What is marketing-mix modelling (MMM)?
How does MMM work?
You feed a model your past: spend by channel, sales, prices, promotions, the time of year. The model looks for patterns between them and estimates what each channel added to sales.
Two ideas often appear inside it1. Adstock captures the lagged effect of advertising on sales. Saturation curves build in diminishing returns: each extra euro in a channel buys less than the one before.
Because the data is aggregated, MMM does not need cookies or user IDs. Not needing user-level data is one reason for its comeback as privacy rules tighten1. It also cannot tell you about one customer or one ad.
What is MMM good and bad at?
An agency measurement guide gives MMM the role of the wider cross-channel view, next to attribution for speed and tests for proof2. That is a fair split.
MMM finds patterns. It does not prove cause. A paper on Robyn, an open-source MMM package, co-written by academics and members of Meta's marketing science team, recommends validating MMM results with experiments and calls calibration with randomised controlled trials the key bias fix in Robyn1. A vendor that sells MMM lists coarse granularity and omitted-variable bias among the pitfalls of traditional models3. Omitted-variable bias means something that moved sales was left out, so a channel gets credit for it.
For ZFD: In our view, an MMM output is a hypothesis about the mix. A test, where one is affordable, is how to check it.
To see how MMM fits with the other methods, read MMM vs MTA vs incrementality tests. Whether it makes sense at your size is covered in do I need marketing mix modelling. The software side is in marketing mix modelling tools.
Easily confused with
Marketing-mix modelling vs Multi-touch attribution (MTA)
Splits credit across touchpoints for each conversion. It is fast and granular. It does not show what would have happened without the ads.
Marketing-mix modelling vs Incrementality
The extra sales a marketing activity caused. A test can measure it. MMM can estimate it, with more assumptions.
Marketing-mix modelling vs Marginal ROAS
The return on the next euro, not the average. MMM is one way to estimate it.
FAQ
Is MMM the same as media mix modelling?
Yes. Different names, same method. We write marketing-mix modelling, with the UK spelling.
Does MMM replace testing?
No. MMM works from correlations in observational data, and the Robyn paper treats calibration with experiments as its main defence against misleading results1.
Do I need user-level tracking for MMM?
No. It works on aggregated data, which is one reason interest has grown as privacy changes limit tracking1.
Let’s find what’s actually working.
Next step. Start with what your numbers say now. Use the Is your ROAS real? calculator.