Google today marked the general availability of Meridian GeoX, its open-source geographic experimentation tool, during a community livestream on the Google Analytics YouTube channel, alongside a set of changes to the core Meridian library covering automated experiment calibration, brand equity modelling, an agentic skills repository and a switch of computation backend.

In Short

Google has finished testing a free tool that runs geographic advertising experiments, where some regions see ads and others do not, and it is now open to anyone. The results of those experiments can be fed straight into Google's open-source budget model, which previously relied more heavily on statistical patterns in historical data. Google says the design work behind those experiments now costs less and runs faster, though the figures come from its own internal testing rather than independent review.

What was said and where

The session ran 34 minutes and 36 seconds and was hosted by Jeff, who introduced himself as streaming from Seattle, Washington. He opened by noting that the Meridian Discord server had passed 1,800 measurement specialists and data scientists and had reached its one-year mark, describing it as one of the larger measurement communities online. Roughly 70 people were watching concurrently, according to a remark he made around the 25-minute point. The recording carried 174 views on the Google Analytics channel, which lists 522,000 subscribers.

Two product managers presented. The first, credited in the automatic captions variously as Lin, Lynn and Shea, is the product manager for Meridian GeoX and said she was streaming from Los Angeles. The second, Katie Monroe, is the product manager for the core Meridian library. Prior PPC Land coverage of Google's May 2026 Ads DevCast episode rendered the same surname as Munro, and the captions on this stream are machine generated throughout, so spellings and several technical terms in the transcript are approximate. Quotations below reproduce the caption text.

That caveat matters more than usual here. The captions render the product name as "Meridian Gox", "geewax" and "Meridian docs" at different points, collapse marketing mix modelling into "MDM framework", and turn branded Google query volume into "GKV" and "GQ". The underlying product names and terms are recoverable from context and from Google's published documentation, but the transcript is not a clean record.

Three pillars, and the numbers attached to them

The GeoX presentation was organised around three claims. The first was framed as cheaper and more reliable testing. According to the GeoX product manager, native multicell execution allows several treatments to be compared against a single shared control group, which removes the need to stand up a separate control for every test. She put the effect of that design on internal benchmarks at "budget savings of more than 31% for large advertisers" when set against other open-source equivalents.

The figure is a vendor number. No sample size, test period, comparison set or definition of "large advertiser" was given during the stream, and the comparison is against unnamed open-source alternatives. Meta's Robyn is the most obvious member of that category, and PPC Land examined in April 2026 the conflict that arises when a media seller supplies the model that grades its own inventory.

The second pillar was transparency. The open-source codebase, according to the same speaker, gives users "complete visibility, where you can audit the methodologies for trust." Design choices are configurable, and the library supports holdback tests for new campaigns, go-dark tests where spend in treated regions falls to zero, and heavy-up designs where spend rises. On speed, she said the JAX framework had cut design generation times "by over 94% versus other open source solutions and legacy solutions."

The third pillar was the connection back to the mix model itself. GeoX experiment results feed into Meridian as Bayesian priors, and the model in turn nominates which channels would gain most from being tested. That loop was described in Google's May 2026 developer episode as the calibration engine for the wider framework, and PPC Land documented the design intent at the time, when general availability was still targeted for later in 2026.

Stratified sampling against matched markets

The methodological argument occupied a large share of the session. GeoX measures uplift between treatment and control geographic groups and compares observed response against a counterfactual, meaning the response expected had no ads been served. To make that comparison efficient, the library clusters geographies into balanced strata rather than assigning them at random.

Asked how this differs from Google's earlier libraries, the GeoX product manager drew the distinction around validation. Matched markets, she said, relies heavily on in-sample fit, using the same historical data both to fit the model and to estimate its variance. The consequence she named is specific: such an approach can "underestimate the actual noise or variance" that a live test will encounter, producing recommendations with a budget or minimum detectable effect that looks attractively low but leaves the test underpowered.

Stratified sampling in GeoX, by contrast, uses out-of-sample validation, evaluating design stability across historical validation windows that were not used to fit the model. The stated result is a design carrying more buffer against real-world volatility and, in her words, "a higher probability of delivering more conclusive results."

