Meridian Agent Skills: Google Ships an Agent-Native MMM

On September 10, Google published a measurement update that folded a lot of separate announcements into one post. Buried in the Meridian section was a line about "new agentic capabilities" — and most of the coverage, reasonably, read that as "AI will help you build models." The documentation says something narrower and considerably more interesting. Google shipped five Meridian agent skills, distributed inside the open-source repository, built on the open Agent Skills specification. Not a panel inside a Google dashboard. A set of instruction packages your existing coding agent can load.

That distribution choice is the part worth thinking about, and it is a different story from the one about AI making marketing mix modeling easier — which, as we will get to, is mostly not true.

What Google Actually Shipped

Five skills, each mapped to a stage of the MMM lifecycle:

Skill	Identifier	Key outputs
Documentation Consultant	meridian-doc-consultant	Grounded technical explanations with references
Model Building	meridian-model-building	Python script, EDA HTML report, saved model (.binpb)
Result Visualization	meridian-result-visualization	MCMC convergence health check, 2-page summary HTML
Budget Optimization	meridian-budget-optimization	Budget optimization report HTML
Scenario Planner	meridian-scenario-planner	Scenario data file, Looker Studio handoff

Installation runs through standard agent tooling rather than anything Google-specific:

npx skills add google/meridian
npx skills add google/meridian --skill meridian-model-building

If you work inside a clone of google/meridian, compatible assistants discover the skills automatically through the workspace manifest at .agents/skills.json. There is no license, no waitlist, and no Google product surface in the loop.

Two other things shipped in the same release and they matter to the argument. Meridian GeoX — the open-source library for running causal geo-experiments across any ad platform, not just Google's — moved from beta to general availability worldwide. And Meridian models can now incorporate brand signals such as Branded Google Query Volume, aimed at attributing longer-term effects from TV and out-of-home.

The Distribution Choice Is the Story

For roughly two decades, new capability from an ad platform arrived in one of two shapes: a screen in the platform's UI, or an endpoint in the platform's API. Both keep the platform in the middle of the interaction. This arrives as a third shape — portable Markdown instructions plus Python templates, conforming to a spec Google does not own, runnable in any agent environment that implements it.

The practical consequence is that the MMM workflow now sits in the same workspace as everything else your agent can reach. If your agent already talks to Google Ads, Meta, and TikTok through MCP connections, the budget-optimization skill's output is no longer something a human re-keys into a media plan. It is a file in the same working directory as the tooling that could act on it. We have walked through building that kind of agent before; what is new is that the measurement side now ships in a form that composes with it.

This is also the second standards-based agent layer to land in adtech in as many months, after IAB Tech Lab's AAMP 2.3. Two data points is not a trend, but the direction is consistent: the interface is migrating from the platform's dashboard into the buyer's agent workspace.

The Case That This Changes Nothing

Here is the deflationary read, and it deserves to be taken seriously because most of it is correct.

What has always gated MMM adoption is not writing the Python. It is assembling clean daily time-series data on spend, impressions, and outcomes across platforms that disagree about what a day is. It is convincing a finance team to move budget on the strength of a posterior distribution. It is the compute — Google recommends GPU resources because MCMC sampling is expensive — and the calendar time a full modeling cycle consumes. An agent that authors script.py for you removes perhaps the fourth-largest obstacle on that list.

The skills' own safety design is the strongest evidence for this view. Google's documentation is explicit that the agent pauses for human confirmation before finalizing column mappings, before setting Bayesian priors such as ROI priors and knot spacing, before launching long MCMC jobs, and before executing generated scripts. Read that list again: those checkpoints sit at exactly the decisions that require statistical judgment. If you did not previously know how to set an ROI prior, a dialog asking you to approve one does not fix that. It relocates the problem and adds a click.

Sharpened all the way, the objection is not that the skills are useless. It is that they lower the effort required to produce a model without lowering the expertise required to produce a correct one — which means the honest first-order prediction is more confidently wrong models, produced faster.

What Survives That Objection

One thing, and it is not "MMM is now easy." It is that the cost of running the model again went down.

Under the old workflow, a modeling cycle was a data-science ticket. That made every re-run expensive enough that most organizations built one model a quarter, presented it, and defended it. A model you build once and defend is a political artifact. A model you can rebuild in an afternoon is an instrument.

