GuideMeasurement and attribution
MMM (marketing mix modelling)
Definition
A statistical model that relates marketing spend, channel by channel, together with outside factors like season and price, to revenue over time. It answers budget questions: what each channel contributes, and what happens if money moves between them.
How it works
Underneath sits regression: the model estimates how revenue moved when each channel’s spend moved, with season and price accounted for. Because it reads weekly totals of spend and revenue, cookie loss and consent banners do not touch it, which is much of its current appeal. It needs a long history, years of it, and it needs spend that has moved within that history. A budget held steady gives the model nothing to learn from; the channels cannot be told apart, and the output turns into an echo of the modeller’s assumptions.
What to watch for
Treat a vendor’s model as a proposal about your business and ask how it earns belief. A fit chart drawn on the data the model was trained on proves memory; it says nothing yet about what the model can predict. The test that counts is prediction: hold recent months out of the training data and ask for the forecast before revealing the answer. Ask what the model assumes, because a Bayesian model fed strong priors and thin data hands the priors back. And the validation that settles arguments is action: when the model calls a channel overfed, cut the channel and watch revenue. That is a holdout test run with real money.
The question you ask
“If we hold the last three months out, how close does the model get to what happened?”