Marketing Mix Modeling
Marketing mix modeling, or MMM, estimates what each channel contributed to revenue by analyzing the relationship between spend and outcomes over time, at the portfolio level. It predates the internet: consumer goods companies used it to read TV and print decades before online tracking existed. Demand for it has returned because tracking individual users, never possible offline, is now unreliable online as well. MMM does not depend on cookies, so cookie loss does not affect it.
How it actually works
The model is built from history: weekly or monthly spend by channel, revenue, and the outside factors that also move sales, such as seasonality, pricing, promotions, and the economy. Statistical methods estimate how much each input contributed, including saturation, the point where another dollar in a channel buys less than the last one, and lag, the delay between spend and its effect.
Its output is a different kind of number than attribution produces. Attribution assigns tracked touchpoints to individual sales and inherits every blind spot of the tracking. MMM estimates incremental contribution directly, which is the question that matters: what did the spend cause that would not have happened without it. The trade is precision for reliability. MMM comes with error bars, and it works at the channel level rather than the campaign level. Both limits deserve honest statement. The intervals can be wide: an estimate spanning from break-even to strongly positive supports a direction rather than a precise reallocation, and a model reported without its intervals overstates what it knows. The harder limit is identification: channel spend series tend to move together, and they are set in response to performance, so the model can struggle to separate one channel's effect from another's, or from the demand that drove the spending in the first place. The method also has data requirements, typically two to three years of weekly spend and outcome history with genuine variation in it. A business without that history, or one that never varied its spend, does not yet have the inputs a trustworthy model needs, and should run experiments to create the variation before commissioning the model.
The estimates sharpen when calibrated against experiments. A holdout test in one channel establishes a ground-truth point of incrementality, and the model is tuned to agree with it. Experiments answer one question with high confidence; the model extends those answers across channels and time. Run together, they replace the platforms' self-graded reports with measurement the platforms do not control.
In practice
At USCCA, measurement built outside the platforms identified over $12 million in spend that was not producing incremental revenue. Platform reporting had graded that spend as working, because platforms count revenue from buyers who were already going to purchase. Portfolio-level measurement asked the counterfactual question instead, and the identified waste changed how the budget was allocated.
Where we come in
We build mix models alongside controlled experiments and holdouts, outside platforms that are incentivized to overreport performance, so budget follows what moves revenue and profitability. We do not stop at the channel level. The same method extends to the creative, and the common features across creative at each step of the customer journey, to see what helps and what hurts instead of what platform attribution credits.
Start a Revenue Health Pre-Assessment →See it in action
Related terms
- Incrementality
- The revenue that would not have happened without the marketing. The counterfactual question platforms cannot answer about their own ads, and the quantity MMM is built to estimate.
- Holdout test
- Turning a campaign off for a matched region or audience and comparing outcomes against where it stayed on. The most direct read of what spend causes.
- Saturation curve
- The declining return on each additional dollar in a channel. Knowing where each channel sits on its curve is what turns a model into a budget allocation.
- Adstock
- The lagged effect of advertising, where spend this week keeps producing in later weeks. Models that ignore it under-credit demand creation and over-credit capture.
- Multi-touch attribution
- The bottom-up alternative: splitting credit for individual sales across tracked touchpoints. Granular where MMM is broad, and limited to what its tags can see.
- Model calibration
- Adjusting a model until it agrees with experimental ground truth. Calibration is what turns the model's estimates into numbers a budget can be allocated on.