The statistical method underneath modern marketing measurement is roughly 140 years old. The application of it to marketing is roughly 50. The recent open-source democratization of it is roughly five. Knowing the arc helps explain why some current arguments — including the one declaring attribution dead — sound novel but aren't, and why the methods now reaching B2B marketers have been refined for decades inside an industry most of us don't work in. Brief, accurate, and aimed at CMOs and marketing leaders, not statisticians.
The word "regression" doesn't come from marketing. It comes from Victorian-era genetics — specifically, a observation about peas and human heights that turned into one of the foundational techniques of modern statistics.
In the 1880s, the British polymath Francis Galton was studying inheritance — how traits pass from parents to offspring. He noticed something counterintuitive: when tall parents had children, the children were on average shorter than the parents. When short parents had children, the children were taller. The extremes regressed toward the average of the population. He published the observation in 1885 and 1886, and the term he used to describe it — "regression toward mediocrity" — stuck.
What Galton didn't fully realize was that the line he'd drawn through his height data — the line that best summarized the relationship between parents and children — was the seed of a much bigger idea. Over the next several decades, Karl Pearson formalized the related concept of correlation (1896), G. Udny Yule extended the math to handle more variables (1907), and Ronald Fisher synthesized the modern regression model in 1922 — building on Carl Friedrich Gauss's earlier method of least squares from the early 1800s.
By the mid-1920s, regression analysis was a recognized statistical method: a way to estimate the relationship between a dependent variable (what you're trying to explain) and one or more independent variables (the factors that might explain it), while quantifying how much of the variation is captured by your model versus left to chance. The math hasn't fundamentally changed since.
Several of the statisticians who developed these methods — Galton, Pearson, and Fisher in particular — were also active proponents of eugenics, and some of their statistical work was originally developed in service of those ideas.
The math they produced is sound and is now used responsibly across science and industry. The history is worth knowing honestly. The methods stand on their own; the people who developed them had views we don't share.
Regression didn't go straight from statistics to marketing. It went through economics first — and the marketing version inherited the shape of the economic version, which is part of why it works the way it does.
In the 1930s and 40s, economists Jan Tinbergen and Ragnar Frisch developed econometrics — the application of statistical methods to economic data, particularly the use of regression to model how economic variables relate to each other over time. They shared the first Nobel Prize in Economic Sciences in 1969 specifically for this work. The point isn't the Nobel — it's that by the late 1960s, the idea of using regression on time-series data to estimate causal-looking relationships between aggregate variables was a fully developed, academically respected discipline.
Marketing scholars and practitioners noticed. In the 1960s and 1970s, they began applying econometric methods to a specific problem: how much of a CPG brand's sales could be attributed to advertising spend, versus pricing, versus promotions, versus seasonality, versus everything else? Statisticians at the University of Chicago and elsewhere built the first formal marketing mix models in the 1970s. Building one took months. The work was done by econometricians, not marketers.
The 1980s saw early commercial adoption — Coca-Cola is often cited as one of the first major brands to use MMM rigorously, and Kraft, Procter & Gamble, AT&T, and Pepsi followed. The pioneers of commercial MMM as a service emerged at the end of that decade: Hudson River Group in 1989 and Marketing Management Analytics (MMA) in 1990. These firms productized what had been a bespoke consulting exercise, and Nielsen and IRI eventually bundled MMM into their standard data contracts. For about thirty years, this was a Fortune 500 capability — accessible only to companies that could afford the data, the consultants, and the year-long modeling cycles.
CPG had three things in abundance that made early MMM possible: huge data volumes from store sales tracking (Nielsen, IRI), broad channel diversity (TV, radio, print, in-store), and short purchase cycles where the lag between ad and sale was measurable in days or weeks.
B2B has roughly none of those things. Which is why MMM was developed somewhere else, refined for decades in a domain that suits it, and only recently started to be adapted for environments it wasn't designed for. The genetic code is CPG.
For a couple of decades, marketing measurement took a different path — one that didn't need regression at all. That detour shaped a generation of marketers, and it's the reason much of the current discourse exists.
When digital advertising emerged in the late 1990s and exploded in the 2000s, it brought a seductive new capability: deterministic tracking. You could see, with apparent certainty, which ad a user clicked, which page they landed on, and whether they bought. The need for statistical modeling seemed to evaporate. Why estimate the relationship between spend and sales when you could just watch the user do it?
This was the era of attribution as an industry name — touch-based, click-based, cookie-based credit assignment. First-touch, last-touch, multi-touch, U-shaped, W-shaped. Vendors built products around assigning credit to specific interactions. The aggregate, statistical approach of MMM looked old-fashioned. It was tied to TV and offline, not the dynamic, trackable digital world. For about fifteen years, MMM was something CPG dinosaurs did. Everyone else was tracking clicks.
The detour ended for reasons that had nothing to do with statistics: cookie deprecation, Apple's App Tracking Transparency, GDPR and other privacy regulations, walled gardens that stopped exposing user-level data, and the realization that deterministic tracking had always been more deterministic-looking than deterministic-actually. Suddenly, the aggregate statistical methods that had been around since the 1970s started looking relevant again — not as a replacement for digital tracking, but as a companion to it.
In the last five years, regression-based marketing measurement has gone from a Fortune 500 consulting engagement to an open-source library anyone can download. That shift is the reason this conversation is suddenly reaching B2B.