This is a live methodological dispute rather than a settled one. A structural estimation preprint posted to arXiv on August 21, 2026 argued that summarising an experiment into a single lift figure discards most of the information the experiment produced, and separate research from a Zalando author found that standard mix modelling practice overstates paid search returns by roughly 2.5 times. The vulnerability that experimental calibration exists to correct is endogeneity, where the same commercial conditions that drive spend also drive sales, and observational models credit the spend.

Google's position on its older libraries was direct. GeoX is now the intended primary experimentation product, and while further methodologies will be folded in first, "the plan is to eventually consolidate solutions" so that the market faces fewer choices. Synthetic difference-in-differences and synthetic control were named as candidates for inclusion, both already described in the accompanying white paper.

Automated calibration, and the four adjustments

The core library section covered what Monroe called one of the largest releases since Meridian opened globally in January 2025. The release itself landed the previous week. PPC Land's reference entry on the framework dates version 2.0.0 to September 2, 2026 in the public changelog, and records both the backend switch and the calibration additions as breaking changes that require pipeline updates.

The friction the calibration module targets is organisational as much as statistical. Experiments and mix models are frequently owned by different teams, and an experiment measures a point in time while a model spans two or three years of history. Translating one into the other, according to Monroe, often leaves users unsure where to begin. The automated module applies a series of transformations, adjusting for differences in spend thresholds and experiment duration among others, with four adjustments in total. Manual control remains available in the open-source library, and the automation was positioned as reducing friction rather than removing choice: "the automation is there to make this a little bit less friction forward."

Google generalised the same methodology beyond its own tooling, so that experiments run under other designs can be brought into Meridian with less manual translation. The adjustments are surfaced visually inside the model health module, which shipped earlier in 2026, in a new section covering calibration and channel recommendations.

On the validation side, a completed model run now returns per-channel recommendations flagging which channels would benefit from further testing. That output is what closes the loop between the two products.

Brand equity enters the model

The second major addition is a full-funnel framework intended to capture the value of upper-funnel investment. The problem it addresses is familiar to anyone defending a television or online video budget: direct measurement captures the immediate response and misses the accumulation.

The framework runs as a multi-stage model. A first-stage model treats brand equity as the outcome variable to estimate what media does to brand equity. A second-stage model then estimates the relationship between brand equity and the final key performance indicator, whether sales or conversions. Combining both gives what Monroe described as the total return on brand media, direct effect plus indirect effect, rather than the direct line alone.

The recommended input signal is branded Google query volume, meaning total search volume for the advertiser's own brand terms over time. Monroe characterised the variable as functioning as both a mediator and a confounder in the relationship between brand channels and performance channels. The data is already available through Google's MMM Data Platform, where it has served as a control variable for search, so incorporating it requires isolating an existing feed rather than sourcing anything new.

That choice deserves scrutiny. Branded query volume is a Google-owned signal, measured on a Google surface, and now positioned as the recommended proxy for brand strength inside a model that allocates budget across Google and non-Google channels. Google has been steadily building out this class of signal: Attributed Branded Searches, which tracks short-term intent created by an ad, arrived in Google Ads and Google Marketing Platform at Marketing Live in May 2026. Advertisers whose brand searches concentrate on other engines, or whose categories generate little branded search at all, will find the hero signal a poorer fit.

An agentic skills repository for model building

The third component is a repository of agentic skills for Meridian, demonstrated during the stream running inside the Antigravity CLI. The skills give step-by-step modelling guidance grounded in Google's developer documentation, prompting through each stage of the workflow and surfacing best-practice guidance at the point of use.

Monroe named two purposes. The first is automating and auditing data and model quality so that issues surface during the build rather than after it. The second is bluntly stated: to "remove the steep learning curve that our documentation has." The documentation is accurate and heavily maintained, she said, and also very dense.

Antigravity is Google's own agent development platform, rebuilt at Google I/O 2026 on May 19 as Antigravity 2.0 with a separate command-line interface. The skills were described as usable with any AI agent rather than tied to that harness. Google has been pushing the same pattern across its advertising tooling: the Google Ads API assistant reached version 4.0 in a release that targets Antigravity and Claude Code by name, shipping first-party advertising tooling into a competitor's coding agent.

A live demonstration was planned and dropped. Markov chain Monte Carlo sampling takes several minutes to complete, which does not fit a half-hour stream, so Google said a dedicated walkthrough video would appear on the Google Analytics YouTube channel within the following weeks. The stated contents include configuring the skills and a command-line interface in under five minutes, automating exploratory data analysis before a model run, and using agentic prompts to troubleshoot convergence and prior sensitivity.

