Meridian (Google), PyMC-Marketing (PyMC Labs) and Robyn (Meta) are the three leading open-source engines for marketing mix modelling. They share the same goal, quantifying the impact of marketing on sales, but they follow three very different philosophies. Two are Bayesian, one is based on ridge regression. Here are the key differences.
Three engines, three philosophies
- Meridian (Google): Bayesian hierarchical regression. Samples the posterior with MCMC (TensorFlow Probability).
- PyMC-Marketing (PyMC Labs): Bayesian regression. Samples the posterior with MCMC (NUTS, four chains).
- Robyn (Meta): Ridge regression combined with an evolutionary search. Nevergrad explores thousands of candidate models.
In short: Meridian and PyMC infer a distribution, Robyn searches for good models.

How are ad-stock and saturation chosen?
- Meridian: estimated as posteriors, using Meridian’s own prior families for decay, half-saturation and slope.
- PyMC: estimated; your configured priors become Bayesian priors and the parameters get full posteriors.
- Robyn: searched; an evolutionary optimiser explores the hyperparameters inside the ranges you set.
In every case, ranges and priors act as guardrails.

What does uncertainty look like?
Meridian and PyMC deliver a full posterior: every output, from contributions and ROI to response curves, comes with a probability distribution and credible intervals. Robyn delivers a model ensemble: a Pareto front of models where none is better on both prediction error and decomposition plausibility, and one model is auto-selected. When lift tests are used for calibration, they are added as a third objective.

Constraints and business priors
- Meridian: paid media effects are parameterised through ROI or contribution priors, with one global prior type.
- PyMC: beliefs are genuine priors on coefficients and transformations; lift tests enter as likelihood terms (national).
- Robyn: sign constraints only (positive, negative or unconstrained), no numeric coefficient bounds; lift tests serve as a calibration objective.
Priors versus sign constraints is one of the key practical differences between the engines.

Trend, seasonality and baseline
- Meridian: a spline-knot baseline, a smooth time-varying baseline that absorbs trend and seasonality, with its own level per geo.
- PyMC: Fourier seasonality (two yearly harmonics by default) plus a constant level. There is no native trend, so drift needs an engineered trend column.
- Robyn: Prophet decomposes sales; trend and season enter the ridge regression, plus a non-negative intercept.
This also means that baselines are not directly comparable across engines.

Geo and panel data
- Meridian: geo-hierarchical by design. Media, control and baseline parameters are market-indexed and shrunk toward population-level distributions.
- PyMC: hierarchical partial pooling. Market coefficients are drawn from a shared distribution, each market with its own intercept and seasonality.
- Robyn: national models only, run on summed data. For panel or market-level work, PyMC or Meridian are the better fit.
With hierarchical models, sparse markets borrow strength from the panel.

Under the hood
| Meridian | PyMC | Robyn | |
|---|---|---|---|
| Order | Saturation → ad-stock | Ad-stock → saturation | Ad-stock → saturation |
| Ad-stock options | Geometric, binomial | 6: geometric, delayed, lagged, Weibull PDF/CDF, binomial | Geometric, Weibull PDF/CDF |
| Saturation options | Hill, AdBudg | 8: Hill, AdBudg, Michaelis-Menten, power, tanh, arctan, negative exponential, log1p | Hill, AdBudg |
| Scaling | Own scheme | Normalised space | Raw units |
Notably, only Meridian applies saturation before ad-stock.
Which philosophy fits?
- Meridian: geo-level Bayesian modelling with priors in ROI terms, hierarchical, with a flexible baseline.
- PyMC: maximum Bayesian flexibility, the richest menu of transformations, hierarchical panels and lift-test likelihoods.
- Robyn: national-level model ensembles and fast exploration via ridge regression and a Pareto front.
Same goal, three very different routes. With MMM Labs you can run all three side by side on one platform and compare the results directly.

Want to know more?
We help brands and agencies choose, build and run the right measurement approach. Let’s talk about your questions in a 30-minute call.
