What is marketing mix modeling (MMM)

MMM starts from data almost every company already has: how much was sold each week and how much was invested in each channel that same week. A regression model is fitted to those series, distributing the variation in sales across the factors that explain it, and advertising is one of them.

The technique is old. It was born in consumer goods, where there was no click to track and television, radio and print could only be judged by their aggregate effect on sales. During the decade of individual tracking it was pushed aside, because it looked less precise than counting conversions one by one.

It is back for a practical reason: individual tracking has narrowed. Consent refusals, third-party cookie blocking and mobile identifier restrictions open gaps that user-based measurement cannot fill. MMM does not have that problem because it never looked at the user. It works with totals, and a total does not depend on anyone accepting a cookie notice.

The open tools published by Google and by Meta have also made entry cheaper, which previously ran through a specialised consultancy by necessity.

The starting point is a table with one row per period, usually a week. The columns hold sales or the chosen business metric, the investment in each channel and the context variables that also move sales: price, promotions, seasonality, public holidays and competitor activity.

Before entering the regression, each channel's investment goes through two transformations. Adstock captures the fact that an ad's effect is not spent on the day it runs and decays over the following weeks. The saturation curve captures the fact that each additional euro returns less than the previous one, so the response flattens once the channel saturates.

With those curves fitted, the model estimates each channel's contribution, its return and the shape of the response to spend. Out of that comes what actually interests a decision maker: how much should move from one channel to another to improve the overall result.

Current models tend to be Bayesian. That allows prior knowledge to be fed in, and in particular results from earlier experiments, which enter as the starting distribution for a channel's return. They also return credible intervals rather than a bare number, which makes visible how much uncertainty remains in each estimate.

MMM answers the question platform analytics cannot answer: what the whole is worth. A Google Ads dashboard knows about Google Ads. A mix model sees television, sponsorship, leaflet distribution, display investment and price inside the same equation, and can therefore compare channels that share no measurement system.

That makes it the natural tool for annual planning and for allocation across media. It also covers what no pixel covers: channels without a click, the pull of brand awareness and the part of the business that happens offline.

The second reason is resilience. A measurement system that depends on consent degrades every time a rule or a browser changes. A model fed with investment and sales totals keeps working the same way, because those figures come from the accounts and from the media plan.

There is a price for that, and it is worth saying before commissioning the project: MMM knows nothing about specific people. It is no use for deciding who to reach tomorrow or for optimising a bid. It is an allocation instrument, not an activation one.

• Gather several years of weekly data before starting and count how many parameters the model will have. With few periods per parameter the estimates come out unstable, however many channels you include.
• Deliberately create variation in investment. A channel that has had the same budget for two years leaves no trace the model can read.
• Include the variables that move sales beyond advertising: price, promotions, distribution, seasonality and competition.
• Calibrate the model with experiments. The return measured in a geographic test enters as prior information and anchors the estimate for that channel.
• Look at the intervals, not only at the estimated value. A channel whose interval runs from almost zero to a lot is not a profitable channel, it is an unmeasured one.
• Refit the model on a fixed schedule and then check whether the previous recommendations held up in reality.

• Reading the result as proven causality. The model separates correlations under the assumption that every variable influencing both investment and sales has been included, and that assumption cannot be checked with any statistical test.
• Raising and lowering every channel at once, always on the same dates. When investments move together the model cannot separate them and allocates credit unreliably.
• Using MMM for activation decisions such as audiences, bids or creatives. That level of detail is not in the data.
• Transplanting the return from an old experiment or another market as prior information without checking whether the period, the audience and the compared scenario match.
• Presenting the model as a substitute for experiments. Without an external reference to calibrate it, an MMM can fit the history very well and still get the allocation wrong.