT, ITP, consent gating, and modelling. Geo experiments do not need any user identifier, which is why Meta, Google, and Uber all lean on them internally to sanity-check platform-reported lift. For any account spending over USD 15k per month, a geo test is the single highest-signal measurement decision available.
Common mistakes
- 1.Comparing unmatched cities (Dhaka vs a rural district). Use synthetic control or matched-market pairing, not naive splits.
- 2.Running during a sale, PR event, or seasonal spike. External shocks destroy the control baseline.
- 3.Choosing regions that share media (national TV, national PR). The control gets contaminated by treatment media spill.
- 4.Reading the test at day 3. Weekly seasonality alone can swing lift by 20 percent; wait for the pre-agreed window.
- 5.Skipping the pre-period baseline calibration. Without at least 8 weeks of pre-test history the scaling factor is a guess.
Put Geo Experiment to work
Free calculators
Related services
FAQs about Geo Experiment
How is a geo experiment different from a Meta Conversion Lift study?
Conversion Lift runs at the user level inside Meta's own attribution frame. A geo experiment runs at the market level across all channels, so it captures cannibalisation of Google, TikTok, and organic that Meta's tool cannot see.
How many cities do I need?
3 vs 3 gives a directional read; 5 vs 5 is the standard for 90 percent statistical confidence on a 10 to 15 percent lift. Small-country accounts (BD, SG) often only have 4 to 6 usable metros and should extend the window to 6 weeks instead.
What tools should I use to analyse the result?
Google's `CausalImpact` R package or Meta's open-source `GeoLift`. Both use synthetic control methods that build a weighted counterfactual from the control cities, which is more accurate than a simple treatment-minus-control subtraction.
Can I run a geo test on a small budget?
Under USD 15k monthly spend the confidence intervals swallow any lift under about 25 percent. Small accounts get more signal from a full channel-off week and a blended MER comparison.
How often should I re-run geo tests?
Once per major channel per quarter, plus one whenever you make a large structural change (moving from manual to Advantage+ Shopping, switching bid strategies, launching a new market).
Related terms
Holdout study proving whether ad spend actually caused conversions vs stealing organic credit.
Total revenue ÷ total ad spend, the blended, attribution-free ROAS.
How credit for a conversion is assigned across ad touchpoints.
Estimated conversions Google or Meta report when direct measurement is blocked by consent or ITP.
Regression model that attributes revenue across channels using aggregated, cookieless data.
Meta's automated ecommerce campaign type with minimal manual controls.