Media mix modeling estimates how much each marketing channel contributed to sales using historical aggregate data. It answers a real question, and it fails in specific, predictable ways.
The method works on aggregates, not individuals
The model takes periodic totals of spending by channel alongside sales, and estimates a statistical relationship between them while controlling for price, seasonality and other factors.
Because it never identifies a person, it is unaffected by cookie restrictions, app tracking limits and privacy regulation. That resilience is the main reason the technique returned to prominence.
It also covers channels that cannot be tracked individually. Television, radio, print, outdoor and sponsorship enter the model on the same footing as digital spending.
Variation in the data is the binding constraint
A regression can only estimate the effect of something that changed. A channel funded at a constant level throughout the period contributes almost nothing the model can measure.
Channels whose budgets move together are similarly problematic. If two are always increased at the same time, the model cannot separate their individual effects reliably.
This is why disciplined variation matters. Deliberately changing spend across regions or periods creates the contrast the estimation requires.
Correlation with demand is easily mistaken for cause
Marketing budgets often rise when demand is already rising, because plans are set from forecasts. The model then attributes to advertising what was actually seasonality or a market shift.
Careful implementations control for known drivers such as pricing, distribution changes, weather and competitor activity, but only for those included in the data.
Omitted variables do not announce themselves. A model that fits history well can still be attributing effects to the wrong cause throughout.
Adstock and saturation are assumptions, not findings
Advertising effects persist beyond the period of spend and diminish as spend rises. Models represent both with mathematical forms chosen by the analyst.
Different reasonable choices produce materially different conclusions about how long an effect lasts and where diminishing returns begin.
Because those parameters drive the recommendations, the assumptions deserve as much scrutiny as the results. A model presented without them cannot be evaluated.
Experiments are what validate the model
The strongest test is a holdout: suspend or increase a channel in randomly chosen geographies and compare outcomes against the model's prediction.
Agreement builds confidence in the whole structure. Disagreement identifies which coefficients are unreliable and where the specification needs work.
Mix modeling and controlled experiments answer different questions. The model allocates across everything at once, and experiments verify particular claims, which is why serious programs run both.