What is media mix modeling (MMM)?

Why a 1960s technique came back

MMM predates digital advertising. It was built for television, where nobody could observe an individual viewer, so the only option was to regress aggregate outcomes against aggregate spend.

Digital made user-level tracking possible and MMM went out of fashion for two decades. Identifier deprecation reversed that. A method that never needed to identify anybody is unaffected by the loss of identifiers, which is why every large advertiser has rebuilt an MMM capability since 2022.

What MMM needs

• Long history. Two to three years of weekly data is the usual minimum; a model on six months will fit noise.
• Genuine variation in spend. A channel held at a constant budget contributes no information — the model cannot separate its effect from the intercept.
• Control variables. Seasonality, price, promotion, distribution, competitor activity. Omit them and their effect is attributed to whatever media correlated with them.
• Aggregate outcomes, from your own system of record rather than from platform reporting.

What it is good and bad at

MMM is good at the strategic question: roughly how much should go to each channel, including offline and channels with no tracking at all. It handles diminishing returns and carryover, which user-level attribution ignores entirely, and it is immune to identifier loss and to platform self-reporting.

It is bad at everything tactical. It cannot tell you which creative, which audience or which placement worked, and its estimates carry wide uncertainty that is routinely discarded when results reach a slide. It is also easy to fool: a model without a control for a price change will happily credit the media that ran alongside it.

MMM and AI media

MMM is currently the most practical way to see AI media's contribution at a total level, because it does not require observing the path — only the spend and the outcome. Where AI spend is large enough and has varied enough, it will appear.

The obstacle is usually scale rather than method: AI budgets are still small relative to the noise in a total-business outcome, so the model cannot resolve them. Until they are material, a geo-based incrementality test is the sharper instrument, with MMM picking the channel up later as it grows.