Michael Kaminsky explains why LLMs can't replace Marketing Mix Models. LLMs predict the next token from patterns; MMMs estimate causal relationships from marketing data. AI can help write MMM code, but validating that the causal estimates are actually correct remains beyond what AI can do.
Chapters
There's a lot of talk about AI and MMM, but where do we really stand heading into 2026? A lot of people ask me, isn't AI just going to do what Recast does? Isn't going to just estimate the MMM and replace what you do? The answer is not really. When most people think of AI, they think of large language models, and they work fundamentally different than MMMs do. Large language models predict the next token, or they generate the next token from a discrete set of possible tokens. MMMs estimate causal relationships from data. It's a fundamentally different problem. They're not doing the same thing. Of course, AI could help you write an MMM or write the code for an MMM, or you could train an MMM using a neural net. But the problem with MMM is not building the model. You can build an MMM in Google Sheets in an hour or two. The problem is validation, knowing that model is right, and AI is not going to help you with the validation piece. It's not going to help you validate the models in practice. It's not going to be able to build a model that is going to pass validation metrics. And that is what really matters with MMM. With MMM, it's not about getting numbers. Anyone can get numbers. It's about getting numbers that are provably right, that are validated to be correct. And we are nowhere near that with AI today.
