Google Meridian: What You Need to Know About Google's Open-Source MMM

In early 2024, Google open-sourced Meridian, its internal Marketing Mix Modeling framework. This was a big deal. Google had been using MMM internally for years to help its largest advertisers measure cross-channel performance. Now anyone can use it.

Meridian is a Bayesian MMM built on top of Python and JAX (Google's machine learning library). It is designed for production use, not just academic research.

What Makes Meridian Different

There are three main open-source MMM frameworks right now: Google Meridian, Meta Robyn, and Google's earlier LightweightMMM (now deprecated in favor of Meridian).

Robyn uses a frequentist approach with Ridge regression and multi-objective optimization. It is fast and produces good results, but it does not give you proper uncertainty estimates. When Robyn says "TV ROAS is 2.1x," there is no confidence interval attached.

Meridian is fully Bayesian. Every parameter comes with a posterior distribution, so you know how confident the model is. For budget decisions involving millions of dollars, that uncertainty quantification is not optional.

Meridian also has native support for reach and frequency data from Google platforms, hierarchical geo-level modeling, and a built-in budget optimizer that accounts for saturation curves.

Who Meridian Is For

Meridian is open source, but "open source" does not mean "easy to use." To run Meridian yourself, you need:

• A data scientist comfortable with Bayesian statistics and Python
• Infrastructure to run MCMC sampling (GPU recommended, 4-8 hours per model run)
• Data engineering to extract, clean, and format your marketing data weekly
• Domain knowledge to set reasonable priors and validate results

Most companies with under $10M in annual ad spend do not have this in-house. That is the gap between "Meridian is free" and "Meridian is accessible."

The Technical Requirements

Meridian expects your data in a specific format: a panel dataset with geo-level or national-level observations at weekly granularity. Each row is one geography-week combination with columns for spend, impressions, revenue, and control variables.

Model fitting uses MCMC (Markov Chain Monte Carlo) sampling via NumPyro. A typical run with 4 chains and 1,000 warmup + 1,000 sampling iterations takes 2-6 hours on a GPU. On CPU, expect 8-24 hours.

After fitting, you get posterior distributions for every parameter: channel contributions, adstock decay rates, saturation curves, and more. Interpreting these requires statistical literacy.

How Spendmix Uses Meridian

Spendmix runs Meridian under the hood but removes the technical barriers. You connect your ad platforms through automated integrations, configure your model through a visual interface, and get results as an interactive report.

No Python. No MCMC tuning. No data engineering. You get the same Bayesian rigor as a $100K consulting engagement, at a fraction of the cost and timeline.

See what the output looks like in our interactive demo report.