In short: Marketing mix modeling estimates each channel's total contribution to an outcome from aggregate weekly data, without tracking individuals. It needs history and variation and produces rough answers, but combined with attribution and incrementality tests it gives the most honest picture a company can assemble.
Key points
- MMM needs at least a year or two of weekly data with real variation in spend across channels.
- Unlike attribution, MMM measures a channel's total effect; unlike incrementality tests, it needs no holdout.
- It goes wrong with too little variation, channels switched on together, long lags, or over-trusted point estimates.
- Small companies can run a simple model with open-source tools published by large technology companies.
- Because it uses aggregate data, MMM survives ad account suspensions that wipe pixel history.
Marketing mix modeling, MMM, is a statistical method that estimates how much each marketing channel, and other factors like price, seasonality, and competitor activity, contributed to an outcome such as sales or sign-ups. It works from aggregate data over time, weekly spend by channel against weekly outcomes, and does not track any individual, which is why it has returned to fashion as individual tracking has declined.
It was built for large consumer companies with years of data and television budgets. A version of it now works for companies much smaller than that.
What it does
Given a history of spend by channel and outcomes over the same periods, the model estimates the relationship between each channel's spend and the outcome, allowing for diminishing returns, delayed effects, and the influence of factors the company does not control. The output is an estimate of each channel's contribution, its return at the current spend, and, with care, what would happen if spend were shifted between channels.
Unlike attribution, which assigns credit to touches on converting customers, MMM attempts to measure the total effect of a channel, including the customers it influenced without a trackable touch. Unlike incrementality tests, it does not require holding anything out; it infers from natural variation in spend and outcomes.
What it needs
Enough history: at least a year or two of weekly data, with real variation in spend across channels. If a channel was always on at the same level, the model cannot tell what it did. Clean, consistent data: spend, impressions where available, and outcomes, all by week, all defined the same way throughout. Knowledge of the external factors: pricing changes, launches, seasonality, competitor events, platform policy changes, so they can be included rather than mistaken for channel effects. And someone who understands both the statistics and the business, because a model will produce a confident answer to a badly posed question.
Where it goes wrong
Too little variation, so the model attributes outcomes to whatever happened to move. Channels that were turned on and off together, so their effects cannot be separated. Outcomes with long lags, where this quarter's result reflects last quarter's spend. Over-trust in point estimates that have wide uncertainty. And the temptation to use the model to justify a plan already decided, by adjusting assumptions until it agrees.
A lightweight version for a small company
A company with a year of weekly data across four to eight channels can run a simple model with open-source tools that several large technology companies have published for exactly this purpose. The results will be rough. They will also be more informative than a platform's attribution report, because they attempt to answer the right question.
The practical approach is to combine three things: attribution for running the week, one or two incrementality tests per quarter on the biggest channels, and a simple mix model refreshed quarterly to check that the whole picture is consistent. Where the three disagree, the disagreement is the finding.
For companies whose tracking is fragile
Companies in restricted categories lose pixel history every time an account is suspended and cannot rely on platform reporting. Aggregate methods do not care. A model built on the company's own weekly spend and outcome data survives every account reset, which makes it unusually valuable for exactly the companies least likely to have heard of it.
lowob takeaway: MMM estimates each channel's total contribution from aggregate data, without tracking individuals. It needs history and variation, produces rough answers, and, combined with attribution and incrementality tests, is the most honest picture a company can assemble.
Common questions
What is marketing mix modeling?
A statistical method that estimates how much each marketing channel, and factors like price and seasonality, contributed to an outcome such as sales. It works from aggregate data over time and does not track individuals.
Can a small company use marketing mix modeling?
Yes. A company with a year of weekly data across four to eight channels can run a simple model with open-source tools. The results will be rough but more informative than a platform's attribution report.
How is MMM different from attribution?
Attribution assigns credit to touches on converting customers. MMM attempts to measure a channel's total effect, including customers it influenced without a trackable touch.
Should I use MMM, attribution or incrementality testing?
Combine all three: attribution for running the week, one or two incrementality tests per quarter on the biggest channels, and a simple mix model refreshed quarterly. Where they disagree, the disagreement is the finding.