All insightsMeasurement & Market Mix Modelling
Consider AI Carefully in Market Mix Modelling
AI is being bolted onto MMM faster than the industry can think it through. Here is what is genuinely useful and what is marketing.
twenty10··6 min read
Every MMM vendor now has "AI" in the deck. Some of it is real and useful. Some of it is the same regression with a new logo. The problem for buyers is telling them apart before signing.
Where AI is genuinely helping
Three areas where machine learning has materially improved modern MMM:
- Bayesian inference at scale. Probabilistic programming frameworks (PyMC, Stan, NumPyro) make it tractable to fit hierarchical Bayesian MMMs with proper uncertainty quantification. This is real. It is also not new - Google's Meridian and Meta's Robyn are both built on it. Treat it as table stakes, not differentiation.
- Automated variable selection and transformation. ML helps explore the combinatorially huge space of which variables to include, which to lag, which to adstock and how heavily. Used well, it cuts model-build time from weeks to days.
- Causal inference methods. Double machine learning, synthetic controls, uplift modelling - these are bringing genuinely new tools to the measurement stack, especially around incrementality.
Where "AI" is mostly marketing
Be sceptical when a vendor promises:
- "AI-powered" attribution that needs no MMM. If they cannot explain how it handles confounders without experimental data, it is correlation in expensive packaging.
- LLMs that "interpret" your MMM. A language model can summarise a chart. It cannot decide whether your TV coefficient is plausible. The interpretation layer in MMM is judgement, not text generation.
- Real-time MMM updated daily by AI. Marketing effects play out over weeks. Daily updates fit noise. Slower is usually better.
- Black-box models that "outperform traditional MMM". If you cannot interrogate why the model recommends what it recommends, no CFO will move £10m on its say-so. Explainability is not optional in this domain.
The right question to ask a vendor
Not "do you use AI". Ask: "show me a model you have built, walk me through every assumption, and explain how you validated it out-of-sample". A team that can do that well - with or without AI - is the team you want. A team that pivots to demoing their dashboard is the team you do not.