Marketing Mix ModellingSeeing the whole budget at once

What MMM is, and why it has returned#

Marketing mix modelling is older than digital marketing. Consumer goods companies used it for decades to decide how to split money between television, print, promotions and distribution. For a while it fell out of fashion, because click-level attribution seemed to offer something better: precise, daily, person-level credit.

That promise has weakened. Consent choices, browser restrictions and walled gardens mean person-level tracking now sees less of the journey than it once did. MMM never needed it. It works on aggregates (weekly spend by channel, weekly sales, prices, seasons, competitor activity), which is why it has become central again to how serious advertisers plan budgets.

When you cannot follow every customer, study the tide instead of the swimmers.

How it works, in plain terms#

An MMM is a regression model with domain-specific adjustments. It looks at how sales moved over time and asks which inputs best explain the movement. If sales tend to rise a week or two after television bursts, and that pattern holds after accounting for seasonality and price, the model attributes part of the rise to television.

Two adjustments make it a marketing model rather than a generic regression. Adstock (or carryover) captures the fact that advertising keeps working after it stops: a campaign this week still influences sales next week, decaying over time. Saturation captures diminishing returns: the first rupee spent on a channel does more than the last, because audiences get saturated.

Together these produce response curves for each channel: how incremental sales change as spend rises. Response curves are the most useful output of an MMM, because they show where the next unit of budget would do the most good.

What you need before you start#

MMM is data-hungry in a specific way. It does not need user-level data, but it does need a long, consistent history with enough variation for the model to separate one channel's effect from another's. Practitioners commonly look for a couple of years or more of weekly data, though the right amount depends on how much your spend and sales vary.

The point about variation is critical and often missed. If a channel's spend was flat every week for two years, the model has nothing to learn from. Deliberately varying spend, by region or by period, makes future models far more reliable.

What MMM is good at#

MMM answers questions attribution cannot. How much do television, outdoor and print contribute? What would happen if we moved budget from one channel to another? Where are we past the point of diminishing returns? How much of our sales is base demand that marketing did not create this period? These are board-level questions, which is why MMM tends to sit with finance and strategy as much as marketing.

It is also privacy-resilient by design. Because it uses no personal data, it is unaffected by cookie restrictions and consent rates, which makes it a natural part of any cookieless marketing measurement strategy.

Where MMM misleads#

MMM is a model of correlation over time, made more credible by structure and priors. It can mislead when channels move together (if search and social always rise in the same weeks, the model struggles to tell them apart), when important factors are missing from the data, or when the modeller's assumptions about carryover and saturation are wrong.

MMM is also slow and coarse. It cannot tell you which ad or keyword works, and it typically updates monthly or quarterly. Treat its outputs as strategic guidance, not as daily optimisation signals.

Calibration: making the model honest#

The strongest modern practice is to calibrate MMM with experiments. If a geo test shows that a channel produced a certain incremental lift, that result can be fed into the model as prior information or used to check its estimates. Where model and experiment disagree, the experiment usually deserves more weight for that channel and period.

Read more on running such tests in our guide to incrementality testing, and on how MMM fits beside click-based credit in marketing attribution.

Turning outputs into budget decisions#

The headline output most people look at is average return by channel. It is the least useful number in the model. Average return tells you how a channel performed across all of its spend; budget decisions are about the next unit of spend, which is governed by the marginal return at the current position on the response curve.

A channel can have an excellent average return and a poor marginal one, because it is already saturated. Another can look mediocre on average and still be the best place for extra money because it sits on the steep part of its curve. Good MMM reporting therefore shows each channel's current position on its curve and the expected effect of moving budget up or down by a realistic amount.

Make reallocations in steps, not leaps. Models are least reliable outside the range of spend they have seen, so a channel that has never been funded at double its current level cannot be confidently predicted at that level. Move gradually, measure, and refresh the model with the new data.

Build, buy or open source#

There are three routes. Specialist vendors build and maintain models, often with useful benchmarks and tooling, at a cost. Open-source libraries, including ones released by large ad platforms such as Meta's Robyn and Google's Meridian, let in-house analysts build credible models; check each project's current documentation and be aware that a platform-built tool is still a tool to be validated, not an oracle. In-house bespoke models offer full control and require scarce skills.

Whichever route you choose, insist on three things: uncertainty ranges on every estimate, transparency about assumptions, and out-of-sample validation (does the model predict periods it was not trained on?). A model that cannot pass those tests should not move a budget.

Published by Fabulous.Media, a network of specialist marketing agencies. Updated 9 October 2026. Platform features change often; check current official documentation before acting on platform-specific detail.