That distinction would still be academic if there were nothing new to calibrate against — and this is why the timing of the GeoX general-availability launch matters more than the skills themselves. GeoX produces causal evidence from a real geo holdout, and those incrementality results can be fed back into the Meridian model to improve its accuracy. That closes a loop: run an experiment, calibrate the model against it, re-run, re-optimize. Every step in that loop except the experiment just got cheaper to repeat.

The brand-signal addition compounds the same effect. Once branded query volume can enter the model, there are simply more specifications worth trying — and more reason to want re-running to be cheap. It is also the addition most in need of discipline, a point Search Engine Journal made well: branded search rises for reasons that include competitor activity, seasonality, promotions, and news coverage. A lift in branded queries is a signal to model, not a result to report.

The Homework-Grading Problem Gets Slightly Worse

Marketing Dive noted the standing objection to Meridian — that Google is supplying the tool that evaluates Google's media. Open-sourcing it is a real mitigation, not a cosmetic one: the priors are yours, the code is auditable, and GeoX explicitly runs across platforms Google does not sell.

But make model production cheaper and you amplify whatever direction the defaults lean, in both directions. The specific new exposure is that the agent now proposes the priors. A proposed ROI prior that a tired analyst approves at 6pm is functionally an assumption nobody chose, and it will show up downstream as a confident number in a budget deck. This is the same failure mode we described for modeled conversions: the model is not wrong, the interpretation of its outputs as ground truth is.

The mitigation is unglamorous and cheap: version-control the generated script and the prior configuration alongside the report, so that six weeks later you can answer "what did we assume?" without re-deriving it.

How to Adopt This Without Getting Burned

1. Treat the prior checkpoint as the review, not a dialog to clear

The checkpoints are where the modeling decisions actually happen. If nobody on the team can defend the ROI prior on its own terms, the model is not ready to inform a budget decision regardless of how clean the output HTML looks.

2. Run the convergence check before anyone sees a number

meridian-result-visualization generates MCMC health checks specifically so that non-converged models get caught. Make that output a required attachment to any result you circulate. A summary report from a model that did not converge is a well-formatted guess.

3. Your data pipeline is still the bottleneck

None of this touches data assembly. If daily spend and outcome data across your platforms are not already landing somewhere consistent, that is the project — not the modeling. Google's parallel announcement made the Data Manager API universal and based it on the IAB Tech Lab ECAPI standard, which is the more boring and more load-bearing half of the same release.

4. Budget the media cost of the holdout before you promise the experiment

GeoX requires daily time-series data and enough geographic variation to design a valid test, and running one means deliberately suppressing or boosting spend in selected markets. That is a real line item. Agree it before the experiment gets scoped, not after.

5. Read Google's efficacy stats as vendor-reported

The 26% incremental ROAS and 11% Search conversion figures in the announcement are Google internal data, footnoted as such. They are directionally useful and they are not independent evidence. Treat them the way you would treat any other platform-published performance claim.

6. Put the outputs where your buying tooling can reach them

The value of an agent-native measurement layer only shows up if the optimization output can flow into the systems that spend the money. If that handoff is still a person reading a PDF into a spreadsheet, you bought a faster report, not a shorter loop.

The Pattern Worth Naming

Google could have shipped this as a Meridian web app with a chat box. It would have been easier to demo, easier to monetize later, and it would have kept the workflow inside Google's surface. Instead the capability shipped as portable packages on a specification anyone can implement, in a public repository, installable with one command.

Read cynically, that is a distribution play: skills that live in every analyst's agent workspace are stickier than a dashboard nobody logs into. Read plainly, it is still the more useful outcome for the buyer, because portability is the property that lets a measurement tool sit next to a buying tool.

The thing to actually take away is smaller than "agentic MMM has arrived" and more durable. Modeling did not get easier this month. Re-running got cheaper, and calibrating against a real experiment got possible for everyone at once. Teams that convert that into a standing loop — experiment, calibrate, re-run, reallocate — will get something out of this release. Teams that use it to generate the same quarterly deck in less time will not, and will have a nicer-looking artifact to be wrong with.

Sources: Google — Drive profitable growth with new data and measurement tools (September 10, 2026), Meridian documentation — Agentic skills, Marketing Dive — Google upgrades Meridian with agentic AI, upper-funnel capabilities, Search Engine Journal — Google Launches Meridian GeoX Globally, google/meridian on GitHub