Meridian GeoX is Google's open-source library for designing and analysing geographic advertising experiments, released to general availability on September 9, 2026, whose results feed into the Meridian marketing mix model as causal priors.

Meridian GeoX is a software library that runs geographic incrementality experiments and hands their results to a marketing mix model. A geo experiment splits a market into regions, changes advertising spend in some of them, keeps others unchanged as a control, and measures the difference in outcomes. That difference estimates how much a channel actually caused, rather than what a correlational model attributes to it. GeoX exists because the dominant privacy-safe measurement method, marketing mix modelling, reads statistical patterns in historical data and cannot on its own separate the effect of advertising from everything else that moves alongside it. The tool is distributed free on GitHub under the same open licensing as Meridian itself, and its stated purpose is to generate what Google calls ground-truth experiment data that anchors a model in tested reality.

The library is publisher-agnostic and cookieless by design. Because a geo experiment measures aggregate regional outcomes rather than tracking individuals, it needs no user-level identifiers and works across any channel where regional targeting is possible, including channels a given advertiser does not buy through Google. That property is central to how Google positions the product: a single experiment framework spanning search, video, social, and offline media, feeding one model.

How a geo experiment works

The unit of a geo experiment is a geographic region, typically a Designated Market Area, a metropolitan region, or a postal cluster, depending on the market. Regions are sorted into a treatment group and a control group. During a defined test window, spend in the treatment group is modified. The control group carries on unchanged and supplies the counterfactual, meaning the outcome that would have occurred without the intervention.

GeoX supports three experiment types. A holdback test withholds a new campaign from the control regions, useful when launching something that was not previously running. A go-dark test cuts spend in treated regions to zero, measuring what disappears when a running channel stops. A heavy-up test raises spend in treated regions, measuring the return on additional investment. These map to different questions a budget owner faces: whether to start, whether to stop, and whether to scale.

The measurement engine is time-based regression, abbreviated TBR, a method Google researchers formalised in 2017. TBR models the relationship between treatment and control response during a pre-test period, then projects the control forward through the test window to construct the counterfactual. Observed treatment response is compared against that projection, and the gap is the incremental effect. An earlier method, geo-based regression, and a newer option, randomized sampling, sit alongside it; the white paper names synthetic difference-in-differences and synthetic control as methodologies slated for inclusion, part of a stated plan to house multiple approaches in one library and let users compare designs before committing.

The step that distinguishes GeoX from Google's older tools is how regions are assigned. Rather than randomising, the default rule clusters geographies into balanced strata, a technique the documentation calls stratified sampling. According to the GeoX product manager on the general-availability livestream, the older matched-markets approach leaned on in-sample fit, using the same historical data to both fit the model and estimate its variance, which can underestimate the noise a live test encounters and produce designs that look powerful on paper but run underpowered in practice. Stratified sampling uses out-of-sample validation across historical windows the model did not fit, which the product manager said yields designs with more buffer against real-world volatility.

Fields, defaults, and configuration

The library exposes its design parameters through a configuration class. The published API reference lists a DesignConfigdataclass whose defaults reveal the intended standard test: experiment type set to holdback, methodology set to time-based regression, geo assignment set to stratified sampling, one treatment cell, a significance level (alpha) of 0.1, statistical power of 0.8, a two-sided test, ten candidate designs returned, and 100,000 candidate splits searched. The single-cell default reflects the simplest case, but the feature Google emphasised at launch is native multi-cell execution: comparing several treatments, such as different publishers or tactics, against one shared control group in a single study. That removes the need to stand up a separate control for every comparison, which is where the headline cost savings are claimed to come from.

The user journey documented by Google runs in seven stages. An optional first step surfaces an experiment suggestion from an existing Meridian model. Data collection follows, requiring historical geo-level daily time series of the key performance indicator, plus historical daily spend by region for designs that modify budgets. The design stage runs the algorithm to produce ranked, viable designs. Implementation sets up geo-targeted campaigns in-platform. Results collection gathers outcome data over the test window, with an optional cooldown period afterward to capture lagged effects. Analysis applies counterfactual modelling and robust inference to judge statistical significance. A final optional step calibrates a Meridian model with the findings, provided the experiment passes data-quality checks.

