The history of marketing mix modeling runs from a phrase coined at Harvard in the late 1940s to Bayesian libraries released by Google and Meta in the 2020s. The “marketing mix” comes from Neil Borden, who borrowed his colleague James Culliton’s 1948 image of the marketer as a “mixer of ingredients”, while the modeling comes from econometricians who, from the 1950s onward, used distributed lag regressions to measure how advertising builds sales over time, an idea Simon Broadbent turned into “adstock” in 1979. Scanner data made the method a commercial product for consumer goods brands from the late 1970s, digital attribution pushed it aside in the 2000s and 2010s, and privacy changes plus free open-source tools brought it back after 2017.
This history is worth a few minutes of anyone’s time, because almost every design choice in a modern library makes more sense once you know which problem it was invented to solve. Adstock exists because 1970s TV advertisers needed to explain sales that kept coming after the campaign ended, the baseline exists because packaged goods brands sold plenty without any advertising, and the current enthusiasm for calibrating models with experiments comes from a decade of research showing how badly observational methods can miss. The timeline below has every event with its source, and the sections after it tell the story in order.
Borden, Culliton and the “marketing mix” (1948-1964)
The phrase came before the model. In 1948 James Culliton, a Harvard marketing professor, wrote a research bulletin on The Management of Marketing Costs that described business executives as mixers of ingredients. His colleague Neil Borden picked up the idea; by his own account he used “marketing mix” in his teaching from around 1949, and he used it in his 1953 presidential address to the American Marketing Association. Borden’s 1964 article in the Journal of Advertising Research, “The Concept of the Marketing Mix”, set out the ingredients he had in mind and credited Culliton for the metaphor.
In 1960 E. Jerome McCarthy’s textbook Basic Marketing: A Managerial Approach condensed that list into product, price, place and promotion, the four Ps that most marketers still learn first. It’s worth keeping in mind that none of this was statistical. The marketing mix was a management idea, a way of saying that the levers interact and have to be blended, and that framing is the reason a true marketing mix model is supposed to include price, distribution and promotions, not just media (I explain the practical difference in media mix modeling vs marketing mix modeling).
Econometricians measure advertising’s lagged effect (1954-1976)
The statistical half of the story starts in economics. In 1954 the economist Leendert Koyck published Distributed Lags and Investment Analysis, which formalised a lag whose weights decline geometrically over time, so that a cause in one period has a fading effect in each period after. Applied to advertising, that’s exactly the idea of carryover, and the “Koyck lag” is mathematically the same thing as the geometric adstock used in Robyn and in the model on the what is marketing mix modeling page. In 1962 Marc Nerlove and Kenneth Arrow gave the economic rationale in “Optimal Advertising Policy under Dynamic Conditions” in Economica, treating advertising as an investment in a stock of goodwill that depreciates.
Two years later Kristian Palda’s The Measurement of Cumulative Advertising Effects (1964) became one of the first book-length attempts to estimate those effects from real sales and advertising data. By 1976 there were enough such studies for Darral Clarke to publish “Econometric Measurement of the Duration of Advertising Effect on Sales” in the Journal of Marketing Research, comparing how long the different estimates said advertising lasted. In my experience the carryover you estimate depends a lot on whether the data is daily, weekly or monthly, which is why I’m cautious about anyone quoting an adstock rate as a fixed fact about a channel.
Adstock, scanners and the packaged goods era (1974-1990s)
The commercial practice needed data, and it arrived at a checkout. On 26 June 1974 a pack of Wrigley’s Juicy Fruit gum became the first product scanned with a Universal Product Code, at a Marsh supermarket in Troy, Ohio. As scanning spread, consumer goods manufacturers could buy weekly, store-level sales data instead of waiting for bimonthly audits. Information Resources, Inc. (IRI) was founded in Chicago in 1979 by John Malec and Gerald Eskin to sell exactly that kind of data, and IRI and Nielsen became the two big suppliers to packaged goods brands, later bundling marketing mix models into their data contracts. IRI merged with NPD in 2022 and the combined company became Circana in 2023.
The same year IRI was founded, Simon Broadbent published “One Way TV Advertisements Work” in the Journal of the Market Research Society (that’s how it’s usually cited; the original isn’t available online), describing advertising as building a stock of effect that decays when the ads stop. Broadbent is credited with coining “adstock”, and the word stuck so well that it’s now the default parameter name in nearly every MMM library. The UK developed its own strong tradition in this period: the IPA (the trade body for UK agencies) established its Effectiveness Awards in 1980, and many of their published cases leaned on econometric evidence that advertising paid back, which is one reason MMM in the UK is still often just called “econometrics” (see marketing econometrics).
The academic reference for this era is Market Response Models: Econometric and Time Series Analysis by Dominique Hanssens, Leonard Parsons and Randall Schultz, first published in 1990 with a second edition in 2001. It pulled together the methods for estimating how sales respond to price, advertising, distribution and competition, and much of what we now call saturation curves, baseline sales and long-term effects is laid out there. If you want to read one book from before the digital era, it’s this one (more in marketing mix modeling books).
