Marketing Mix Modeling Is Back: Building the Modern Measurement Stack

Marketing mix modeling is back, and the reason is simple: the measurement method that replaced it for fifteen years no longer works the way it used to. User-level, click-based attribution depended on following individuals across sites and devices — and privacy changes, browser restrictions, and walled gardens have made that increasingly impossible. MMM never needed to track individuals in the first place. That's why a decades-old technique is suddenly the most relevant tool in the stack.

But "MMM is back" is the easy part. The harder, more useful question is how it fits with the methods you already use. The answer isn't to pick one. It's to build a stack where each method does what it's actually good at.

Why did MMM come back now?

Because signal loss broke the alternative. Deterministic attribution assumes you can observe most of the journey. As we covered in our post on third-party cookies in 2026, that assumption no longer holds — Safari and Firefox block third-party cookies by default, App Tracking Transparency suppressed device IDs across iOS, consent rules limit what you can collect, and the walled gardens report their own results inside their own boundaries.

MMM is structurally immune to most of that. It uses aggregate, time-series data — weekly spend by channel, plus outcomes like sales or sign-ups, plus external factors like seasonality and promotions — and statistically estimates each channel's contribution. No user-level tracking, no cookies, no device IDs. The same forces breaking attribution leave MMM essentially untouched.

The market has noticed. According to December 2024 IAB data cited by eMarketer, 56% of US ad buyers said they would focus at least somewhat more on MMM in 2025. A separate eMarketer and TransUnion survey found nearly 47% of US brand and agency marketers plan to invest in MMM over the following year. This isn't a niche revival; it's a mainstream shift in how serious advertisers measure.

How do MMM, incrementality, and attribution fit together?

They answer different questions, on different timeframes, with different strengths. The mistake is treating them as competitors. They're a triangulation system, and each covers the others' blind spots.

Attribution is your daily optimization layer. It's granular and fast — which keyword, which creative, which audience. It's also the most damaged by signal loss and inherently correlational: it tells you what happened before a conversion, not what caused it. Useful for steering campaigns, dangerous as the basis for big budget decisions.

Incrementality testing is the causal anchor. You run a controlled experiment — a geo holdout, where you turn a channel off in some regions and compare against matched control regions, or a user-level holdout where a segment is deliberately excluded — and measure the actual lift. This is the closest thing to ground truth, because it isolates cause rather than correlation. Geo experiments are especially valuable for channels where user-level holdouts are impractical, like TV, audio, and large-scale digital.

MMM is the strategic allocator. It takes the whole picture — every channel, online and offline, plus seasonality and external factors — and estimates each channel's contribution and diminishing returns. It answers the question attribution can't: given a fixed budget, where should the next dollar go?

The methods reinforce each other. Incrementality results calibrate the MMM so it reflects real causal lift, not just historical correlation. The MMM provides the strategic frame that attribution data fills in tactically. And attribution flags day-to-day shifts that might warrant a new experiment. This is the core of modern analytics and attribution practice — and it's what lets performance reporting move from "here's what the platforms claimed" to "here's what actually drove the business."

What about the open-source MMM tools?

This is the development that made MMM practical for more than just the biggest advertisers. Two free frameworks now anchor the space.

Google Meridian is an open-source MMM framework built on Bayesian causal inference. Google made it available to everyone in early 2025, as the successor to its earlier LightweightMMM library. Its defining strengths: it reports a full posterior distribution for each channel's ROI — a central estimate with a credible interval that honestly reflects uncertainty rather than a single false-precision number — and it can integrate incrementality experiment results as priors, so the model is calibrated against real causal lift. It's Python-based and available on GitHub.

Meta Robyn is Meta Marketing Science's open-source MMM package, hosted on GitHub under facebookexperimental. It uses machine learning techniques — ridge regression, evolutionary algorithms for hyperparameter tuning, time-series decomposition — to estimate channel efficiency, adstock, and saturation. It was built first in R and now also ships a Python version, and it's MIT-licensed.

A few honest caveats. "Open-source" means the code is free, not that the work is. Both tools require statistical expertise to set up correctly, sufficient data history (typically a couple of years of weekly data), and meaningful variation in spend to produce trustworthy estimates. A poorly specified MMM doesn't fail loudly — it produces confident, wrong numbers. The model is the easy part; the discipline around inputs, validation, and calibration is where the value is.

How should you actually build the stack?

Start with the decision, not the tool. The biggest waste in measurement is building sophisticated models nobody acts on.

1. Define the decisions you need to make. Budget allocation across channels is an MMM question. "Is this channel even working" is an incrementality question. "Which creative should we scale" is an attribution question. Match the method to the decision.

2. Fix your data foundation first. MMM needs clean, consistent spend and outcome data by channel and week. Most organizations underestimate how much cleanup this requires. This is unglamorous marketing infrastructure work, and it determines whether everything downstream is trustworthy.

3. Run incrementality tests on your biggest line items. Before you trust any model's channel estimates, validate the channels that consume the most budget with a geo or holdout experiment. These results become both a sanity check and a calibration input for the MMM.

4. Layer in MMM for strategic allocation. Once you have clean data and a couple of calibration experiments, an MMM — Meridian, Robyn, or a vendor model — turns it into forward-looking budget guidance.

5. Keep attribution for daily optimization, but stop treating it as truth. It's directional. The strategic decisions come from the MMM-plus-incrementality layer.

The bottom line

MMM came back because the privacy era broke the alternative, not because it's a fad. But the win isn't swapping attribution for MMM — it's building a measurement system where attribution optimizes day to day, incrementality proves causality, and MMM allocates the budget, each calibrated against the others. Open-source tools like Meridian and Robyn have lowered the cost of entry dramatically. The hard part was never the model. It's the data discipline and the willingness to act on the answer.

Sources

• FAQ: What marketers need to know about marketing mix modeling — eMarketer
• Nearly half of US marketers plan to invest in MMM over the next year — eMarketer
• Meridian is now available to everyone — Google Blog
• google/meridian — GitHub
• facebookexperimental/Robyn — GitHub
• All you need to know about geo holdout testing — Funnel
• Update on Plans for Privacy Sandbox Technologies — Privacy Sandbox (Google)