Meridian GeoX: How Google’s Open-Source Geo Experiments Work

01

Overview

Google made Meridian GeoX generally available globally on September 10, 2026, turning its previously beta geo-experimentation library into a production-ready part of the broader Meridian measurement stack.

Meridian GeoX is an open-source framework for measuring the incremental impact of marketing by comparing geographically separated treatment and control groups. Unlike click attribution, it is designed to estimate causal lift: what happened because marketing changed, rather than simply which touchpoint received credit.

02

What Meridian GeoX does

GeoX helps marketers design, run and analyze geographic incrementality experiments. Google documents support for holdback, go-dark, heavy-up and multi-cell experiment designs. The library is publisher-agnostic, so the same methodology can be used to evaluate campaigns across different advertising platforms rather than only Google media.

03

Why the global GA matters

GeoX existed earlier in beta, so the September 10 event is not the first appearance of the product. The material change is general global availability inside Meridian. Google now positions GeoX as a direct way to generate causal evidence that can be used independently or fed back into Meridian Marketing Mix Models.

04

How a GeoX experiment works

A typical workflow has several stages:

1. Gather historical geo-level response and spend data.
2. Choose an experiment type and define operational constraints.
3. Generate candidate treatment/control assignments and inspect statistical power and minimum detectable effect.
4. Run the campaign intervention across the selected geographies.
5. Collect test-period outcomes.
6. Analyze the result with counterfactual modeling and design-aware inference.
7. Use the estimated incremental effect for decision-making or Meridian MMM calibration.

GeoX primarily uses time-based regression for counterfactual analysis and reports metrics such as incremental conversions, percentage lift, confidence intervals, p-values and incremental conversion per dollar. If revenue is used as the response, the resulting measure can be interpreted similarly to incremental ROAS.

05

Supported experiment designs

Holdback tests withhold incremental media from selected geographies. Go-dark tests reduce or stop existing media in treatment regions. Heavy-up tests increase media pressure. Multi-cell designs let teams compare more than one treatment against a common control, reducing the time and cost required for separate experiments.

The best design depends on business objectives, existing spend and operational constraints. A design that is technically available is not automatically statistically valid for every advertiser.

06

Data and feasibility requirements

Geo experimentation requires sufficient geographic variation, historical data and a test large enough to detect a plausible effect. GeoX calculates minimum detectable effect during the design process and includes validation and quality checks for issues such as missing data, sparse conversions, duplicate rows and unstable geo relationships.

Google's documentation requires daily data. The library also applies constraints around available geographies and pre-test duration. This means small advertisers or businesses with very limited regional volume may not have enough statistical power for a useful test.

07

Open-source installation

Meridian GeoX is published under Apache 2.0 and is available on GitHub and PyPI. The standalone package can be installed with pip install --upgrade meridian-geox. Teams using it alongside Meridian can install the combined package through the documented Meridian extras path.

Python 3.10 or later is required for GeoX alone; Google documents Python 3.11 or later when using GeoX together with Meridian MMM. CPU-based JAX installs automatically, while GPU-backed JAX can be used for heavier simulations.

08

Meridian MMM integration

The strongest strategic use case is calibration. Marketing mix models estimate channel contribution over time, but they can still contain uncertainty or biased priors. GeoX provides experimental evidence that can be converted into priors for Meridian, grounding the model in observed causal lift.

Meridian 2.0 can also recommend channels that would benefit from a new incrementality experiment when the model shows high uncertainty, potential bias or implausible ROI behavior.

09

Pricing

GeoX itself is open-source software and Google does not publish a standalone license fee for the library. That does not make geo experiments free. Real cost can include campaign budget changes, analyst time, infrastructure and the opportunity cost of withholding or redistributing media.

10

Important limitations

Geo experiments measure the effect of a controlled intervention under specific conditions. They do not automatically explain every customer journey, identify individual users or replace all attribution and MMM workflows.

Results also depend on design quality, geographic isolation, seasonality, sufficient statistical power and accurate conversion/spend data. A poorly powered test can produce an inconclusive answer even when the software works exactly as intended.

11

Meridian 2.0 context

Google's September measurement announcement also highlights Meridian 2.0 improvements such as native GeoX calibration, brand-equity signals, agentic skills and a JAX backend. Google's speed and efficiency figures for the new backend are vendor-reported and should not be treated as universal independent benchmarks.

12

Who should use Meridian GeoX

GeoX is most relevant for growth teams, performance marketers, marketing scientists and agencies that spend enough across regions to run controlled geographic tests and want causal evidence beyond platform attribution. Smaller businesses may get more value from simpler conversion measurement until they have enough scale for geo experimentation.

13

Bottom line

Meridian GeoX gives marketers an open-source way to design and analyze cross-platform incrementality tests and then use those results to improve Meridian MMM calibration. Its move to global general availability makes it substantially more important as a practical measurement tool, but successful use still depends on experiment design, statistical power and clean data—not just installing the library.