Google Meridian is an open-source marketing mix modelling framework released by Google in 2024. It uses Bayesian statistics to estimate channel ROI. It is free to use but requires a data scientist or specialist partner to implement it.
Open-source in this context means the underlying code is publicly available and free. Anyone can download it, inspect it, and build models with it. That transparency is one of Meridian's main selling points. You can see exactly how the model is built, rather than trusting a black box produced by a software vendor.
What Meridian actually does
Meridian builds a Bayesian marketing mix model using your media spend, sales, and control variable data. It estimates the contribution of each channel to your revenue, produces ROI estimates with confidence intervals, and runs budget optimisation scenarios. It handles adstock, saturation, and geographic variation within a single modelling framework.
One feature that distinguishes Meridian from earlier tools is its native support for geo-level modelling. Rather than running a single national model, Meridian can model each region separately and pool information across regions. This produces more reliable estimates when your media activity varies by geography.
How Meridian compares to Meta Robyn
- Both are open-source Bayesian MMM frameworks, so the philosophical approach is similar
- Meridian is written in Python and uses the Pymc or JAX probabilistic programming libraries
- Robyn is written in R, though Python wrappers exist
- Meridian has stronger native geo-modelling support out of the box
- Robyn has a larger existing user base and more community documentation as of mid-2026
- Both are actively maintained and receive regular updates
The conflict of interest question
It is a fair observation that Google built a tool for measuring the ROI of advertising channels, including Google's own channels. Google has addressed this by making the model design decisions public and auditable. The priors built into Meridian are documented, and nothing in the code systematically favours Google channels.
Using any tool from a media vendor deserves scrutiny. The mitigation is transparency: with open-source tools, your team or your independent partner can inspect every assumption. Ask your partner to review the default priors in Meridian and confirm they are appropriate for your category before accepting results.
What you need to run Meridian
Running Meridian requires a Python environment, familiarity with Bayesian modelling concepts, and access to computing resources capable of running Monte Carlo sampling (the mathematical process Bayesian models use to estimate probability distributions). This is not a tool you hand to a marketing analyst and expect results by Friday. You need a data scientist with MMM experience.
Most brands access Meridian through a specialist consultancy that builds and maintains the model on their behalf. The consultancy manages the technical complexity, while the marketing team focuses on reviewing outputs and making decisions.
Who should consider Meridian
Meridian is a strong choice for brands that want a transparent, auditable model and are working with a partner who has Python-based MMM experience. It is particularly well suited for brands with significant geographic variation in their media activity. It is less well suited for teams that want a quick build or lack access to experienced modelling expertise.
Is Meridian free to use?
The Meridian code itself is free and open-source. The cost comes from the expertise required to implement it: data scientist time, computing infrastructure, and the analytical work needed to interpret and act on the results.
Does Google see our data if we use Meridian?
No. Meridian runs on your own infrastructure or your partner's infrastructure. You do not send data to Google when you use it.
How does Meridian handle channels that are not Google products?
Meridian is channel-agnostic. You can include any channel for which you have weekly spend data, including TV, radio, out-of-home, print, Meta, TikTok, and others. The model does not privilege Google channels in its structure.