Marketing mix modelling (MMM) is a statistical technique that estimates how much each marketing activity contributed to a business outcome, usually sales or revenue, using aggregated data over time. A model takes weekly or daily totals - television spend, search spend, price, promotions, holidays, weather - and works out which movements in those inputs line up with movements in sales. It exists because most advertising cannot be traced from exposure to purchase at the level of an individual. MMM sidesteps the problem by never looking at people at all.

How a mix model works

A mix model is a regression equation. The outcome, say weekly revenue, sits on one side. On the other sit a baseline (the sales a brand would make with no marketing), seasonal terms, control variables such as price and competitor activity, and one term per media channel. Fitting the model to history yields a coefficient for each channel, which converts into a contribution: the share of sales the model credits to that channel.

Two transformations handle advertising's non-linear behaviour.

  • Adstock captures carryover. An ad seen this week can still influence purchases weeks later, so each week's spend is smeared forward with a decay rate. The British media researcher Simon Broadbent formalised adstock in 1979. A decay of 0.5 means half of a week's effect persists into the next.
  • Saturation captures diminishing returns. A Hill curve or similar function makes each additional pound of spend produce less than the one before, so the model can tell a channel near its ceiling from one with room to grow.

A hypothetical example shows the output. A retailer feeds three years of weekly data, 156 observations, into a model with five channels. It reports that television drove 14% of revenue at a return of 1.9, paid search 9% at 3.4, and that the next 100,000 pounds in search would return only 1.2 because the channel sits on the flat part of its curve. That last figure, the marginal return, is what budget optimisers try to equalise across channels.

Data is the main barrier. Most practitioners want two to three years of weekly history, and PPC Land reported that Google's Meridian needs at least 12 to 18 months of historical data. Spend must also vary, or the model has nothing to learn from. The analyst Sebastian Hewing proposed a budget-based framework in which MMM is reserved for advertisers spending more than 1 million euros or dollars a month.

Who builds the models

On the buy side, MMM has traditionally been commissioned by advertisers from specialist consultancies, analytics arms of media agency groups, and research firms such as NIQ and Circana. An IAB Australia landscape published in September 2025 profiled twelve providers, from Analytic Partners and Gain Theory to the open-source frameworks.

Three free libraries now dominate in-house work. Robyn, from Meta Marketing Science, is an R package (version 3.12.1 as of October 2026; its Python port is labelled a beta). It fits ridge regression and tunes adstock and saturation settings with Meta's Nevergrad evolutionary optimiser, scoring candidate models on three objectives: NRMSE (prediction error), DECOMP.RSSD (how far each channel's share of effect strays from its share of spend) and, when experiments are supplied, MAPE.LIFT (disagreement with those experiments).

Meridian, from Google, is a Bayesian hierarchical model that can work at regional level and accepts reach and frequency inputs as well as spend. Its central feature is the prior: an analyst states a belief about a channel's return before fitting, and the data updates it. PyMC-Marketing, maintained with support from the consultancy PyMC Labs under an Apache 2.0 licence, is a Bayesian Python library offering configurable adstock and saturation functions, time-varying media effects and budget optimisation.

On the sell side, platforms and retailers supply data feeds for advertisers' models. Amazon's MMM application and API, for example, reached general availability across 14 markets in May 2026.

Calibration against experiments

Observational data alone often cannot pin down a channel's true effect, so modern practice anchors models with incrementality tests: controlled experiments that measure a causal effect directly. Incrementality results enter a model in two main ways. In Bayesian tools they become priors; a lift test showing a channel return of 2.5 sets the model's starting belief near 2.5. In Robyn, they enter as a calibration objective that penalises models disagreeing with the test.

Meta's May 2025 white paper "Building a Suite of Truth" placed MMM in the middle of a ladder of incrementality, below randomised experiments, and described calibration either as a directional check or as direct input into coefficients. The UK's IPA took a similar position in a framework published in March 2025.

From marketing mix to mix models

The phrase "marketing mix" predates the models. Neil Borden of Harvard Business School used it in teaching from the late 1940s, building on James Culliton's 1948 description of the marketing executive as a "mixer of ingredients", and set it out in "The Concept of the Marketing Mix" in the Journal of Advertising Research in 1964. Econometric models of advertising and sales date from roughly the same period; an IAB white paper of April 2026 described MMM as in use since at least the 1960s. Broadbent's adstock work in 1979 gave models a standard treatment of carryover, and through the 1980s and 1990s MMM became a staple for consumer goods brands measuring television against retail scanner sales.

Digital advertising pushed MMM aside. Click tracking appeared to measure conversions directly, and Hewing recalled first encountering MMM in 2007 as attribution was emerging. The revival came with privacy restrictions, the loss of mobile identifiers and the prolonged uncertainty over third-party cookies.

Google researchers Yuxue Jin, Yueqing Wang, Yunting Sun, David Chan and Jim Koehler published "Bayesian Methods for Media Mix Modeling with Carryover and Shape Effects" in 2017, warning that with small samples the choice of priors can heavily influence, and bias, the results. Meta's Robyn followed; the earliest documented user, the mattress company Resident, began running it in December 2020. Google's LightweightMMM (2022) and PyMC-Marketing (2023) followed, and Google unveiled Meridian in March 2024. Meridian opened to all advertisers on January 29, 2025, after which LightweightMMM was marked as unsupported.

Why it matters to marketers

MMM answers the budget question that attribution cannot: how to split money between television, search, social, retail media and everything else. Working on aggregates, it needs no consent to track individuals. The French trade body Alliance Digitale argued in January 2026 for pairing MMM with attribution and experiments as cookie-free tracking faltered.

