Marketing mix modeling is the measurement method that survived the privacy era, and the industry has rediscovered it with the enthusiasm of a team that has run out of alternatives. It needs no cookies, no device IDs, no user-level data at all, and it covers every channel you spend on — including the ones no pixel can see. It also cannot tell you a single thing about which of your ads worked, and that gap is where most MMM programs quietly disappoint the people who paid for them.
What MMM actually does
An MMM is a regression built on aggregate history. You feed it two or three years of weekly data — spend by channel, plus price, promotions, seasonality and whatever else moves your category — and it fits a model that explains the sales line. The output splits revenue into a base (what you would have sold with no advertising at all) and incremental contributions per channel.
Two mechanics do the real work. Adstock models carryover: a TV flight keeps paying out for weeks after the money stops, so the model spreads its effect forward with a decay rate. Saturation models diminishing returns: the tenth million in a channel does not buy what the first did, so each channel gets a response curve that flattens. Those two curves are what let an MMM answer the strategic question — where does the next dollar go?
Why it came back
- Signal loss made it necessary — Apple’s App Tracking Transparency and the long decline of third-party cookies broke the user-level pipes multi-touch attribution ran on. MMM never used those pipes, so it never broke.
- Open source made it affordable — Meta’s Robyn and Google’s Meridian both ship free, along with PyMC-Marketing on the Bayesian side. What used to be a six-figure consulting engagement is now a repository and an analyst.
- It covers the uncoverable — retail media, out-of-home, podcast reads, linear TV, sponsorships. If it has a spend line, it goes in the model.
What it costs and how long it takes
The licence may be free; the programme is not. A credible model wants roughly 104 to 156 weeks of clean weekly history, which disqualifies brands younger than two years and any account that changed its data structure last spring. Expect eight to twelve weeks to a first readout, most of it data engineering rather than modelling.
Then there is the variance problem: if your channel budgets moved in lockstep for two years, the model cannot separate them, because nothing in the history distinguishes their effects. MMM rewards advertisers who already varied their spend and punishes the disciplined ones who never did.
A worked example
A brand spends $4M a year across five channels and the model reports a base of 62% of revenue. Of the incremental 38%, search takes the largest contribution but sits far up the flat end of its saturation curve, while a $300K connected-TV line is still on the steep part of its own. The recommendation: move roughly $400K from search to CTV, for a projected gain of about 6% on total contribution.
That is a valuable answer, and nothing else produces it. Note what it is not. It is a statement about channels: CTV deserves more money. It does not say what to put in the CTV slot.
The creative blind spot
Inside an MMM, creative is not a variable. It is noise absorbed into the channel coefficient. A quarter in which three strong ads and six weak ones ran in the same channel produces one blended number, and that number is then used to set next quarter’s budget for the channel — as if the assets were interchangeable.
This matters more than the modelling literature admits. Creative is the largest controllable driver of ad effectiveness in most published analyses, and the spread between the best and worst asset in a flight is usually wider than the spread between the channels being modelled. An MMM that shifts 10% of budget between channels is optimising the smaller lever while the larger one goes unmeasured.
It is also too slow to help. MMM reads history; by the time a quarter is in the model, the media is spent. Nothing in the process tells you, before production, which of six concepts deserves the budget. That question belongs to concept testing and, once the assets exist, to creative testing.
How to use it well
- Calibrate it with experiments — an MMM fits correlations in observational data, so its coefficients drift without a causal anchor. Both Meta and Google recommend feeding incrementality test results back in as priors. An uncalibrated model is a confident opinion, not a measurement.
- Refresh it, don’t enshrine it — quarterly refreshes keep the curves current. An annual model is a historical document by the time anyone acts on it.
- Keep the layers separate — MMM allocates budget across channels, incrementality validates that the channel pays, and pre-testing decides what runs inside it. Substituting any one for another is how testing budgets get wasted.
Used properly, marketing mix modeling answers the budget question better than anything else available in 2026. Just don’t ask it the creative question. It was never built to hear it.
Frequently asked questions
Is marketing mix modeling the same as media mix modeling?
In practice the terms are used interchangeably and both are abbreviated MMM. Purists reserve “marketing mix” for models that include price, promotion and distribution alongside media, and “media mix” for models limited to advertising channels. Most commercial models include the non-media variables either way.
How much data do I need to run an MMM?
Roughly two to three years of weekly observations — 104 to 156 data points — with consistent definitions throughout. You also need genuine variation in spend across channels; if budgets moved together for the whole period, the model cannot attribute their effects separately.
Does MMM replace attribution?
No. They run on different clocks. Attribution is daily and useful for tactical pacing; MMM is quarterly and answers strategic allocation. Incrementality experiments are the causal calibration layer that tells you how much to trust either one.
Can MMM tell me which ad performed best?
No. Creative variation is absorbed into the channel coefficient, so an MMM reports a single blended effect for everything that ran in a channel during the period. Isolating individual assets requires creative-level testing, ideally before the media is committed.
Is open-source MMM good enough, or do I need a vendor?
Robyn and Meridian are production-grade and used by large advertisers, so the modelling is rarely the constraint. Analyst time is: data preparation, validation, and the judgement to spot an implausible result. Teams without that in-house are usually better served by a vendor.
