Marketing mix modeling software: open-source tools, paid vendors and what they need

Three open-source packages do most of the public MMM work. Each is free to use, and each is a code library rather than an app: someone on your side writes the code, prepares the data and reads the output. Here's what each project's own docs say, as of October 2026:

Robyn's site says it uses Facebook's Prophet library to separate trend, seasonality and holidays, and includes a budget allocator. Google says Meridian is free, that its code is fully visible, and that it supports reach and frequency data and query volume as a control. PyMC-Marketing's docs say it's Apache-2.0 licensed and also covers customer lifetime value models.

Robyn's analyst guide describes three ways to run an MMM: hire a third-party vendor to do it end to end, use a software-as-a-service platform that automates the modeling, or build it in-house. It says vendors are usually the most expensive and in-house takes the most of your own time. Meta also runs an MMM Preferred Partner Program and lists its badged partners on its Measurement Partner site.

A few vendors, in their own words as of October 2026:

• Recast says it's powered by a proprietary Bayesian MMM, designs and analyzes GeoLift experiments, and feeds every lift finding back into the model. It says it's a badged Meta Measurement Partner. See the Recast MMM guide for a fuller summary from getrecast.com.
• Haus describes a Causal MMM updated every week, next to its incrementality experiments.
• Northbeam describes MMM Plus for forecasts and budget scenarios, alongside its multi-touch attribution.

The pages above don't list prices, so ask. Also ask which sales the model uses as its outcome, and whether Amazon is in it.

The docs are specific about data. As of October 2026:

• History. Robyn's guide says a model needs at least two years of weekly data, and more (four to five years) if you only have monthly data. Meridian's docs give two years of weekly data for a geo-level model, three years for a national one, and at least three years if the data is monthly.
• Weekly rows. Both recommend weekly data. Meridian says weekly data balances variation against noise better than daily or monthly.
• Spend that varies. Robyn's guide gives the example of a channel whose spend stayed flat while sales moved: the model has trouble telling what that channel did. It suggests leaving out channels without enough variation or volume.
• Sales and spend that add up. Meridian requires the outcome (sales or conversions) and media to be summable across regions and weeks, and spend for each paid channel.
• Regions, if you have them. Both prefer data by state or metro area. Meridian says geo-level data gives tighter estimates, and national data is an acceptable fallback.

Robyn's guide also estimates how long a first study takes, step by step. Added up, that's 12 to 22 weeks, with data collection alone at 4 to 6.

A model infers cause from history, and history is tangled: brands spend more when they expect to sell more. Calibration feeds the model the result of an experiment, a holdout test that measured one channel directly, so the model's estimate for that channel stays near what the test found. Each tool does it differently:

• Robyn recommends people-based or geo-based experiments, naming Meta Conversion Lift and GeoLift, as ground truth. It adds the gap between model and experiment as a third score when it picks models. You enter each result as a point estimate: start date, end date and incremental response.
• Meridian uses experiment results to set each channel's prior, the model's starting belief about its ROI. Google says there's no single formula for this, and provides a CalibrationBuilder that adjusts for how old, how long and how big the experiment was.
• PyMC-Marketing adds lift test measurements to the model before fitting it, with an add_lift_test_measurements step.

Robyn and Meridian give similar advice on which tests to use. Match the test to the model: same channel, same outcome, a similar period. Robyn says to skip results without enough confidence and prefers longer tests. Meridian adds a subtle point: a model measures a channel against zero spend, so a test that only cut spend measures something different.

A model can only credit a channel for sales in its outcome column. If you sell on your own site and on Amazon, and the model uses site sales alone, it can't see Meta ads selling on Amazon. Use total sales, or model each store as its own outcome.

Calibration has the same catch. A test that only counts site orders will pull the model toward an answer that leaves Amazon out. A geo test that counts orders on both stores by region gives a result that fits a total-sales model.

• Under two years of weekly data. That's below Robyn's and Meridian's stated minimums. Start with a holdout test on your biggest channel.
• Someone on staff writes R or Python. Robyn, Meridian or PyMC-Marketing cost nothing but time. Budget for the data work.
• No analyst, many channels, years of data. A paid vendor or SaaS platform does the work. Ask how it calibrates with experiments and what the outcome includes.

The MMM vs. incrementality testing guide covers when a model is enough and when you need a test. The marginal ROAS calculator shows diminishing returns on your own spend history.

Spillmark runs incrementality tests: geo holdouts that measure how many Amazon sales Meta ads cause. It locks the design before the test starts. Each result is kept with its test dates, measured lift and uncertainty range, which covers what Robyn's and Meridian's calibration steps ask for. The method is on how lift is calculated.