Calibration is the reason the two products are paired. A completed GeoX experiment produces an incremental effect that enters Meridian as a Bayesian prior, constraining the model's estimate for that channel toward the tested value. The relationship runs both ways: a completed Meridian model returns per-channel recommendations flagging which channels would gain most from a test, closing a loop where the model nominates experiments and the experiments discipline the model.

Origin and evolution

The methodology predates the product name by more than a decade. Google researchers published "Measuring Ad Effectiveness Using Geo Experiments" in 2011 and the time-based regression follow-up in 2017, and released open-source implementations in the repositories google/matched_markets, google/trimmed_match, and an R package for geo experiment analysis. Meridian GeoX consolidates that lineage into a single named library inside the Meridian ecosystem.

The immediate history is compressed into 2026. Google disclosed GeoX on May 5, 2026, in a pre-Google Marketing Live measurement push announced through a blog post by Gaurav Bhaya, VP and GM of Buying, Analytics and Measurement at Google, alongside Meridian Studio and a Data Manager update. At that point the tool was framed as beginning testing later in the year, with no broader availability date. Google's May 28 developer episode described GeoX as the calibration engine for the wider framework while it ran in closed beta. The tool moved into its own GitHub repository with guides, Colab notebooks, and an API reference over the following months. On September 2, 2026, version 2.0.0 of the core Meridian library switched the default computation backend from TensorFlow to JAX, added prior calibration from incrementality experiments including GeoX, and introduced channel calibration recommendations. General availability was marked on September 9, 2026, during a community livestream on the Google Analytics YouTube channel.

The general-availability event carried specific vendor claims. The GeoX product manager put budget savings above 31% for large advertisers against unnamed open-source alternatives, and design-generation speed improvements above 94%, attributed to the JAX backend. Both figures came from Google's internal benchmarking rather than independent review, and no sample size, test period, or definition of "large advertiser" accompanied them.

Why it matters for the marketing community

Mix modelling absorbed a large share of measurement budget after privacy restrictions degraded user-level attribution, and confidence has not tracked the spending. TransUnion and EMARKETER research published in October 2025 found 54.1% of marketing professionals reporting no year-over-year change in measurement confidence and 14.3% reporting a decline, with 49.5% already running mix models. IAB State of Data 2026 found up to 75% of buy-side decision-makers rating current measurement methods as underperforming. Incrementality testing is the correction the market has converged on, and its cost has fallen: Google cut its minimum incrementality experiment budget to 5,000 dollars in November 2025 from thresholds approaching 100,000 dollars, and vendors moved in parallel through 2026.

GeoX changes who owns the calibration path. An advertiser running a geo test through GeoX, feeding it into Meridian through an automated module built by the same company, and reading the output in a dashboard built on Google's own visualisation product is operating a measurement stack in which every layer originates with a firm that also sells a large share of the media being graded. The code is auditable, a real difference from closed vendor tooling. But CIMM warned on July 31, 2026 that default assumptions embedded in open-source frameworks still carry allocation consequenceswhether or not anyone inspects them.

Limitations and disputes

The most direct methodological challenge concerns what a geo experiment gives up when its result becomes a single number. A structural estimation preprint posted to arXiv on August 21, 2026 argued that collapsing an experiment into one lift figure discards most of the information the experiment produced. Separate research from a Zalando author found that standard mix modelling practice overstates paid search returns by roughly 2.5 times, estimating observational carryover at 0.50 against a true value near 0.20. The vulnerability that experimental calibration exists to correct is endogeneity, where the commercial conditions that drive spend also drive sales, and an observational model credits the spend. GeoX addresses that bias but inherits a debate over how much of the experiment's signal survives the handover.

A second concern is the brand-strength signal Google recommends inside the wider framework. The full-funnel module added in version 2.0.0 uses branded Google query volume as its input for brand equity, a signal measured on a Google surface and used to allocate budget across Google and non-Google channels. Advertisers whose brand searches concentrate on other engines, or whose categories generate little branded search, will find that proxy a poor fit. The vendor-number problem applies throughout: the 31% and 94% claims will hold or fail on independent replication that has not yet happened.

