If you need an introduction to Marketing Mix Model please read here. A marketing mix model is econometrics at its core: regression that reads your past data and estimates how much each marketing channel contributed to sales. Two marketing-specific adjustments make it fit the job: adstock (advertising keeps working after the ad runs) and saturation (returns bend as spending grows). And in modern practice the estimation is Bayesian, producing probability ranges instead of single numbers.
The econometric core: learning from variation
The heart of the model is multiple regression: finding relationships between things that moved together. Think of modeling how weather affects crop yields: temperature, rainfall and sunshine each pull on the harvest, and the model untangles their separate effects. In our case the weather is spend per channel, and the crop is revenue.
The model studies weeks where spend moved, because of campaigns, seasonality, or budget changes, and checks whether sales moved with it. Because all channels enter the analysis simultaneously, each channel’s effect is estimated while holding the others constant. It might find that a 10,000 euro increase in paid search reliably preceded a measurable sales lift, while a similar increase in display produced no detectable effect.
Control variables do the housekeeping: promotions, price changes, holidays and seasonality soak up the sales swings advertising didn’t cause, so channels don’t steal credit for the Christmas rush. We’ve written about them separately: control variables in MMM.
One requirement matters above all: variation. A channel that spends the same amount every week teaches the model nothing: no movement, no signal. That’s why measurement-minded teams vary spend deliberately, and why a perfectly flat always-on budget is the hardest thing in marketing to measure.
A good model also tells you how sure it is. It might say: we’re 80 percent confident this channel’s true return sits between 1.8 and 2.6 euro per euro. That honesty about uncertainty isn’t a weakness: it’s the part budget decisions actually need.
What is Adstock in MMM?
Adstock is a simple and fundamental idea in MMM: an ad’s effect doesn’t stop the moment it airs. It decays over time, and the model applies a decay curve to capture that.
Channels decay differently. Search clicks convert fast: most of the effect lands within days. Brand TV lingers: someone who saw a campaign two weeks ago is still buying this week. The model estimates each channel’s decay speed from your data rather than assuming it.
A concrete, illustrative read: suppose the model credits TV with 10,000 euro of one month’s sales. Perhaps 4,000 of that landed while the campaign aired, and 6,000 arrived over the following weeks as the effect decayed: the adstock tail. Miss the tail and you under-credit TV by more than half, which is exactly what naive week-by-week comparisons do.
What is Saturation in MMM?
Saturation is the other marketing reality baked into the model: response to spend is a curve, not a line. The audience you can reach cheaply gets reached first; each extra euro buys a little less.
The model fits a saturation curve per channel from your history. The output that matters is the marginal return at your current spend: what the next euro buys, as opposed to the average of every euro so far. In practice the model says something like: at 100,000 euro of monthly spend in this channel, the marginal return is about 1.2 euro per euro, down from an average of 2.0. Averages flatter saturated channels; marginals price the next decision.
Each channel’s curve has its own shape, and the model learns it rather than assuming it. That’s the difference between “spend more on the best channel” and knowing when the best channel stops being best.
Why modern MMM went Bayesian
Bayesian estimation, in plain words: instead of one best-fit number, the model maintains a whole distribution of plausible values. Every output arrives as a range with stated confidence: “TV most likely lifted sales 8 to 12 percent” rather than “TV lifted sales 10 percent.”
Two practical wins follow. Priors: the model can start from what’s already known, a previous model’s results, an experiment’s finding, rather than from a blank slate. Knowledge accumulates across refreshes instead of being re-derived. Uncertainty as feedback: when the model is unsure, you can see it, and plan conservatively: budget against the lower end of a range, or test before betting.
This is the route Google’s open-source Meridian takes; Meta’s Robyn reaches similar ends via ridge regression with an evolutionary search over model configurations. Bayesian versus frequentist is a genuine statistical debate; for marketing decisions, the practical verdict is easier: ranges you can plan around beat points you have to trust.
The shift to Bayesian methods is, at bottom, a shift toward honesty: presenting uncertainty as part of the answer rather than hiding it behind a confident-looking number.
From methods to trust: validating an MMM
Building a sophisticated model is one thing. Proving it works is another, and the machinery only earns budget authority through validation.
The internal checks: holdout tests, where the model is trained on history up to a point and asked to predict weeks it never saw; and stability across refreshes, where a trustworthy model’s estimates move gradually as new data arrives rather than lurching.
The external check outranks both: the incrementality test. Take the model’s biggest claim, change spend in one place while everything else holds still (matched regions are the classic design), and watch what actually happens. A model that predicted the experiment’s result has earned real trust; one that didn’t just taught you something valuable about its blind spots.
Questions to ask about any MMM’s method
- Bayesian or frequentist, and why? Bayesian gives probability ranges and accumulates knowledge through priors; frequentist gives best-fit estimates with confidence intervals. Either can be done well: the “why” reveals whether the vendor thought about it.
- How are adstock and saturation estimated? From your data, or assumed from defaults? Estimated beats assumed. In modern MMM methods, users can define the adstock and saturation of each channel or can use the defaults.
- Which control variables, and who chose them? Seasonality, promotions, pricing, competitors. Although control variables have the reputation that are difficult to get access to, we do recommend that a typical MMM model should have the right control variables.
- How does new data enter? Full re-run from scratch, or an informed refresh that carries forward what the model learned? The second preserves accumulated knowledge.
- What validation has this model survived? Holdouts, stability, and above all experiments. A model that’s never been tested against reality is a hypothesis with a dashboard.
See MarSci’s marketing mix modeling solutions or talk to MarSci.
Frequently Asked Questions
Regression at the core: estimating each channel’s effect on sales from historical variation, with adstock and saturation transformations layered on, and in modern practice Bayesian estimation producing probability ranges rather than single point estimates. Control variables like seasonality and promotions keep those estimates honest.
The carryover effect of advertising: an ad keeps influencing sales after it runs, decaying over time. The model estimates each channel’s decay speed from data, so sales arriving weeks after a campaign still get credited to the spend that caused them.
Because budget decisions need honest uncertainty: Bayesian models output probability ranges (“most likely 8 to 12 percent”) instead of lone numbers, and priors let each refresh build on what previous models and experiments already established rather than starting blank. Ranges you can plan around beat points you must trust.
ML techniques appear inside modern MMMs, but pure prediction isn’t the goal. Budget decisions need causal, explainable channel effects with stated uncertainty, which is what the econometric structure provides: adstock, saturation and controls aren’t decoration, they’re the causal reasoning. Structure is what makes the estimates decision-grade.
No. You need the concepts: variation, adstock, saturation, ranges, and the confidence to ask vendors the method questions in this guide. Plain-language answers exist for every one of them; a vendor who can’t give them is the red flag, not your math.