Picture the version of marketing mix modeling that lived in corporate memory until recently. A consultancy collected two years of spend data, disappeared for a quarter, and came back with a bound report and a chart that said television returned 1.8 times its cost. By the time the chart landed, the media plan it described was already history. The model was treated like an annual physical: useful, expensive, slow, and not something you consulted before making a decision.
That technique is now one of the fastest growing tools in marketing measurement. In a July 2025 survey by eMarketer and TransUnion, 46.9 percent of US marketers said they would invest more in MMM over the next year, and when the same marketers were asked to name their most reliable measurement methodology, MMM was the top answer at 27.6 percent. A method with econometric roots in the 1960s is beating newer, flashier approaches on the question of which one practitioners believe. That is worth understanding, because the reason is not nostalgia.
What MMM actually is
Marketing mix modeling is a top-down statistical model. It takes an outcome you care about, usually sales or revenue, and relates it to the things that might explain the movement: spend on each marketing channel, plus the factors outside marketing that move sales anyway. Price changes. Promotions. Seasonality. A competitor's launch. Weather, for some categories. The model is usually regression-based, and it works on aggregate data, weekly spend and weekly sales across a market or a region.
The detail that matters most is what MMM never touches. It does not know who you are. It does not need a cookie, a device ID, a login, or a consent signal. It looks at the total picture, "the brand spent this much on each channel each week, and sales did this," and works backward to an estimate of how much each channel contributed. There is no individual journey in the math. That single design choice, invisible for decades, turned out to be the thing that saved it.
The model also has to handle two awkward facts about advertising. Spend does not convert to sales instantly, so MMM uses an adstock or carryover term to capture how an ad keeps working for weeks after it runs. And channels do not return the same rate forever, so the model uses a saturation curve to capture diminishing returns, the point where the next dollar into a channel buys less than the last one. Those two ideas, carryover and saturation, are what let MMM answer the question a budget owner actually has: not "did this work" but "what happens to sales if I move a million dollars from one channel to another."
Where it came from
The econometric machinery underneath MMM predates marketing. Ragnar Frisch and Jan Tinbergen built the foundations of econometrics in the 1930s, applying statistical methods to economic models, work that won the first Nobel Prize in economics in 1969. Those methods migrated into marketing research through the 1950s and 1960s, when companies started using regression to connect advertising spend to sales. Kraft was an early adopter, and consumer packaged goods led the way because CPG brands already had the raw material: Nielsen tracked their sales, prices, and advertising.
The discipline became rigorous in the 1970s, when MIT's John Little built what is often described as the first true marketing mix model, using regression to quantify how advertising and price promotions moved sales. Through the 1980s and 1990s, MMM was simply how large CPG companies, the Procter and Gambles and Coca-Colas, planned media. It was mainstream, respected, and entirely normal.
Then digital arrived and bolted a second measurement tradition onto the first. The web could record clicks. Clicks could be tied to individuals through cookies. By the late 2000s, multi-touch attribution promised something MMM could never offer: a person-level, touch-by-touch account of every conversion, refreshed daily, apparently precise to the dollar. Against that, MMM looked like a relic. It was aggregate where digital was granular, quarterly where digital was real-time, a range where digital gave you a number. For roughly a decade, the deterministic, cookie-based approach felt like the winner, and MMM was quietly filed under "what CPG does."
Why it is back
The comeback is not a marketing fashion. It is a direct consequence of the cookie-based model breaking.
Apple's App Tracking Transparency, introduced with iOS 14.5 in 2021, let users opt out of cross-app tracking, and most did. Safari and Firefox block third-party cookies by default. Google spent years promising to deprecate them in Chrome and then reversed course, but the damage to industry confidence was already done. The signal that multi-touch attribution depends on, a continuous record of one identifiable person moving across sites and apps, has been shredded. Industry estimates put MTA coverage at 30 to 60 percent of its 2020 levels. A model that needs to see the whole journey now sees a fraction of it and models the rest.
MMM walked through that wall untouched. It never needed identity, so the signal collapse that broke MTA did not degrade it at all. The privacy regulation that is a structural threat to attribution is, for MMM, simply not a relevant problem. That is the core of the renaissance: of all the measurement methods marketers had, MMM is the one that was privacy-safe by design before privacy was the issue. It survived the storm by accident of its 1960s architecture.
The adoption numbers reflect this. eMarketer and TransUnion data shows nearly half of US brand and agency marketers already investing in MMM. The technique that looked dated is the one practitioners now name as their most reliable, precisely because it works under any regulatory regime. For the contested obituary of the method it is replacing, see our piece on multi-touch attribution dying.
