For over a decade, measuring digital marketing was relatively simple: every click left a trail, every conversion had an owner, and looking at the last click was enough to decide who got the credit. That world is disappearing. Between the end of third-party cookies, tracking restrictions on iOS, and ever-tighter privacy rules, the tools that relied on following the individual user have lost much of their visibility.
This is the context in which Marketing Mix Modeling (MMM) has come roaring back. It is not a new technique — it comes from econometrics applied to retail and consumer goods — but it suddenly became relevant for companies that were once comfortable with click-based attribution. The promise is simple: understand how much each channel contributes to sales using only aggregated data, without needing to know anything about any specific person.
This article explains what MMM is, why it makes sense again, how it works under the hood, and what you need to get started — without promising magic and without hiding the limitations.
What Marketing Mix Modeling is
MMM is a statistical technique that relates what you spend on marketing to what happens to the business. In practice, you build a model — usually a regression — that tries to explain an outcome variable (sales, revenue, number of orders) from several input variables: spend on each media channel, price, promotions, seasonality, and external factors such as weather or competitor activity.
Instead of following a user's journey, MMM looks at the whole. It asks: "in the weeks when we invested more in television, did sales rise more than we would expect?" It repeats that question for each channel, across months or years of history, and separates out the effect of each one. The result is an estimate of the contribution of every euro invested — and, with it, a basis for deciding where to invest next.
Why MMM is back in fashion
The most obvious reason is privacy. MMM works with aggregated data — weekly totals of spend and sales — and never needs to identify anyone. It does not depend on cookies, on device IDs, or on individual consent, so it is immune to the changes that are dismantling user-based attribution.
There is a second, less-discussed reason: MMM sees what digital attribution cannot. It can measure the effect of offline channels (television, radio, billboards, sponsorships), of brand awareness, and even of factors you do not control, such as the weather or a public holiday. Where click attribution is short-sighted — it only counts what is clickable and trackable — MMM gives a top-down view of the whole business.
MMM and attribution are not the same thing
It is worth not confusing the two. Multi-touch attribution works bottom-up: it follows individual events and distributes conversion credit across each person's various touchpoints. It is granular, almost real-time, but blind to everything that is not trackable and increasingly constrained by privacy.
MMM does the opposite: it works top-down, with aggregated data and a long time window. It is slower (it needs history) and less granular (it will not tell you what to do tomorrow morning), but it is robust to privacy and covers everything, online and offline. Many mature teams do not choose between the two — they use MMM for the big allocation decisions and attribution for day-to-day optimization, and cross-check both with incrementality tests to validate.
How it works under the hood: adstock and saturation
Two concepts separate a serious MMM from a naive regression. The first is adstock, or the carry-over effect: the impact of an ad does not vanish on the day it airs. Someone who sees a campaign today may buy two weeks from now. The model has to spread the effect over time, with a tail that decays gradually.
The second is saturation, or diminishing returns. The first thousand euros in a channel usually deliver a lot; the next thousand deliver less; beyond a certain point, spending more barely moves the needle. A good MMM models this curve for each channel — and that is precisely where the value lies, because it reveals which channels still have room to grow and which are already saturated.
Putting the pieces together, the model splits sales into two parts: the base — what you would sell even without any marketing spend, driven by brand, distribution, and habit — and the incremental, the part marketing actually generated. It is the incremental that is worth optimizing.
What data you need to gather
MMM is demanding on history but modest on granularity. You do not need personal data; you need consistent time series, ideally two to three years, at weekly frequency. The minimum usually includes:
- Spend per channel per week — paid media, but also the cost of owned channels wherever possible.
- Outcome variables — sales, revenue, or orders, at the same time granularity.
- Price and promotions — discounts and campaigns distort everything if left out.
- Seasonality and calendar — holidays, peak seasons, Black Friday.
- External factors — weather, economic indicators, relevant competitor actions, when available.
The quality of these series matters more than the sophistication of the model. A history with gaps, mislabelled channels, or unrecorded promotions produces a confident, wrong model — the worst of both worlds.
From model to decision: optimizing the budget
An MMM is only worth the effort if it changes decisions. Once each channel's response curves are estimated, the next step is optimization: given a total budget, how do you allocate it to maximize return? The maths is about equalizing the marginal return across channels — taking from what is already saturated and putting it where there is still return to grow.
The useful output is not a magic ROI number, but scenarios. "If we keep the budget but reallocate it like this, we expect +8% incremental revenue." "If we cut 15%, this is where it hurts least to take from." These are conversations a marketing lead can have with a finance lead — and that is the real deliverable of MMM.
Common mistakes when building an MMM
The most frequent mistake is confusing correlation with causation. If you always spend more on paid search at the sales peak, the model may credit that channel with merit that actually belongs to seasonality. Hence the importance of including context variables well and, whenever possible, validating with experiments.
Other classic mistakes: history too short for the model to learn; channels aggregated at a level that hides differences (lumping all "digital" into a single number); ignoring adstock and saturation, treating everything as linear; and, perhaps the most dangerous, presenting the estimates as certainties. An MMM returns ranges, not absolute truths — communicating the uncertainty is part of using it well.
Mini-case: reallocating 20% of the budget at a retail brand
Picture a retailer with both an online and an in-store presence, investing around 1.2 million euros a year in marketing, split across paid search, social, television, and paper catalogues. Digital attribution gave almost all the credit to paid search, and the temptation was to cut television, which "did not show up" in the reports.
Building an MMM with three years of weekly history changed the picture. Paid search was saturated: the last euros delivered little. Television, invisible to attribution, showed a long adstock and a real contribution to the base of online sales — people saw the ad and then searched for the brand directly. On this basis, the team reallocated around 20% of the budget from saturated paid search to television and to an underinvested social channel, without increasing the total spend. Over the following two quarters, estimated incremental revenue rose by close to 7%, with the same investment. It was not magic: it was stopping over-investing where there was no longer any return.
Tools to get started
You do not need to buy an expensive platform to experiment. There are mature open-source libraries: Robyn, from Meta, and Meridian, from Google (which succeeded LightweightMMM), let you build an MMM with a semi-automated approach and good documentation. They require statistics and programming skills — typically in R or Python — but they remove the licence-cost barrier.
For a first iteration, the sensible thing is to start small: one market, one product line, clean two-year data. A simple model the team understands and trusts is worth more than a sophisticated model no one can explain.
In practice
Marketing Mix Modeling is not the silver bullet that replaces everything else, but it is the most solid answer to the question privacy made hard: "is my marketing actually working, and where?" It works with aggregated data, respects privacy by design, and sees the whole business — online, offline, and brand.
If today you rely almost solely on click attribution, the next step is not to throw it away, but to complement it. Gather two to three years of weekly history, start with a simple model, validate it with a real test whenever you can, and use it for the decisions that weigh most: where to allocate the budget. That is where MMM earns its keep.
