What Google Meridian is, in plain terms

Meridian is Google's open-source marketing mix model, written in Python and built on Bayesian inference. It works from aggregated weekly data per market, usually spend, impressions, sales and a set of outside factors, rather than from user-level tracking. Much of its appeal comes from that design. Privacy rules keep tightening and user-level signal keeps degrading, while an aggregate model depends on neither.

The Bayesian part deserves a moment of attention before you commit. You write down what you already believe about a channel as a prior, and the model updates that belief against the data. Teams with real commercial knowledge get to encode it. Teams without it spend the review meeting defending numbers somebody typed in because the notebook wanted a value.

Google also shipped the parts that open-source projects usually leave out: documentation it keeps current, a Python API for retraining models and generating outputs, and a certified partner network of roughly 30 providers for teams who would rather not build in-house.

Why the barrier to marketing mix modeling dropped

Meridian keeps absorbing work that used to sit around the model rather than inside it. It handles non-media variables such as pricing and promotions, which matters more than it sounds. Early MMM builds regularly credited a media burst for a lift that a price cut had produced, and real budgets moved on the strength of it. You can also set contribution priors per channel and choose how each channel's adstock decays, so the specification follows what you know about your own media.

Around the model, Google added a Scenario Planner for people who do not write Python, then put Meridian inside Google Analytics 360, in the same interface as campaign reporting. Whatever you make of the strategy, it takes marketing mix modeling out of the data science corner and puts it in front of the people who own the budget.

Running a model has become the easy part. Getting a CFO to move spend on the strength of one is the same conversation it always was, and no amount of tooling shortens it.

Where Meridian is strong

Meridian treats digital media the way planners do, working with reach and frequency instead of raw impressions. Frequency effects are real and rarely linear, and a model that only sees impressions will flatten them into an average that describes nobody. It also pulls in Google's own media and search query data, which saves setup time when your media is Google-heavy.

Geography is the other strength, and for Belgian advertisers it is the interesting one. Meridian was designed for geo-level modeling, so Flanders and Wallonia can be read separately instead of averaged into one national figure that fits neither. Google extended this with GeoX, which feeds the results of geo experiments into the model. The experiments carry the causal weight there, and the model puts them in context.

Meridian or Robyn?

Meta's Robyn is the obvious alternative, and the two have converged more than the early comparisons suggested.

Meridian is the shorter path for a first model, particularly with a digital-heavy mix and media data already sitting in Google's platforms. The defaults are sensible and you will reach a defensible result faster. Every default is also an opinion, which is fine for as long as you know which opinions you have accepted.

Robyn rewards teams who want to specify everything themselves and who model unusual economics. It asks for someone in-house who enjoys that kind of work, and it punishes teams who do not have one.

In practice the choice comes down to three questions:

• How much of your media already sits inside Google's platforms?

• Who owns and rebuilds this model six months from now?

• How unusual is your business in pricing, seasonality or channel mix?

We have built models with both tools. What decided the outcome each time was the quality of the input data and whether anyone had agreed, before the modeling started, which decision the model was supposed to inform.

What Meridian will not solve

A model that converges is not a model that is right. Someone has to decide if the priors are defensible, if the fit is real or the model has memorized noise, and if the reallocation it recommends is one the business can execute in the next planning cycle. That judgment is the expensive part of MMM, and it has not become cheaper.

Native integration with Google's platforms saves weeks of pipeline work, then hands you back the questions that actually take time. Is the offline sales file clean? Is that hole in the spend data a zero or a missing value? Data preparation is still where most of the effort goes and where most of the risk sits.

Neutrality is the one worth sitting with. Open source gives you transparency, not independence. The easier it becomes to run a marketing mix model inside Google's environment, on Google's data, the more deliberate you have to be about what goes in and which conclusions you accept without a second source.

Attribution deserves its own paragraph, because the two get conflated. A marketing mix model compares channels at an aggregate level over months. Sales attribution follows individual journeys and touchpoints. Teams that run both end up with a picture neither one produces alone.

What this looks like on a real budget

For an energy provider, we modeled TV, radio, press, search and social against B2C sales. The channel ranking turned out to be the least useful output. What the client used were the curves: several channels carried their effect for weeks after the spend stopped, and others had reached the point where each extra euro returned steadily less. On the simulations, reallocating the existing budget pointed to a sales lift above 5% with no additional spend. The full story is in our marketing mix modeling case study.

We built that one on Robyn, which is rather the point. The value came out of the questions we agreed with the client before any modeling started: which channels saturate, what a realistic carryover looks like, and which part of the budget can genuinely move next quarter.

So, is Google Meridian a game-changer?

For access, yes. A team with clean data and a decision waiting on an answer can build a credible marketing mix model without a procurement cycle. Free MMM libraries existed before Meridian, Robyn among them, but Meridian is what made that route the obvious one instead of the adventurous one.

For everything downstream of the model, no. Meridian will not tell you whether your priors are honest or whether the reallocation it proposes survives contact with commercial reality. Those questions need people who have answered them before, and they decide whether an MMM project ships or dies in a notebook.

If you are weighing it up, start smaller than feels satisfying. One market, one decision, two to three years of weekly data, and someone in the room who will argue with the output. Scale after that.

Questions we hear

Is Meridian free?

The code is open source, so the license costs nothing. Your budget goes to the infrastructure to run it, the data preparation, and the expertise to set priors and validate the output. Running Meridian through Google Analytics 360 is a separate conversation, since 360 is a paid enterprise product.

Is Meridian only worth it if most of our media is Google?

No. Meridian models any channel you have data for, including TV, radio, print and other platforms, and Google is explicit that it is meant to be cross-channel. The Google integrations save you setup work rather than deciding whether the tool fits.

How much data does a marketing mix model need?

Two to three years of weekly data per market is a fair rule of thumb, covering media spend, impressions where you have them, sales, and the outside factors that move your category: pricing, promotions, seasonality, weather.

Can we run this without a data science team?

A first model, yes, with the Scenario Planner or a certified partner. Sustaining it is harder, because the model needs rebuilding as your mix, pricing and market move. That is usually the point where data-driven marketing decisions either become a habit or stop.