Mix modelling came back, and it is not what you remember

Marketing mix modelling was declared obsolete when digital attribution arrived. Attribution promised per-impression precision; mix modelling offered a coarse quarterly estimate built on aggregates. It was an easy choice and, in hindsight, the wrong one.

Mix modelling came back for a reason that has nothing to do with statistics. It runs on aggregate data you already own, so it does not degrade when a browser changes policy, a platform restricts reporting or a user declines tracking. It is privacy-durable by construction.

What it is genuinely good at

• Allocating a whole budget across channels, including the ones with no digital signal at all.
• Estimating diminishing returns, which tells you not just where to spend but where the next euro stops working.
• Capturing carryover, so brand investment is not scored as if its effect ended on the day it ran.
• Surviving measurement regime changes, because nothing it needs can be taken away by a policy update.

What it cannot do, whatever a vendor says

It cannot see a single creative. It cannot tell you which audience segment responded. It cannot produce a reliable answer with eighteen months of data and six channels that all moved together, because it has no variation to learn from. And it cannot, on its own, distinguish correlation from causation: two channels that were always budgeted together will never be separable.

A mix model that has never been checked against an experiment is a well-dressed opinion.

The part that makes it trustworthy

Calibration. You run experiments on two or three channels, obtain causal estimates, and use them as priors or as validation for the model. Where the model agrees with the experiment, you trust its estimates for the channels you could not test. Where it disagrees, you have found either a modelling error or a more interesting problem.

Without that loop, mix modelling is a regression with good marketing. With it, it is the closest thing this industry has to a map.