Disambiguation

Meridian is the marketing mix model itself, an aggregate Bayesian framework that estimates channel contributions from historical data. GeoX is the experimentation library that calibrates it; the two are separate products with separate repositories.

Incrementality is the general concept of measuring advertising's causal effect above a baseline. GeoX is one implementation of it, specific to geographic experiments, sitting alongside ghost bidding, user-level holdouts, and always-on approaches.

A holdout study withholds ads from a randomly selected group of users and compares outcomes. GeoX withholds or modifies ads across regions rather than user cohorts, which is what makes it cookieless and cross-publisher.

Matched markets is the older Google approach GeoX supersedes, relying on in-sample fit and pairing similar regions. GeoX replaces that assignment logic with stratified sampling and out-of-sample validation.

Recent developments

At general availability on September 9, 2026, Google marked GeoX as generally available globally and detailed the surrounding version 2.0.0 release: an automated calibration module applying four adjustments to translate an experiment into model priors, a full-funnel brand equity framework, an agentic skills repository for AI coding agents, and JAX as the default backend running at roughly twice the speed and four times the memory efficiency of TensorFlow. The GeoX product manager confirmed the intent to consolidate Google's several experimentation libraries into GeoX over time. Time-varying media return on investment, allowing effectiveness to shift across a modelled period, was confirmed as planned for later in the fourth quarter of 2026.

Timeline

  • 2011 - Google researchers publish "Measuring Ad Effectiveness Using Geo Experiments"
  • 2017 - The time-based regression follow-up methodology is published
  • March 2024 - Google unveils Meridian as an open-source marketing mix model on a limited basis
  • January 29, 2025 - Meridian opens globally after testing with hundreds of brands
  • November 11, 2025 - Google lowers the minimum incrementality experiment budget to 5,000 dollars
  • February 19, 2026 - The Meridian Scenario Planner launches as a code-free budget interface
  • April 29, 2026 - Version 1.6.0 adds an optional JAX backend
  • May 5, 2026 - Meridian GeoX and Meridian Studio are disclosed ahead of Google Marketing Live, with GeoX testing set for later in 2026
  • May 20, 2026 - Meridian is integrated into Google Analytics 360 at Google Marketing Live 2026
  • May 28, 2026 - A developer episode confirms GeoX closed beta is running with customers
  • July 31, 2026 - CIMM warns that default priors in open-source frameworks carry budget consequences
  • August 21, 2026 - A structural estimation paper on recovering mix model parameters from geo-experiments is posted to arXiv
  • September 2, 2026 - Version 2.0.0 makes JAX the default backend and adds prior calibration from incrementality experiments
  • September 9, 2026 - Google marks Meridian GeoX general availability on a community livestream

Summary

Who: Google, through its Meridian measurement team, built and released Meridian GeoX. The audience is advertisers, agency measurement teams, and data scientists who run marketing mix models and need causal evidence to calibrate them.

What: An open-source library that designs and analyses geographic incrementality experiments using stratified sampling and time-based regression, supports holdback, go-dark, and heavy-up designs with native multi-cell execution, and feeds results into the Meridian marketing mix model as Bayesian priors.

When: Disclosed on May 5, 2026, run in closed beta through the middle of the year, and marked generally available on September 9, 2026, alongside the version 2.0.0 core-library release dated September 2, 2026. The underlying geo-experiment methodology dates to Google research published in 2011 and 2017.

Where: Distributed through GitHub and Google's Meridian developer site, with experiments run in-platform across any channel supporting regional targeting, and results optionally imported into Meridian.

Why: Marketing mix models estimate media contribution from correlations across time and geography, an approach biased when the conditions driving spend also drive sales. Geographic experiments produce causal evidence that can constrain those estimates. Google's stated aim is to shorten the path from running an experiment to calibrating a model, at a moment when measurement confidence has stalled despite rising investment.