Digital attribution takes over, and its limits show (2000s-2010s)
When advertising moved online, many marketers stopped needing MMM, or thought they did. Ad servers and analytics tools could follow individual users from impression to click to purchase, and attribution models split the credit for each conversion between touchpoints. That was faster and far more granular than a quarterly regression, and for a decade the natural home of MMM shrank back to big consumer brands with TV budgets. I ran a growth agency in this period that spent over $10 million a year on Facebook ads, and almost all of our day-to-day measurement was platform reporting and attribution.
However, research during the 2010s showed how fragile that measurement was. Randall Lewis and Justin Rao’s 2015 paper in the Quarterly Journal of Economics, “The Unfavorable Economics of Measuring the Returns to Advertising”, argued that individual sales are so noisy that even very large experiments struggle to pin down advertising returns precisely (see the Lewis and Rao explainer). Brett Gordon, Florian Zettelmeyer and colleagues then compared observational methods against large randomized experiments at Facebook (published in Marketing Science in 2019) and found that the observational estimates often failed to recover the experimental lift, the argument laid out in Facebook experiments vs observational methods. Attribution’s rule-based models were also quietly retired: Google Analytics stopped offering first click, linear, time decay and position-based models in November 2023. The comparison is covered in MMM vs multi-touch attribution.
Google’s papers, privacy and open source (2017-2024)
In 2017 a group of Google researchers published three papers that shaped modern MMM. Yuxue Jin, Yueqing Wang, Yunting Sun, David Chan and Jim Koehler’s “Bayesian Methods for Media Mix Modeling with Carryover and Shape Effects” combined a delayed adstock and a Hill saturation curve in a Bayesian model and showed how much the priors matter when data is limited (our walkthrough of Jin et al.). David Chan and Mike Perry’s “Challenges and Opportunities in Media Mix Modeling” listed the ways MMMs go wrong (explained here), and Sun et al. extended the model to geo-level hierarchical data (explained here). Those papers are the direct ancestors of LightweightMMM and Meridian.
Then the privacy changes arrived. GDPR applied across the EU from 25 May 2018, and Apple’s App Tracking Transparency came into force with iOS 14.5 on 26 April 2021, requiring apps to ask before tracking users across other companies’ apps and websites. For advertisers who relied on Facebook and app attribution, a large part of what they could see disappeared, and demand for a method that only needed aggregate data came back fast. The tools followed: Uber open-sourced its Orbit forecasting package in 2020; Meta officially released Robyn in November 2021, with ridge regression and an automated search over adstock and saturation settings; Google published LightweightMMM on GitHub in February 2022; and PyMC Labs made the first release of PyMC-Marketing in February 2023. Google opened its successor to LightweightMMM, Meridian, to selected advertisers in limited availability on 7 March 2024, and in December 2024 Meta announced a Python version of Robyn in beta.
This was also the period when I started teaching MMM. The early Vexpower course taught marketing mix modeling in Excel and Google Sheets first, building adstock and regression by hand, then moved on to lessons on Robyn in R and LightweightMMM and Orbit in Python, along with log-log, geo-level, Bayesian and deep learning models. The current course has been rebuilt around 30 lessons and one invented brand, Northstar, which you can follow in the marketing mix modeling example.
Calibration, Meridian and AI (2025-2026)
Google made Meridian available to everyone on 29 January 2025, and LightweightMMM’s README now says it’s no longer supported and recommends switching; Google archived the repository on 19 January 2026. The cookie story ended with a whimper rather than a bang: in April 2025 Google said Chrome would keep its current approach to third-party cookie choice, and on 17 October 2025 it announced it was retiring most of the Privacy Sandbox APIs that had been meant to replace cookies. Third-party cookies are still blocked by default in Safari, consent rates still limit tracking, and the case for aggregate measurement hasn’t weakened.
The most important shift of the last two years is that experiments have moved inside the model. Meridian 2.0, released on 2 September 2026, added prior calibration from incrementality experiments such as its own GeoX geo tests, and PyMC-Marketing supports calibrating with lift tests; this is the subject of calibrating MMM with experiments. The other shift is AI. PyMC Labs described an AI agent for running MMMs on PyMC-Marketing in February 2025 and released a beta in November 2025, and general-purpose coding agents can now write a working model from a CSV in minutes. In my experience they’re good at the code and at the documentation, and they make the same mistakes a junior analyst would with controls, holdouts and causal claims, which is what AI marketing mix modeling covers.
Why the history of marketing mix modeling matters for your model
If you read a Meridian or Robyn configuration with this history in mind, most of it stops being mysterious. The adstock parameters are Koyck and Broadbent; the saturation curves are the diminishing returns that Hanssens and colleagues documented; the trend and seasonality terms are the baseline that packaged goods modelers always needed; the priors are the Bayesian approach from Google’s 2017 papers; and the calibration settings are the field’s answer to Lewis and Rao and the Facebook experiments. For a structured way to learn each of those pieces, the Vexpower marketing mix modeling course follows roughly the same order, and the research papers page collects the key papers with links. The method is older than it looks, and that’s mostly reassuring: the problems people hit today are problems the field has been working on for seventy years.
