Google Meridian: GeoX MCP Skills Update

Google's latest Meridian update, announced on 2 September 2026 and covered by Search Engine Journal and Marketing Dive over the following two weeks, makes GeoX generally available worldwide. It also adds a full-funnel media mix modelling (MMM) option, an agent toolkit built on the Model Context Protocol, and a faster JAX engine. Google says Meridian has passed 1 million downloads in the 18 months since launch.

Media mix modelling often sits in a quarterly slide that few people can rerun. This update targets that problem, with tests that feed the model, brand signals for upper-funnel spend and help for the analyst who has to debug it.

What changed

GeoX is Google's open-source library for geo experiments, and it works across ad platforms. Its results can be fed into Meridian through automated calibration, so a lift test can anchor the model.

The full-funnel option adds brand equity signals, such as branded Google query volume, to show how video and TV spend affects later sales. The MCP toolkit lets AI agents such as Antigravity CLI guide modelling and fix errors as they appear. Marketing Dive adds that the agent features help audit data quality, and that the update extends measurement to TV and out-of-home.

Where it fits in marketing analytics work

If you build MMM in-house, the agent toolkit may shorten the slow part, which is chasing data errors. If you buy MMM from a vendor, the same update is a reason to ask which of these features they support.

Search Engine Journal cautions that branded query volume does not establish causation, because competitor activity, seasonality and news coverage all move demand too.

Try this

Run a readiness check before you touch any new feature.

First, confirm you have daily time series for spend and your main KPI by channel. SEJ lists daily data and enough geographic variation as requirements for GeoX, and recommends GPU resources for Meridian's heavy modelling.

Second, list each channel and note whether it has ever had a lift test. Channels without one are where calibration will help most.

Third, fit your last model twice, with and without branded query volume, and compare the estimated return on upper-funnel video. A large swing tells you the brand signal is doing real work in the model, and you should explain it before anyone acts on it.

Fourth, point an agent at a past model in a sandbox and time how long it takes to find a known data problem. Compare that with your usual debugging time.

The takeaway

Do the four-step check this week and write down which channels lack a lift test. That list is your case for running one.

---

Does your team build MMM in-house or buy it, and what slows you down most? Tell me in the comments, and subscribe for more practical measurement posts.