The Marketing Mix Model Renaissance: Why Google and Meta Open-Sourced Their Playbooks

What marketing mix modeling actually is, and is not

Marketing mix modeling predates the digital advertising it is now being asked to rescue. In its standard form it treats a business outcome, sales or bookings observed as a time series, as the dependent variable, and regresses it on the marketing inputs recorded over the same window: spend by channel, price, distribution, seasonality, and exogenous factors such as weather or macroeconomic conditions. Two transforms distinguish it from a naive regression. An adstock or carryover transform encodes the fact that an impression can influence a purchase days or weeks later; a saturation transform encodes diminishing returns, the empirical reality that the tenth thousand dollars of spend on a channel rarely returns what the first thousand did.

The consequential property is what the method does not require. It never follows an individual user from ad to conversion. It reads populations and periods, not people. That is the reason it has re-emerged: a measurement method that regresses aggregate outcomes on aggregate inputs is structurally indifferent to whether cookies, device identifiers, or cross-app tracking still function. When the substrate that individual-level attribution depends on erodes, an aggregate causal-inference method loses nothing, because it was never standing on that substrate in the first place.

The tracking substrate that MMM was built to outlast

The renaissance is not intellectual fashion; it is a response to a dated, documented event. In April 2021, Apple's App Tracking Transparency framework required apps to obtain explicit permission before tracking users across other companies' apps and websites. Industry measurement reports describe a sharp reduction in individual-level tracking accuracy that followed, and the consistent direction of those reports matters more than any single figure.

The specific magnitudes cited across ad-tech vendor reporting should be read as directional rather than settled: figures include roughly three-quarters of iOS users declining cross-app tracking, an estimated 30 to 40 percent drop in Meta ad-tracking accuracy, and pixel-based e-commerce revenue attribution falling from a capture rate in the region of 80 to 95 percent down toward 60 to 70 percent. These numbers originate with firms that have a commercial interest in the narrative, so the correct posture is to treat them as evidence of a real, dated degradation of a known magnitude range, not as precise constants. What is not in dispute is the industry's response: a pronounced shift toward server-side and consented signals, such as conversion APIs, and toward aggregate methods, marketing mix modeling and geo-experiments, that do not depend on individual identifiers at all.

Why observational attribution was never the ground truth

To understand why the platforms reached for an aggregate causal method rather than patching the old one, it helps to see that the old one had a deeper problem than privacy. The dominant alternative, following individual users and assigning credit along their path, is an observational method: it reads who converted after seeing an ad, not what the ad caused.

The most rigorous test of that method against reality remains Gordon and colleagues' 2019 study. The authors ran fifteen large-scale randomized field experiments on Facebook, drawing on more than 500 million user-experiment observations and 1.6 billion ad impressions, and compared the true causal lift each experiment measured against what standard observational attribution methods would have estimated from the same data. The observational methods frequently produced effect estimates in the wrong direction or of the wrong magnitude, and this held even after conditioning on rich demographic and behavioral covariates. The implication is uncomfortable and clarifying at once: the problem with individual-level attribution was never only that tracking broke. Even when the tracking worked, the method it fed could systematically miscredit. Privacy loss removed a tool that was, on the evidence, unreliable for causal claims to begin with.

Multi touch attribution versus marketing mix modeling

The pivot is best understood as a change of method, not a change of vendor. Multi touch attribution is identity-based: it stitches together the touchpoints an identifiable individual encountered and distributes credit across them, and it therefore inherits both the fragility of identity resolution and the observational bias documented above. Marketing mix modeling is population-based: it asks what the aggregate outcome would have looked like at different levels of aggregate input, and infers the contribution of each channel from that relationship.

The "MMM vs MTA" framing that now circulates in the measurement community is really a question about what kind of claim is being made. Multi-touch attribution answers "which touchpoints did this converting person see," a descriptive question about a recorded path. Marketing mix modeling attempts "how much did this channel contribute to the outcome," a causal question about a counterfactual. The first degrades directly with signal loss and, per the field-experiment evidence, was never a trustworthy causal estimator. The second was designed for the counterfactual question from the start and is unmoved by the loss of individual identifiers. That asymmetry, more than any single privacy headline, is why the largest platforms chose to rebuild using the second method and to publish the rebuild.