Modern MMM is a different product
If MMM were back in its old form, this would be a smaller story. A slow quarterly consulting deliverable, even a privacy-safe one, is still slow. What actually changed is that the product modernized while it was out of fashion.
Two shifts did most of the work. The first is open-source frameworks. Google announced the wide launch of Meridian in January 2025, an open-source, geo-level, Bayesian MMM library that is actively developed and free to use. PyMC-Marketing, from PyMC Labs, is the other heavily maintained framework, fully Bayesian with support for custom priors. Meta's Robyn, released in 2020, did much to popularize open-source MMM in the first place; it still ships updates but Meta labels it experimental, and momentum has shifted toward Meridian and PyMC-Marketing. A capability that used to mean a six-figure engagement is now a library you can install, which collapses the barrier to entry. Our comparison of open-source MMM frameworks goes deeper on the trade-offs between them.
The second shift is Bayesian methods, and this one is about honesty as much as speed. A classic regression hands you a point estimate: television returned 1.8 times its cost, full stop. That number carries false precision, because the model is never that sure. A Bayesian model expresses its answer as a distribution, a credible interval rather than a single figure. The result reads as "the return is most likely between 1.4 and 2.3 times, centered near 1.8." That is less satisfying and far more useful. It tells a budget owner where the model is confident enough to act and where it is guessing, which channels to scale and which to test first.
The cadence changed too. Cloud computing lets modern MMM process larger datasets and refresh far more often than an annual review allowed. Commercial platforms now describe weekly or near-continuous refreshes and scenario planning you can run before committing a budget, rather than a verdict delivered after the fact. The honest framing is that data preparation, not modeling, is still the slow part. As one measurement lead put it, the model is roughly 20 percent of the work and consolidating the data is the other 80 percent. MMM has gone from annual to almost-continuous, but it has not become instant, and anyone promising real-time MMM is overselling.
The honest limits
MMM is not a universal answer, and the renaissance comes with conditions that matter.
It needs scale and it needs variation. The model learns by watching spend change and seeing what sales do in response. If a brand's budget is flat and its channel mix never moves, there is not enough variation for the model to separate signal from noise. Two years of weekly data is only about 104 observations, which is a small sample for a model with many channels to untangle. MMM also struggles with very few channels, because it cannot isolate an effect it has nothing to contrast against. One industry guideline, from measurement firm Circana, is that a brand's marketing budget should be at least fifty times the cost of the modeling project before MMM is worth the spend. A small brand running two channels on a flat budget should not buy it yet.
It is correlational, not causal. MMM observes that spend and sales moved together; it does not run an experiment. When marketers raise spend across several channels at once, or push hard exactly when demand is already rising, the model can credit a channel for sales it did not cause. This is why the serious guidance, including from Google's own research, is to pair MMM with experiments. Incrementality and geo tests produce a real causal measurement that can be fed back into the model as a calibration, anchoring the correlational estimate to experimental ground truth.
And it cannot do tactical work. MMM operates at the channel level on a weekly grain. It will not tell you which keyword, which creative, or which audience to adjust this afternoon. That is not a flaw to fix; it is the wrong tool for that job.
Where it fits
The useful way to think about modern measurement is not picking a winner. It is a stack, and the consensus framing gives each method the question it is good at. MMM sets strategic budget allocation across channels and captures brand and upper-funnel effects that touch-based tracking never sees. Incrementality and geo experiments provide causal proof and calibrate the model. Platform and multi-touch signals, where consent still allows them, handle tactical optimization inside a channel. When the three roughly agree, a budget decision is defensible. When they disagree, that disagreement is information.
The technique that looked obsolete turned out to be the load-bearing piece. It is older than the cookie, and it outlived it.
Council summary
This post argues that marketing mix modeling did not so much return as get reclassified: the same aggregate, identity-free method that once felt like a slow relic is now the most reliable measurement tool marketers have, because privacy rules broke the cookie-based attribution that briefly outshone it. The depth is in the contrast. The piece traces the lineage from 1930s econometrics through John Little's 1970s models, explains adstock and saturation in plain terms, and then shows precisely how open-source frameworks and Bayesian credible intervals turned a six-figure annual deliverable into something a team can run and re-run. It is honest where it counts: MMM is correlational, needs spend variation and scale, cannot do tactical work, and is not real-time. The reader's takeaway is to stop hunting for one winning measurement method and build a stack, with MMM setting strategic budget allocation, incrementality tests calibrating it, and platform signals handling tactics. Verified against eMarketer, Forrester, Circana, and Google research; publishable.
Comments