Google's Meridian GeoX Just Went Live:
What It Actually Changes (and What It Doesn't)

What Google actually shipped

On September 9, 2026, Google took its geo-incrementality testing library, Meridian GeoX, out of beta and into general availability, announcing the move on a livestream on the Google Analytics YouTube channel. Alongside it, Google shipped Meridian 2.0.0, the underlying open-source MMM framework, with JAX replacing TensorFlow as the default computation backend.

For anyone running or evaluating a marketing mix model, this is worth pausing on — not because GeoX is revolutionary, but because of what its release validates about where measurement is heading, and what it quietly leaves unresolved.

Meridian GeoX is an open-source library for designing and analyzing geographic advertising experiments: split a market into regions, change spend in some, hold others as a business-as-usual control, and measure the difference in outcomes. That's a standard geo holdout design. What's new is the packaging — GeoX results are built to feed directly into Meridian, Google's open-source marketing mix model, as causal priors. Meridian 2.0.0 landed the same week, with JAX as the default backend in place of TensorFlow. Google claims roughly twice the processing speed and four times the memory efficiency as a result — a meaningful upgrade for anyone running Meridian on large, multi-year datasets with many geos and channels.

This isn't the only measurement infrastructure that shifted this same month. Google's own platform-owned MMM push — folding Meridian directly into Google Analytics 360 — is part of the same broader story, one we examined in detail here, including how it compares with Meta's parallel retreat from Robyn.

Why "causal priors" is the part worth understanding

The core weakness of any marketing mix model is that it's fundamentally correlational: it's fitting a statistical relationship between spend and outcomes across historical data, and that relationship can be confounded by seasonality, competitor moves, or channels that happen to run together. A geo experiment produces something different — a controlled, causal estimate of what a specific channel actually did in a specific window.

Feeding that causal estimate back into the model as a prior is how you get calibrated MMM: instead of asking the model to infer causality purely from correlation, you're giving it a real, experimentally grounded anchor point to reason from. This is the same logic behind Meta's Robyn calibration and most serious MMM practice today — Google packaging it as a one-click, integrated workflow is what's genuinely new here, not the underlying statistics. For a closer, vendor-neutral comparison of how Meridian and Robyn actually stack up as modeling tools, see our earlier breakdown.

The numbers Google is touting — and why to treat them carefully

Google's product manager claimed GeoX delivers budget savings of more than 31% for large advertisers compared with unnamed open-source alternatives, and cuts experiment design generation time by more than 94%. Both numbers were presented without a disclosed sample size, test period, comparison set, or definition of what counts as a "large advertiser," and neither figure has been independently replicated.

That's not a knock on GeoX specifically — it's a reminder that applies to every vendor benchmark in this space, including from measurement platforms that position themselves as neutral. Ask any vendor — Google included — to show you the validation data, not just the headline percentage.

The gap GeoX doesn't close

Here's the part that matters most for anyone whose media mix isn't Google-native. GeoX experiments are naturally scoped to the channels Google can instrument and where you're willing to run geo splits — typically Google/YouTube spend, and other digital channels you can control at the geo level. A geo test on your YouTube budget tells you what YouTube did in that window. It doesn't tell you how that YouTube spend interacted with a linear TV flight, an out-of-home campaign, or a radio buy running the same weeks in the same markets — because those channels sit outside the experiment, and often outside the advertiser's own MMM entirely if the model was built digital-first.

Offline media doesn't get a free pass just because it's harder to instrument. TV, OOH, radio, and print all move the same demand a geo-split YouTube test is trying to isolate, and a model that treats them as an afterthought — or leaves them out — will misattribute the very incrementality GeoX is designed to surface. Causal calibration only works if it's calibrating a model that actually contains your whole mix. That's exactly the kind of blind spot we've also seen show up on the data side — see, for instance, how a single upstream methodology change at Nielsen quietly rewrote local TV ratings in the same two-week window as GeoX's launch.

The trust data backs triangulation, not single-method bets

According to EMARKETER's 2026 research on incrementality, 60% of US senior marketing decision-makers trust independent incrementality testing most as a measurement method, compared with 40% for marketing mix modeling and 37% for in-platform reporting. Read that carefully: incrementality testing wins the trust contest specifically because it's a check on the other methods, not a replacement for them. Nobody in that survey is saying MMM should go away — they're saying it needs causal grounding.

That's exactly the workflow GeoX formalizes for Google's own channels. The open question for most advertisers is whether their MMM is built to accept that kind of causal grounding across every channel they actually buy — not just the ones a single ad platform can run an experiment on.

What this means for your stack

A few practical takeaways if you're evaluating your own measurement setup in light of this release:

• Don't mistake platform-native incrementality testing for complete measurement. A geo test run by the platform selling you the media is a useful data point, not an independent audit.
• Use experiment results as calibration inputs wherever your MMM can ingest them — whether they come from GeoX, a vendor-run geo holdout, or your own in-house test.
• Press any MMM vendor, including Google, for validation methodology, not headline percentages. Ask what MAPE, R-squared, and holdout accuracy look like on your own data, not a hypothetical "large advertiser."
• Make sure offline channels get the same causal treatment as digital ones. If your model can't absorb a TV or OOH geo test the same way it absorbs a YouTube one, the model has a structural blind spot, not just a data gap.

GeoX is a real step forward for Google's own measurement stack, and the shift to causally calibrated MMM it represents is the right direction for the industry. But the release is also a useful stress test for a question every marketer should be asking their own measurement provider: does your model actually span the media you buy, offline and online together — or just the part one platform can see?

That's the specific gap Animo is built to close. Animo runs a single, always-on causal MMM across TV, out-of-home, radio, print, and every digital channel at once, so incrementality evidence from any source — a geo test, a holdout, a regional rollout — calibrates one model that reflects your real mix, not a digital-only slice of it. If you're rethinking how experiment data should feed your model this quarter, we're happy to walk through how that works on your own data.

Sources

• Google Meridian GeoX general availability announcement, Sept 9, 2026 livestream: searchenginejournal.com
• Meridian GeoX exits beta, 31% savings claim, 94% faster design generation, no independent verification: ppc.land
• Meridian GeoX general availability and savings claim, confirming detail: relevantaudience.com
• Official Google announcement of Meridian GeoX: business.google.com
• Meridian GeoX explainer / background on geo-experiment-to-MMM workflow: ppc.land and ppc.land
• Meridian 2.0.0 JAX backend change, changelog: github.com
• EMARKETER: 60% trust incrementality testing most, 40% MMM, 37% in-platform reporting: emarketer.com