If you’ve decided to build a Marketing Mix Model (MMM), an important first step is to choose what platform you want to use. There are quite a few options, ranging from simple to complex, with different benefits to each
You’ve probably come across these libraries already: Google’s Meridian, Meta’s Robyn, PyMC Marketing and a handful of others.
This post walks through these main contenders, what they do well, where they fall short, and factors you should consider before choosing.
What is open-source MMM?
Before we get into the platforms, a quick clarification on what we mean by an open-source Marketing Mix Model. These are coding libraries and frameworks. You write the model yourself (or adapt one), run it on your own infrastructure, and retain full control over the inputs, assumptions, and outputs. This is categorically different from self-serve tools, where the model runs behind the scenes, and you interact through a dashboard.
Open-source gives you transparency. You can inspect every prior, every transformation, every assumption baked into the model. That transparency matters when you’re trying to convince stakeholders that your channel attribution is credible, or when a client’s data has intricacies that a one-size-fits-all solution would paper over.
The trade-off is that it requires statistical knowledge to implement well and is more time-intensive. But done right, open-source MMM tends to be more reliable than its automated rival, as you can create something truly bespoke for your needs.
The main platforms
Google’s Meridian
Google publicly launched Meridian in early 2025, and it is a fully Bayesian framework.
Meridian’s standout feature set includes geo-level hierarchical modelling, reach and frequency inputs, ROI priors, and a time-varying intercept to capture gradual shifts in baseline. These are all genuinely useful capabilities.
It also comes with an impressive dashboard output and benefits from Google’s documentation and support infrastructure.
Meridian’s limitations include a lack of support for time-varying media parameters, and prior customisation is constrained to a fully flexible Bayesian system. For businesses with unusual structures or a strong prior reason to model a channel differently, those constraints can be frustrating.
When to use Meridian: Choose Meridian when you want a fully Bayesian framework with clear dashboard outputs. It is ideal if you need features like reach and frequency inputs or geo-level hierarchical modelling, but do not require time-varying media parameters or highly flexible custom priors
Meta’s Robyn
Meta’s Robyn has built up a large community of users since its launch. It was one of the first widely adopted open-source MMM packages and, for many organisations, it remains the default starting point.
Robyn uses the simpler Frequentist method instead of a more advanced Bayesian approach. It automates the complex process of tuning a model by quickly cycling through thousands of configurations (such as adstock decay and saturation) and automatically selecting a set of good models. This is very helpful when you need a quick readout of your ad performance and don’t have the time for manual tuning.
Robyn also produces clean one-page output summaries, including contributions, saturation curves, and a budget allocator, which are designed with clients in mind rather than a data science team.
Robyn’s trade-offs are well-known. It uses Ridge regression rather than a Bayesian approach, which gives you rougher uncertainty estimates. For ranking channels, this often isn’t a problem, but it is less ideal for justifying budget decisions. It can also calibrate against experimental results, but only as a point estimate that nudges the model in the right direction. It does not bake the result directly into the model itself.
When to use Robyn: Robyn is best when you need a quick readout of ad performance and lack the time for manual tuning. It is a great starting point for teams with lower statistical expertise who benefit from automated model selection and clean, client-ready summary outputs designed for stakeholders rather than data scientists.
PyMC Marketing
PyMC Marketing is the MMM module built on top of PyMC, a mature probabilistic programming library for Python. It utilises a fully Bayesian model, allowing for advanced capabilities.
For example, you can define prior beliefs about your channel effects, run MCMC (Markov Chain Monte Carlo) sampling to get posterior distributions, and end up with uncertainty estimates on everything, including contributions, ROAS, saturation curves and adstock decay.
The key strength of PyMC Marketing is flexibility. Unlike some other platforms, it doesn’t impose a fixed model structure. You can define your own adstock transformations, build hierarchical structures across geographies or product lines, incorporate incrementality test results into the model and extend the model however your client’s business logic demands. There are a few limitations.
And if PyMC Marketing’s built-in API still doesn’t fit, you can step outside it entirely and build a fully custom PyMC model from scratch, defining your own model structure with no constraints at all. Neither Robyn nor Meridian offer anything comparable.
It does require more statistical literacy than some alternatives, but the PyMC Marketing website offers detailed documentation and a growing library of examples, making it increasingly accessible.
PyMC Marketing has no ties to any advertising platform. For clients who are sensitive to the idea of their MMM being built by the same company running their paid search, that independence matters.
When to use PyMC: When you need model flexibility, and you have the statistical literacy and Python knowledge to make use of the flexibility of PyMC. It can build completely bespoke models that, with enough time investment, will give you the best indication of marketing performance possible.
Platform comparison
| PyMC Marketing | Meta Robyn | Google Meridian | |
|---|---|---|---|
| Language | Python | R | Python |
| Inference method | Full Bayesian | Ridge regression + optimisation | Full Bayesian |
| Uncertainty quantification | Full posterior distributions | Confidence intervals | Full posterior distributions |
| Customisable priors | Highly flexible | N/A | Limited |
| Geo-level modelling | Yes | No | Yes |
| Time-variant media parameters | Yes | No | No |
| Incrementality Test Calibration | Yes – custom priors or through likelihood | Yes – through optimisation | Yes – ROI priors |
| Budget optimisation | Yes | Yes | Yes |
| Model flexibility | Maximum | Moderate | Moderate |
| Active development | Yes | Yes | Yes |
| Statistical expertise required | High | Low–moderate | Moderate |
How we choose a platform for our clients
We’re platform-agnostic. When we start working with a client, we work together to define the objectives of the MMM projects and identify the channels to be considered. Using this information, we discuss with all stakeholders and decide the preferred route forward.
If most of a client’s budget is in the Google ecosystem and they are keen to know how reach & frequency affect results, then Meridian might be the best option.
If they have many different channels and plenty of test results to feed into a model, then PyMC might be best, as we can create something flexible enough and truly bespoke.
If the business is newer and has few channels, then Robyn can be appropriate if a quick readout is desired and there is no need for the more advanced features.
