Google Meridian vs Meta Robyn for Marketing Mix Modeling

Google Meridian vs Meta Robyn is a choice between two open source marketing mix modeling libraries that reach their estimates by different routes, and three facts about your own team settle it: the language your analysts work in, whether your sales and media data is split by region, and how much budget sits in channels that report reach and frequency. Our position, from the two projects' documentation and repositories as read on 7 October 2026, is that Meridian is the better default for a Python team with regional data, while Robyn still suits an R team modeling one national market on mostly digital spend. Both estimate the incremental effect of media from aggregate history, the question our guide to what incrementality is approaches from the experiment side.

How do the two compare side by side?

A Google Meridian vs Meta Robyn comparison for marketing mix modeling rests on seven attributes, and every cell below comes from the projects' own pages.

The Meridian README states that "Python 3.11-3.13 is required to use Meridian" and recommends at least one GPU. The Robyn README says the package is available in R and Python, then adds: "Please note that the current Python version is a LLM-translated Beta version and might encounter bugs." That Python version lists R and the glmnet package as prerequisites, and the R package in turn requires the Python library Nevergrad. CRAN lists Robyn 3.12.1, published on 2 July 2025, under an "MIT + file LICENSE" license.

What does Meridian include out of the box?

Google Meridian's open source MMM features, as the introduction in the Meridian documentation lists them, are a design for causal inference, priors that carry your business knowledge, control variables such as Google Query Volume, geo-level modeling, a reach and frequency methodology, and optimization with scenario planning.

Reach and frequency is the feature for which we found no close equivalent in Robyn. The reach and frequency page of the Meridian documentation says the option applies to "any channel with reach and frequency data", that the MMM Data Platform provides that data for YouTube, and that it must sit at the same geo and time granularity as the KPI. Robyn's nearest feature is the curve calibrator in the Robyn documentation, marked beta, whose "first use case is to calibrate the response saturation curve using cumulative reach and frequency data as input". The same page, last updated in December 2024, describes a reach and frequency allocator as "currently a prototype and not yet available".

How do the two projects describe themselves?

Google Meridian's official documentation calls it an open-source marketing mix modeling framework built by Google, and a Meridian and Robyn comparison should set Meta's own description beside it. The Robyn README calls the package "an experimental, semi-automated and open-sourced Marketing Mix Modeling (MMM) package from Meta Marketing Science" and says it is "especially suitable for digital and direct response advertisers with rich data sources". Meridian's README routes support through GitHub, where "The Meridian team responds to these questions weekly in batches" and pull requests are "very difficult for us to merge". Release dates are the other signal: the Meridian repository's citation block gives version 2.1.0 and the year 2026, while Robyn's features documentation says it was updated in December 2024 for version 3.12.0 and the CRAN release above dates from July 2025. We found no newer Robyn release on the pages we read, and for a model you expect to refresh for years that gap deserves weight.

Bayesian sampling or ridge regression with an evolutionary search?

Meridian fits one Bayesian model and returns a posterior distribution, so every ROI arrives with a credible interval. Its README warns that the No U Turn Sampler "can be compute intensive", which is why the GPU recommendation exists. Robyn's documentation says it "uses the L2-form of the regularisation, also known as Ridge regression", and searches the adstock, saturation and penalty hyperparameters with "TwoPointsDE, an evolutionary algorithm from Nevergrad". The search minimizes several errors at once: prediction error (NRMSE), a business error called DECOMP.RSSD that "represents the difference between share of spend and share of effect for paid media variables", and a calibration error when experiments are supplied. Robyn recommends a minimum of 2000 iterations in each of at least 5 trials, returns a set of Pareto-optimal candidates, and bootstraps ROAS or CPA within clusters of similar models for a 95% interval.

The practical difference is where analyst judgment enters. With Robyn someone picks among candidate models after the run, and DECOMP.RSSD favors candidates whose effect shares resemble the current spend shares, a metric Robyn's own documentation acknowledges is controversial. With Meridian the judgment moves into the priors, set before the run and visible to a reviewer. We prefer the second arrangement for any model a finance team will question.

How does each one use incrementality tests?

The calibration page of the Meridian documentation defines calibration as "the process of using experiment results and other domain knowledge to set channel-specific priors", cautions that "There is no single formula to translate an experiment result into a prior", and offers a CalibrationBuilder that "automates the translation of incrementality experiments into calibrated ROI priors". Robyn treats the experiment as a target: the search minimizes MAPE.LIFT alongside its other errors, and the Robyn Analyst's Guide notes that "currently, Robyn only accepts point-estimate as a calibration input".

A prior has a standard deviation, so a noisy test can enter as weak evidence, while a point estimate enters at face value. Our guides to running a quarterly geo-holdout test and analyzing an incrementality test result cover the testing side, and the multiplier method handles the separate job of correcting platform-reported numbers.

One hypothetical dataset under each project's own rule

Take a hypothetical retailer, invented for illustration: 104 weeks of history, 8 paid channels, 4 control variables, and sales and media available for 40 regions. The amount of data page in the Meridian documentation counts data points per parameter (channels plus controls plus knots, plus one per geo beyond a baseline in a geo model, ignoring Adstock and Hill parameters), and the model specification in the Meridian documentation sets the default knots: one for a national model, one per time point when there are multiple geos. Robyn's guide gives two rules of thumb: a recommended ratio of "1 independent variable : 10 observations", and rows at about 7 to 10 times the number of columns.

Google's own example calls four data points per parameter "too low to estimate the model reliably" and says roughly 15 might yield "some directional information", so the national Meridian model at 8.0 is thin. The Robyn model at 8.7 falls inside its guide's 7 to 10 range yet short of the 10 observations per variable the guide recommends, which leaves little room to add variables. The regional data changes the picture for Meridian only. Google explains that the geo hierarchy shares information across geos, and because "Meridian uses partial pooling (hierarchical modeling)" the effective figure lies between the strict and lenient views, here between 6.7 and 26.8. On national data both libraries face the same squeeze, and the remedy is fewer channels and controls.

Which one should you pick?

Pick by team skills first. An R team gets a CRAN package in Robyn, and a Python team should choose Meridian over Robyn's Python beta. Data shape comes next: with sales and media by region, Meridian is the one whose documentation describes modeling the regions together, and with national data alone Robyn's Prophet decomposition and semi-automated search make a first model less work to specify. Channel mix is the tiebreaker. Heavy YouTube or other spend with reach and frequency data, or a program where organic brand interest needs a control such as Google Query Volume, points to Meridian; a mostly digital, direct response mix is the case Meta describes Robyn as especially suitable for.

The optimizers are close enough that we would not choose on them. The budget scenarios in the Meridian documentation cover a fixed budget, with each channel held between 70% and 130% of its starting spend by default, and a flexible budget to a target ROI or marginal ROI. Robyn's allocator offers maximum response or a target ROAS or CPA.

Both are open source, so the budget a paid platform would absorb can go to analyst time, compute and the experiments that calibrate the model. That work sits with our data and analytics service, and the Attribution Doctor is a further resource for the attribution side of your measurement.

Sources

• Google, Meridian repository and README
• Google, An introduction to Meridian
• Google, Model specification
• Google, Reach and frequency
• Google, Calibrate treatment priors
• Google, Budget optimization scenarios
• Google, Amount of data needed
• Meta Marketing Science, Robyn repository and README
• Meta Marketing Science, Robyn Key Features
• Meta Marketing Science, An Analyst's Guide to MMM
• CRAN, Robyn package page