JAX by default, and the Scenario Planner question

Meridian now runs on JAX out of the box. According to Monroe, internal testing puts it at twice the speed and four times the memory efficiency of the TensorFlow backend it replaces. Those figures match what Google published when JAX arrived as an option in version 1.6.0 on April 29, 2026, at which point the company stated the default would switch during the second half of 2026. The switch has now happened, and the TensorFlow path is deprecated.

Optimisation engine changes were also described, generating Scenario Planner dashboards in minutes when run on GPUs. The Scenario Planner, a code-free budget interface that opened as a beta with per-channel spend constraints in August 2026, remains available to the whole open-source community.

One naming point drew a question from the chat. The dashboard is built in Data Studio, formerly Looker Studio, following Google's April 11, 2026 decision to retire the Looker Studio name. Monroe acknowledged the confusion the rename created. The workflow it enables is a handoff: technical staff connect the model to the dashboard, and non-technical stakeholders then run budget scenarios across quarters or an annual planning cycle without executing code.

What Meridian still will not do

Two limits were stated plainly. The first concerns forecasting. Meridian optimises against historical periods, asking counterfactual questions about spend that already happened. It does not project future conditions, because, according to Monroe, fluctuation and business factors make those projections unreliable. Advertisers holding their own forward view can pass future data into the optimiser through a new data argument and receive an optimal allocation for that period, but the model supplies no assumptions of its own: "we don't make those assumptions on behalf of the user."

The second is timing. 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.

Why this matters

Mix modelling has absorbed a large share of measurement budget since privacy restrictions degraded user-level attribution, and confidence in the results has not kept pace with the investment. 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 and 46.9% planning to increase that investment. 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 that market has converged on, and the cost of running it has fallen sharply. Google cut its minimum experiment budget to 5,000 dollars in November 2025 from thresholds approaching 100,000 dollars. Vendors moved in parallel across 2026, and geographic and household-level holdout designs became standard inventory in measurement pitches.

What today's change alters is ownership of 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 result in a dashboard built on Google's visualisation product, is running a measurement stack in which every component originates with a company that also sells a large share of the media being graded. The code is auditable, which is a genuine difference from closed vendor tooling, and CIMM warned on July 31, 2026 that default assumptions embedded in open-source frameworks still carry allocation consequences whether or not anyone reads them.

For agencies and in-house teams, the practical consequence is a shorter path from experiment to budget decision, and a longer list of Google-supplied defaults sitting inside that path. The 31% and 94% figures will hold or fail on independent replication that has not yet happened.

Timeline

Summary

Who: Google, through the Meridian team on its Analytics Discord and YouTube channels. Two product managers presented: the product manager for Meridian GeoX, named inconsistently in the automatic captions, and Katie Monroe, product manager for the core Meridian library. The host was identified only as Jeff. The audience is advertisers, agency measurement teams and data scientists running marketing mix models.

What: General availability of Meridian GeoX, an open-source geographic incrementality testing library using stratified sampling and time-based regression with native multicell designs, plus four additions to the core Meridian library: an automated experiment calibration module applying four adjustments, a full-funnel framework using branded Google query volume as a brand equity signal, an agentic skills repository for AI coding agents, and JAX as the default computation backend. Claimed gains are budget savings above 31% for large advertisers, design generation faster by more than 94%, and runtimes twice as fast at a quarter of the memory.

When: The livestream took place on September 9, 2026 and ran 34 minutes and 36 seconds. The core library release it described, version 2.0.0, is dated September 2, 2026 in the public changelog. Time-varying media return on investment is slated for later in the fourth quarter of 2026, and a dedicated agentic skills walkthrough video is expected within weeks.

Where: Delivered on the Google Analytics YouTube channel and the Meridian Discord server, which has passed 1,800 members. The tooling itself is distributed through GitHub and Google's Meridian developer site, with dashboards running in Data Studio.

Why: Mix models estimate media contribution from correlations across time and geography, an approach exposed to bias when the conditions driving spend also drive sales. Geographic experiments produce causal evidence that can be imported as priors to constrain those estimates. Google's stated aim is to shorten the path between running an experiment and calibrating a model, and to consolidate its several experimentation libraries into one. The commercial context is a measurement market in which confidence has stalled despite rising investment, and in which a growing share of the calibration stack originates with a company that also sells the media being measured.