Demand is rising faster than capability. Research cited by PPC Land found 46.9% of marketers planned to increase MMM investment, while only 15% of teams had implemented it. Output from mix models feeds metrics such as iROAS and the long-term ROI multiplier, which differ from platform-reported ROAS because they attempt a causal estimate.

Limitations and disputes

The deepest problem is endogeneity. Budgets rise when demand is expected to rise, so spend and sales move together for reasons other than advertising. Endogeneity usually inflates the returns of demand-capturing channels. A simulation by the Zalando researcher Niklas Heusch, posted to arXiv on August 21, 2026, found a standard model reported a paid search ROAS of 10.61 against a true 4.20, and 8.41 even with the true confounders supplied. Ryan Dew, Nicolas Padilla and Anya Shchetkina argued in a 2024 paper that saturation and time-varying effects are often statistically indistinguishable.

Multicollinearity is the practical version of the same issue: when television and search rise together at every launch, the model cannot apportion credit between them. Priors and calibration help, but critics note they move assumptions rather than remove them. Robyn's DECOMP.RSSD objective, by design, favours models whose effect shares resemble spend shares.

Structural fit is disputed too. The IAB's April 7, 2026 white paper argued that standard MMM is misaligned with retail media, because always-on spend disappears into the baseline and cost-per-click buying does not map to the impression inputs models expect.

Governance is the third front. Platforms that sell media now also supply the models that allocate budgets across them. Commenters in a 2026 debate over Robyn asked whether platform-built tools favour their makers, and one practitioner, Dan Rubel, said MMM "feels more and more like religion". The Coalition for Innovative Media Measurement argued in July 2026 that default priors and embedded assumptions can reconfigure data asymmetry, and proposed a standard model factsheet.

Not the same as

  • Data-driven attribution assigns credit for individual tracked conversions across the touchpoints a user saw. It works bottom-up on user-level paths; MMM works top-down on aggregates and covers untracked channels.
  • Media mix modelling is often used interchangeably. Strictly, it covers only media inputs, while marketing mix modelling also models price, promotion, distribution and product.
  • Incrementality testing measures a causal effect through a controlled experiment for one channel at one time. MMM estimates all channels at once from history, and tests are commonly used to calibrate it.
  • Econometrics is the broader discipline. MMM is one application of its methods.

Recent developments

At Google Marketing Live on May 20, 2026, it announced Meridian inside Analytics 360. Version 2.0.0 of the Meridian library, with JAX as its default computation backend, followed on September 2, 2026, and the geo-experiment tool GeoX reached general availability on September 9, with Google claiming experiments over 31% cheaper. NIQ announced on September 30, 2026 that Meridian now runs inside its Cadence platform.

The direction is towards models calibrated by experiments and wired into planning tools. Whether that resolves the identification problem or relocates it into priors remains unsettled as of October 2026.

Timeline

  • 1948: James Culliton describes the marketing executive as a "mixer of ingredients".
  • 1964: Neil Borden publishes "The Concept of the Marketing Mix" in the Journal of Advertising Research.
  • 1979: Simon Broadbent introduces adstock.
  • 2017: Google researchers publish "Bayesian Methods for Media Mix Modeling with Carryover and Shape Effects".
  • December 2020: Resident begins running Meta's Robyn alongside its in-house model.
  • February 14, 2022: Google's LightweightMMM is first released on PyPI.
  • February 9, 2023: PyMC-Marketing is first released on PyPI.
  • March 2024: Google unveils Meridian.
  • August 14, 2024: Dew, Padilla and Shchetkina post "Your MMM is Broken".
  • January 29, 2025: Meridian becomes available to all advertisers.
  • March 2025: The IPA publishes a framework treating experiments as the calibration anchor for MMM.
  • May 28, 2025: Meta publishes "Building a Suite of Truth".
  • September 2025: IAB Australia publishes its MMM vendor landscape.
  • January 2026: Alliance Digitale publishes its measurement white paper.
  • April 7, 2026: The IAB publishes its white paper on MMM and retail media.
  • May 2026: Amazon's MMM API reaches general availability in 14 markets.
  • May 20, 2026: Google announces Meridian in Analytics 360.
  • July 2026: CIMM proposes a standard MMM factsheet.
  • August 21, 2026: Heusch's paid search simulation appears on arXiv.
  • September 2, 2026: Meridian 2.0.0 is released.
  • September 9, 2026: Meridian GeoX becomes generally available.
  • September 29, 2026: PyMC-Marketing 1.2.0 is released.
  • September 30, 2026: NIQ integrates Meridian into Cadence.

Summary

  • Who: Advertisers, agencies, consultancies and research firms build mix models; Meta (Robyn), Google (Meridian) and PyMC Labs (PyMC-Marketing) maintain the main open-source tools; platforms and retailers supply data.
  • What: A regression-based method that estimates each marketing input's contribution to sales from aggregate time-series data, using adstock and saturation transformations, increasingly calibrated with experiments.
  • When: The marketing mix concept dates from 1948 to 1964 and the models from the 1960s; the technique revived from around 2020, with Meridian opening globally on January 29, 2025.
  • Where: Upstream of media buying, in annual and quarterly budget planning across all channels, including offline media that cannot be tracked.
  • Why: Privacy rules and signal loss have weakened user-level tracking, and MMM measures channel contribution without identifiers, though endogeneity and governance disputes remain unresolved.