In 2020, Meta released Robyn, an open-source MMM library written in R. For the first time, the math that had been guarded behind consulting fees was free to download. Robyn used ridge regression and adstock modeling and was specifically designed to let advertisers measure the contribution of their media spend, including spend on Meta's own platforms — which was Meta's commercial motivation for releasing it, though the tool itself was genuinely useful regardless.
In 2022, Google released LightweightMMM, a Bayesian MMM library in Python. In January 2025, Google followed with Meridian, the official successor to LightweightMMM and Google's more advanced, actively-maintained open-source MMM framework. PyMC Labs built PyMC-Marketing, a Python Bayesian MMM library with a strong open-source community. Uber contributed Orbit KTR. (Meta has since discontinued active development of Robyn; Meridian and PyMC-Marketing are now the two most actively maintained options.)
What changed isn't the math. The math has been sitting there since the 1920s, and the marketing application since the 1970s. What changed is access. A technique that required a $250,000 consulting engagement in 2010 is now a free download. AI assistance is making the surrounding judgment work — data preparation, model specification, interpretation — increasingly tractable for teams that don't have a data scientist on staff. The democratization is real, and it's relatively new.
"Open source" does not mean "easy to use." Downloading Meridian or PyMC-Marketing gets you the math. You still need data prep, statistical judgment, and someone who can interpret what the model produces.
Forrester and others note that successful open-source MMM programs still require strong in-house data engineering and data science capacity, or a services partner. The democratization is at the math layer, not the operating layer — though that's changing too.
Francis Galton observes "regression toward mediocrity" in his studies of inheritance, and invents the first crude form of linear regression.
Karl Pearson develops correlation theory. G. Udny Yule extends regression to handle multiple variables. The framework starts to look modern.
Ronald Fisher unifies Gauss's least squares with the Pearson/Yule correlation framework. The math underneath all modern regression analysis is essentially in place.
Tinbergen and Frisch develop econometrics — the application of regression to economic time-series data. They share the first Nobel Prize in Economic Sciences in 1969.
Researchers begin applying econometric methods to marketing data, primarily in CPG. The University of Chicago is one center of early work. Building a model takes months.
Coca-Cola, Kraft, P&G, AT&T, and Pepsi adopt MMM as part of marketing planning. The method requires a dedicated team of econometricians.
Hudson River Group (1989) and Marketing Management Analytics (1990) launch as the first MMM-as-a-service firms. Nielsen and IRI eventually bundle MMM into their data contracts.
Digital advertising's deterministic tracking promises individual-level attribution. The aggregate statistical methods of MMM look old-fashioned. A generation of marketers learns to think in clicks, not coefficients.
Cookie deprecation, ATT, GDPR, walled gardens, and the realization that deterministic tracking was never as deterministic as it looked. MMM and other aggregate methods come back into the conversation.
Meta releases Robyn under an open-source license. For the first time, advertisers outside the Fortune 500 can download and run MMM-class statistical models for free.
Google's LightweightMMM (2022), PyMC Labs's PyMC-Marketing, and Uber's Orbit KTR expand the open-source ecosystem. Bayesian MMM becomes the modern default.
Google releases Meridian as the official successor to LightweightMMM — Bayesian, geo-level, with integrations for Google Search and YouTube data. Mainstream attention follows.
AI assistance starts to make the surrounding judgment work — data prep, model specification, interpretation — tractable for teams without dedicated data scientists. The math layer was solved decades ago; the operating layer is solving now.
A few things worth taking from the arc.
First — these methods are old, well-tested, and not faddish. When someone pitches you regression-based marketing analytics in 2026, they're not selling something new. They're selling the application of a hundred-year-old statistical framework that's been refined inside the most rigorous quantitative marketing organizations in the world for fifty years. The novelty is the accessibility, not the math.
Second — the methods were developed in CPG, not B2B. The genetic code is short-cycle, high-volume, broad-channel consumer marketing. B2B has different shape — longer cycles, smaller volumes, narrower channel mix. The methods adapt, but the adaptation is genuine work. Treating CPG MMM as a turnkey B2B solution is a category error.
Third — the digital attribution era was the detour, not the destination. Cookie-based, deterministic, individual-user tracking was a fifteen-year experiment that's now ending. The methods coming back into the conversation aren't new — they're the older, more durable methods that predate the detour. If you've been in marketing long enough to remember when MMM was something only big CPG brands did, that's where the field is heading again, with better tools.
Fourth — the democratization is recent and ongoing. Most of what's now possible for a B2B team to do with these methods wasn't possible five years ago. That's worth knowing, because some of the current discourse — including the "attribution is dead, MMM is the future" argument — is reacting to capabilities that genuinely didn't exist when most of today's marketing leaders learned their craft. The skepticism that comes from "we tried MMM in 2010 and it didn't work for us" is honest, but the world has changed.
The math is 100+ years old. The marketing application is 50+ years old. The democratization is 5 years old. The B2B fit is still being worked out.
This document is background — the historical context for the rest of the Field Notes publication. If you've read this far and want the operational arguments: the main hub makes the case that the category is misnamed, the Progression companion helps you place your team on the maturity map, and the MMM companion goes deeper on whether marketing mix modeling specifically is right for your situation.