Media Mix Modeling (MMM)

By the end of this article, you should be able to: say what a media mix model estimates and from what data, explain carryover and diminishing returns, allocate a budget on marginal rather than average return, and explain why an MMM needs experiments to be trusted.

What an MMM does

A media mix model estimates how much each marketing channel (TV, search, social, video) contributed to sales, using aggregate data: sales and spend per channel, week by week, sometimes by region. It also controls for things that move sales without advertising, such as price, seasonality, holidays and the economy.

Teams use it when they can't follow individual customers from ad to purchase. That's increasingly common: TV and billboards were never trackable, and privacy changes have made online tracking patchy too. An MMM needs only the totals.

It answers three questions:

What did each channel contribute last year?
What return did each channel earn?
Where should next year's budget go?

The model at its core is a regression of sales on transformed spend. Two transformations matter, because advertising doesn't work linearly or instantly.

Advertising carries over

Money spent this week keeps working next week. People remember an ad, or act on it later. So a model shouldn't use raw weekly spend $x_t$ . It uses adstock, spend with a memory:

$\text{adstock}_t = x_t + \lambda \cdot \text{adstock}_{t-1}$

Here $\lambda$ (lambda), between 0 and 1, is the share of last week's effect that carries into this week. With $\lambda = 0.5$ , half of the effect carries over, then a quarter, then an eighth. Over all the weeks it adds up to $1/(1-\lambda)$ times the first week's effect, so with $\lambda = 0.5$ a dollar works twice as hard in total as it seems to in its own week. A model without adstock credits that later effect to whatever else happened in the following weeks.