Google Meridian and Meta Robyn: the playbooks went open

The strongest evidence for the thesis is a fact about behavior, not opinion: the two largest advertising platforms open-sourced their internal builds of marketing mix modeling, a step well beyond simple endorsement. A company protects what it believes confers advantage and publishes what it wants the market to adopt. Both firms chose to publish.

Geo-experiments and incremental ROAS: the causal check that needs no pixel

A model is only as trustworthy as what calibrates it, and the calibration that pairs naturally with marketing mix modeling is the geo-experiment. Rather than holding out individual users, a geo-experiment holds out matched geographic markets: some regions keep receiving a campaign while comparable regions do not, and the difference in outcomes, estimated with a synthetic-control method, is the incremental lift. Because the unit of holdout is a place, not a person, the method sidesteps both privacy erosion and the last-touch bias of observational attribution, and it yields incremental ROAS, the return on the revenue a campaign genuinely caused rather than the revenue that merely occurred nearby.

Meta's open-source GeoLift library is the widely used version of this method. An independent head-to-head simulation study of open-source geo-testing tools reported GeoLift's statistical coverage at 92 to 95 percent, closest to the 95 percent target, and its false-positive rate at 3 to 5 percent, the lowest of the tools compared. That study is a vendor-run simulation rather than peer-reviewed academic work, so those comparative numbers belong in the emerging, industry-grade tier and should be cited as one study's finding, not a settled result. The broader point stands on firmer ground: geo-experiments are used industry-wide as the causal anchor that keeps an aggregate model honest, and the two methods are designed to work together, the experiment measuring truth on a slice, the model generalizing it across the whole.

The caveat: national playbooks, local businesses

The evidence that Meridian and Robyn exist, and that both were open-sourced as tracking collapsed, is established. The evidence that either performs well for a single-location small business is not. Both models were built and validated on the data environments their authors inhabit: large brands with substantial, multi-market, multi-year spend histories and enough signal to fit a stable regression. A med-spa or a home-services contractor running one or two channels in one metropolitan area occupies a very different statistical world.

This is a scale problem, not a moral one. Aggregate causal methods, like the controlled experiments they are calibrated against, have a power floor: they need enough variation across enough periods or markets to separate a real effect from noise. A single small firm frequently sits below that floor. A national-scale method does not downscale cleanly, and a firm-level media-mix number is not something a single small business's own data can support. The credible path is aggregation, pooling many comparable small businesses so that the statistical power no single one of them possesses can be recovered across the set, in the empirical-generalization tradition of marketing science rather than the single-firm regression. Until such pooled evidence is published, any per-business causal claim deserves the skepticism the field-experiment record has taught the discipline to apply to convenient numbers.

What the renaissance means for reading a marketing number

The marketing industry's founding parable about measurement is a line usually attributed to John Wanamaker: half my advertising is wasted, I just do not know which half. The attribution is itself unverified. Quote investigators trace the earliest documented match to a secondhand account of a 1919 speech, and the same sentiment has been attached to William Lever and William Wrigley among others. The parable is apposite precisely because it is apocryphal: the discipline's origin story about not knowing what works is a claim nobody can source, and that is a fitting emblem for a field that has repeatedly adopted convenient metrics faster than it has verified them.

Read against that history, the open-sourcing of Meridian and Robyn is not a triumphal end to the measurement problem. It is the latest correction in a long pattern: a convenient, computable metric, individual-level attribution, was treated as truth at scale until a harder method, randomized experiment, showed it wandering from ground truth, and the industry moved toward an aggregate causal method that had been quietly reliable all along. The correct posture for a business is neither to venerate the new models nor to dismiss them, but to ask of any marketing number the questions this episode vindicates: is it causal or merely observed, is it calibrated against an experiment, and does the data behind it actually reach the scale the claim requires. A measurement practice that answers those questions in writing, and reports uncertainty rather than hiding it, is continuing this corrective tradition rather than claiming to have ended it.