Marketing mix modelling

Attribution follows individuals. MMM does not look at individuals at all. It takes aggregate history — spend per channel per period, outcomes per period, plus whatever else moved, like price, season, promotion and competitor activity — and estimates how much each input contributed.

Because it never touches user-level data, it is untroubled by cookie loss, consent refusal and tracking prevention. That is why a technique from the pre-digital era came back: it was the thing least damaged by everything that broke.

What it is good at. Channels attribution cannot see at all — television, radio, out-of-home, sponsorship. Long-horizon effects. Diminishing returns, since a model fitted across a range of spend levels can say something about what the next unit of spend is likely to do.

What it is not. MMM is a model, not a measurement. Change the specification, the time window or the control variables and the answer moves. It cannot tell you which creative worked, which audience, or what to do tomorrow — the granularity is not there.

How the three fit together. Attribution is fast, granular and biased toward whatever sits nearest the purchase. MMM is slow, coarse and unbiased by tracking loss. A holdout test is the only one that produces causal evidence, and it is the most expensive. Serious teams use them to check each other rather than picking one — and when MMM and attribution disagree, the disagreement is the finding, not a fault to be reconciled away.