Marketing mix modelling tools: Meridian, Robyn and PyMC-Marketing
For the wider choice of method, see MMM vs MTA vs incrementality tests. We compare the tools on criteria and rank none.
What do MMM tools do?
They take weekly or daily data on spend, sales and other drivers, and estimate how each contributes. Some, such as PyMC-Marketing, model adstock, the carry-over of advertising over time, and saturation, where each extra euro buys less3.
The output is an estimate of return by channel. Robyn, for example, also reports marginal returns for extra spend2. That is marginal ROAS: the return on the next euro rather than the average.
MMM works from observational data, where many activities move together. The Robyn paper says this makes cause and effect hard to separate, and recommends checking results against experiments2.
How do Meridian, Robyn and PyMC-Marketing compare?
| Criterion | Meridian | Robyn | PyMC-Marketing |
|---|---|---|---|
| Who publishes it | Google1 | Meta2 | PyMC Labs, which also sells consulting3 |
| Open source | Yes1 | Yes, MIT licence2 | Yes, Apache-2.0 licence3 |
| Calibration | With incrementality experiments1 | With randomised controlled trials2 | With lift tests3 |
| Also covers | Reach and frequency, budget optimisation1 | Ridge regression, multi-objective model selection2 | Adstock, saturation, budget optimisation3 |
Each cell is what the project says about itself. It is not a ranking, and we do not rate the vendors behind them.
What is MMM calibration?
Calibration means checking the model against an experiment. An experiment measures lift: the difference in outcome between an exposed group and a control group. All three projects describe calibration with experiments123.
The Robyn paper calls calibration with randomised controlled trials the key way Robyn limits bias today2. In our view, a result from a holdout or geo test can pin the model down where the data alone cannot.
How do we implement it?
Start with the decision
Name the budget decision the model should inform. Then check your data history, how much spend varies and whether you can run an experiment.
Name an analyst
Someone builds, checks and re-runs the model. Open source means no licence fee, not no cost.
Calibrate with an experiment
Check the model against a holdout or geo test, so it does not rest on the data alone.
Agree how to read it
Leadership accepts in advance that results may disagree with platform ROAS.
When should you not use it?
Skip it, for now, when:
Our article on whether you need marketing mix modelling goes through the decision. If you want a second pair of eyes first, you can start with a free measurement audit.
What does it cost in effort?
We give no figures. "Open source" means no licence fee, not no cost.
- An analyst. Someone who can build, check and re-run the model.
- Data. Clean spend and sales history, by channel and period.
- Calibration. Experiments take time and budget of their own.
- Trust. Leadership must accept results that may disagree with platform ROAS.
Pitfalls
Treating output as fact
Google's own announcement admits MMMs are not perfect1.
Whose model is it?
A commentary piece on an ad-industry news site asks who benefits when a media owner builds your MMM5. It is a question to ask, not proof of bias.
No calibration
A model with no experiment behind it is a guess with error bars.
Buying the argument from a competitor
A measurement vendor that sells an alternative lists MMM's limits, from data history to channel-level granularity6. Useful list. Read it knowing who wrote it.
Our take: Selective
Our view: Selective. This is our rating, not a market fact or an industry scale.
The tools exist and are open. What varies is whether your brand can feed and read them. A Forrester analyst's blog says going it alone with Meridian takes strong in-house data skills and trust between marketing and finance4. The same post notes that MMM results sometimes contradict campaign-level ROI and ROAS4.
The Robyn paper says an input dataset customarily holds at least two years of weekly data2. For direct-to-consumer businesses, it says daily data over a shorter period can be enough2.
Our view: pick the tool after you know the decision and have data to match; the tool is the smaller question.
FAQ
Meridian or Robyn?
We do not pick. Compare on criteria: who builds and reads the model, what data you have, and how you will calibrate.
Is there a best MMM software, or one for enterprise?
No single answer, and we do not rank tools. Before trusting a "best MMM" list, check whether its author sells one of the tools. Choose by your data, team and decision.
Is open-source MMM free?
Can MMM be calibrated with experiments?
What is an incrementality factor, and does MMM use it?
It is a ratio of incremental conversions measured in a test to conversions a platform reported. See incrementality factor.
Sources
- Meridian (blog Google, 7 mar 2024)Primary source
- Packaging Up Media Mix Modeling: An Introduction to Robyn's Open-Source Approach (arXiv 2403.14674)Primary source
- pymc-labs/pymc-marketing (GitHub README)Primary source
- Is Google's Meridian The Right Open-Source MMM Solution For You? (Forrester)Industry publication
- Meta's Robyn: who really benefits when a platform builds your MMM (ppc.land)Blog or media
- The Rise and Fall of Media Mix Modeling (INCRMNTAL)Vendor
Let’s find what’s actually working.
Next step. Want a view on whether MMM fits your budget decision? Talk to us about your budget decision. Or read about our Measurement & incrementality service.