GeoLift vs Meridian GeoX vs CausalImpact: what each tool is for
The method matters more than the tool. For the full picture, read incrementality for mid-size ecommerce brands.
What does each tool do?
GeoLift. An R package from Meta's open-source repository, using synthetic control1. Synthetic control builds a control from a weighted mix of untreated regions that matched the test region before the change. Its documentation calls it a research tool from Meta's Marketing Science team, meant for cases where people-based measurement is not feasible, and says Meta recommends people-based experiments where possible4.
Meridian GeoX. Google describes it as an open-source incrementality solution for geo experiments, with designs such as holdback, go-dark and heavy-up2. Its results can be converted into priors to calibrate the Meridian marketing mix model, and it can also run on its own2. Trade press reports that Google marked its general availability in September 20265.
CausalImpact. A package that estimates what a metric would have done without an intervention, using a time-series model. Its documentation says it assumes the control series were not themselves affected by the intervention3.
None of them runs the experiment for you. They help you choose markets, check the design and read the result.
Why do e-commerce brands look at them?
Because a platform's own numbers are credited, not proven. A geo test compares regions where ads changed with regions where they did not.
Google's own help centre compares test types. It says a geo-based lift test needs a higher budget and has a noisier split than a user-based one6. Open-source tools give a brand another route, with no platform grading its own work.
They are also free to download. That is not the same as free to run.
How do they differ in practice?
| GeoLift | Meridian GeoX | CausalImpact | |
|---|---|---|---|
| Owner | Meta open-source repo, MIT licence1 | Google, open source2 | Google R package3 |
| Core method | Synthetic control methods1 | Time-based regression, stratified sampling2 | Time-series effect estimate3 |
| Main use | Design and read a geo test | Plan geo experiments, calibrate Meridian | Estimate an effect after the fact |
| Key assumption | Control regions match test | Sensible markets and design | Control series unaffected3 |
One measurement vendor argues that matched-market tests fall short and that synthetic control is better7. That is a vendor view, not a settled question.
A simulation by an MMM vendor, on synthetic data, found CausalImpact more prone to false alarms than GeoLift8. It is one study, by a vendor, so we take the direction and leave the figures out.
How do we implement it?
Check analyst and data
Confirm someone can run and check the work, and that you can get sales by region. If not, stop here.
Match tool to question
The three tools do different jobs. Choose by your data, your markets and the question, not by name.
Design before you test
Choose the regions and check the smallest lift the design could see. Watch for spillover and seasonality.
Read it as an estimate
Report the result with its range. Accept that the answer can be inconclusive.
When should you not use them?
Skip them, for now, when:
What does it cost in effort?
No money figures here.
- Analyst time. Someone has to run and read it.
- Data. Sales by region, over enough history.
- Design. Choosing markets and checking minimum detectable effect (MDE): the smallest lift the design can reliably see.
- Patience. A test runs for a while, and the answer can be "inconclusive".
In small countries there may be too few regions to test well.
Pitfalls
Picking regions by gut
Poor matches give poor controls.
Spillover
Ads reach people in control regions, or the other way round.
Seasonality
A promotion or holiday in one region distorts the read.
Reading a tool's output as proof
It is an estimate with a range.
Mixing up names
"Geo lift" is a result. GeoLift is a package.
Our take: Selective
We rate these tools Selective. This is our view, not a market fact.
They are useful in the right hands. They need an analyst who can work in R or Python and sales data by region. Without that, the output is easy to misread.
We do not rank the three. Which fits depends on your data, your markets and the question.
Our view: the biggest risks are not in the software. They are the choice of markets, spillover between regions, and seasonality.
FAQ
Are GeoLift and Meridian GeoX the same?
Does Meta recommend GeoLift over people-based tests?
No. GeoLift's documentation says Meta recommends people-based experiments where possible, and positions GeoLift for cases where they are not feasible4.
CausalImpact vs GeoLift: what is the difference?
What does GeoLift market selection mean?
It is the design step: choosing which regions get the change, which serve as control, and what lift the design could detect before you spend. Most of the risk sits here, not in the analysis.
What is CausalImpact for?
It estimates the effect of an intervention from a time series, assuming control series were not affected3.
Do I need an analyst?
Yes, in-house or at an agency. These are code libraries, not dashboards.
Which one should I use?
We do not rank them. It depends on your data and question. Start with the design: which regions, how long, and what lift you could detect.
Sources
- GeoLift (GitHub, facebookincubator)Primary source
- Meridian GeoX (Google for Developers)Primary source
- CausalImpact (dokumentacja pakietu R)Primary source
- GeoLift docs: introPrimary source
- PPC Land: Meridian GeoX exits beta (09.2026)Blog or media
- Comparing lift types (Google Ads Help)Primary source
- Haus: Matched Market Tests Don't Cut ItVendor
- Recast: Geo-experiment Simulation StudyVendor
Let’s find what’s actually working.
Next step. Want a view on whether a geo test fits your budget? Talk to us about your budget decision. Or read about our Measurement & incrementality service. You can start with a free measurement audit.