Marketing mix modelling is back at the centre of the CFO conversation- twenty10Multi-touch attribution alone can't survive cookie deprecation: triangulation is the new default- twenty10Geo lift tests and incrementality studies are calibrating the next generation of MMM- twenty10Clean rooms + MMM: how retailers and CPGs are closing the measurement gap- twenty10The UK MMM landscape: who's serious, who's selling a dashboard- twenty10Comparing marketing mix modelling providers in the UK- twenty10Brand vs performance: measuring long-term equity effects in a P&L your CFO can read- twenty10Marketing mix modelling is back at the centre of the CFO conversation- twenty10Multi-touch attribution alone can't survive cookie deprecation: triangulation is the new default- twenty10Geo lift tests and incrementality studies are calibrating the next generation of MMM- twenty10Clean rooms + MMM: how retailers and CPGs are closing the measurement gap- twenty10The UK MMM landscape: who's serious, who's selling a dashboard- twenty10Comparing marketing mix modelling providers in the UK- twenty10Brand vs performance: measuring long-term equity effects in a P&L your CFO can read- twenty10
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:

  1. 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.
  2. 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.
  3. 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.

FAQ

Frequently asked questions

Where is AI genuinely useful in MMM?
Bayesian inference at scale via probabilistic programming (PyMC, Stan, NumPyro), automated variable selection and transformation that cuts build time from weeks to days, and causal-inference methods like double machine learning and synthetic controls used alongside incrementality testing.
Which AI MMM claims should buyers be sceptical of?
AI-powered attribution that needs no MMM, LLMs that 'interpret' models, real-time daily MMM updates (which fit noise because marketing effects play out over weeks), and black-box models that outperform on paper but cannot be interrogated for a CFO sign-off.
What is the right question to ask an MMM vendor?
Not 'do you use AI'. Ask them to show a model they have built, walk through every assumption and explain how they validated it out-of-sample. Teams that pivot to demoing a dashboard instead are the ones to avoid.

More in Measurement & Market Mix Modelling