What is Marketing Mix Modeling?
Marketing Mix Modeling (MMM) is a statistical technique for measuring the effect of marketing spend on revenue. It uses historical data on channel spend, revenue, and external factors to estimate how much each channel contributed to business outcomes and how those contributions change as spend is reallocated.
Unlike multi-touch attribution, which works at the user level, MMM operates on aggregate data: total weekly or daily spend per channel and total weekly or daily revenue. It does not need to track individual users, which makes it resilient to cookie loss, privacy restrictions, and cross-device complexity.
Why MMM Has Returned to Prominence
MMM was the dominant marketing measurement method before digital advertising made user-level tracking possible. It fell out of fashion in the 2000s and 2010s as click-based attribution became ubiquitous.
It has returned to prominence because user-level tracking is increasingly unreliable:
- Apple’s iOS 14.5 App Tracking Transparency framework restricts mobile attribution
- Third-party cookie deprecation in Chrome limits cross-site tracking
- Intelligent Tracking Prevention in Safari shortens cookie lifespans
- Regulatory regimes such as GDPR and state-level privacy laws constrain user identification
MMM sidesteps all of these constraints because it does not require user-level identification.
How MMM Works
At a simplified level, MMM fits a regression model that explains revenue as a function of:
- Spend per channel over time, typically with a transformation to capture diminishing returns
- Adstock, which represents the decaying effect of past spend on current sales
- External factors such as seasonality, price changes, promotions, and macroeconomic conditions
- A baseline that captures revenue that would occur regardless of marketing
The output is an estimated contribution per channel, a response curve showing how revenue changes with spend, and predictions of what would happen if the budget were reallocated.
Three Common Approaches
MMM implementations vary, but three approaches dominate current practice:
Bayesian MMM uses Hamiltonian Monte Carlo sampling to fit the model. It produces uncertainty intervals around every estimate, which is useful for risk-aware budget decisions. It requires more compute than frequentist methods but is increasingly practical with modern tooling.
Ridge Regression MMM uses regularised linear regression to fit the model efficiently. It is faster and often paired with evolutionary algorithms (such as NSGA-II) to search for optimal budget allocations.
Ensemble MMM combines multiple approaches and averages their outputs. Ensembles tend to produce more stable estimates than any single method because they hedge against the assumptions of individual models.
MMM vs Multi-Touch Attribution
MMM and MTA answer different questions and have different strengths:
MMM answers “what is the causal contribution of each channel to revenue, and how should I reallocate the budget?” It is resilient to tracking loss and captures offline influences, but it operates at a channel level rather than a campaign or keyword level.
MTA answers “which touchpoints were present in the journeys of customers who converted?” It is granular and actionable at the tactical level but depends on tracking and is vulnerable to privacy changes.
Most mature measurement stacks use both. MMM guides the overall channel mix. MTA guides within-channel optimisation.
Budget Optimisation with MMM
A key output of MMM is the channel response curve, which shows the expected revenue at each level of spend per channel. Response curves are typically concave, reflecting diminishing returns: the first pounds spent on a channel yield a high marginal return, but each additional pound yields less.
The optimal budget allocation is the point at which the marginal return per pound of spend is equal across every channel. Moving a pound from a channel with low marginal return to one with high marginal return increases total expected revenue.
MMM budget optimisers use the response curves to search for the allocation that maximises forecasted revenue subject to a total budget constraint. The result is a specific, quantified reallocation recommendation, not a qualitative hunch.
Limitations
MMM has real limitations:
- It requires enough historical data to fit the model, typically 1 to 3 years of weekly data with meaningful spend variation.
- It is sensitive to the assumptions built into the model specification, particularly the choice of adstock and saturation functions.
- It can confuse correlation with causation if an external factor moves together with marketing spend in ways the model does not capture.
- The tactical granularity is limited, since it operates on aggregate channel spend rather than individual campaigns.
These limitations are mitigated, not eliminated, by careful model specification and by triangulating MMM outputs with other measurement methods such as incrementality testing.