Media Mix Modeling: What It Can and Cannot Tell You

Media mix modeling is a statistical approach that estimates how much each marketing channel contributes to a business outcome, typically sales or revenue, by analyzing historical data across channels over time. It does not track individual users. It works at an aggregate level, using regression analysis to isolate the effect of each input while controlling for variables like seasonality, price changes, and economic conditions.

That distinction matters more than most people realize. MMM is not attribution. It is not a dashboard. It is a model, which means it is a structured approximation of reality built on assumptions. Understanding what those assumptions are, and where they break down, is the difference between using MMM well and being misled by it.

Why MMM Has Come Back Into Fashion

Media mix modeling has been around since the 1960s. It was standard practice at large consumer goods companies for decades, then fell out of fashion when digital attribution came along and promised something more exciting: the ability to see exactly which ad drove which sale, in real time, at the individual user level.

That promise was always overstated. Digital attribution models, particularly last-click, have been systematically misleading marketers about channel contribution for years. They overweight the final touchpoint, ignore offline influence, and cannot account for the brand awareness that made someone searchable in the first place. I spent years watching brands cut brand budgets because the performance data told them to, only to see their search volumes quietly erode six months later.

Then third-party cookies started disappearing. iOS privacy changes made click-based attribution increasingly unreliable. Walled gardens like Meta and Google started returning less data. Suddenly, aggregate modeling that does not depend on individual tracking started looking attractive again. That is the honest reason MMM is back in the conversation. Not because marketers suddenly discovered its virtues, but because the alternative got worse.

If you are thinking about measurement more broadly, the Marketing Analytics hub covers the full landscape, from GA4 configuration to attribution frameworks to reporting infrastructure.

What MMM Actually Measures

At its core, MMM uses regression analysis to decompose your sales or revenue into component parts. You feed it time-series data: weekly or monthly spend by channel, alongside your outcomes data, and a set of control variables. The model then estimates the contribution of each channel to the outcome, after accounting for everything else.

Those control variables are where most implementations go wrong. A well-specified MMM will include things like seasonality, base sales trends, pricing changes, promotional activity, macroeconomic indicators, and competitor actions where data is available. A poorly specified one will attribute the effect of a summer sales spike to whatever channel happened to be running at the time.

The outputs you typically get from an MMM are channel contribution percentages, return on ad spend estimates by channel, saturation curves showing diminishing returns at different spend levels, and sometimes scenario planning tools that let you model what happens if you reallocate budget. These are genuinely useful. But they are estimates, not measurements. The model does not observe what would have happened if you had spent differently. It infers it from the data you have.

I have seen MMM outputs presented to boards as though they were audit-grade financial figures. They are not. They are informed estimates with confidence intervals that often do not get shown in the executive summary. That gap between what the model says and how it gets communicated is one of the most persistent problems in how organizations use this tool.

The Data Requirements Are Non-Negotiable

MMM is data-hungry. The minimum viable dataset is typically two to three years of weekly data across all channels, with consistent definitions throughout. If your business changed how it categorizes spend midway through, or if you have gaps in the data, or if a channel was only active for four months, the model will struggle to isolate its effect reliably.

This is a real constraint that rules out MMM for a lot of businesses. Startups and younger brands often do not have the history. Companies that have been through acquisitions, rebrands, or significant operational changes may have data that is technically there but not comparable across periods. Businesses with highly seasonal or irregular spend patterns need even more data to produce stable estimates.

Early in my career, I learned a version of this lesson the hard way. We were trying to make sense of channel performance with less than a year of clean data, and the analysis kept producing results that contradicted what we knew from direct observation. The data was not wrong. There just was not enough of it to separate signal from noise. No amount of analytical sophistication fixes a thin dataset. You are just producing confident-looking nonsense faster.

The quality of your spend data matters as much as the quantity. If your channel definitions are inconsistent, if agency fees are sometimes included and sometimes not, if you are mixing net and gross figures, the model will pick up those inconsistencies and try to explain them as signal. Getting your data house in order before commissioning an MMM is not optional preparation. It is the work.

MMM vs. Attribution: Different Questions, Not Competing Answers

One of the more frustrating debates in marketing analytics is the framing of MMM versus attribution as though you have to pick one. You do not. They answer different questions, operate at different time horizons, and have different blind spots. Used together, they are more useful than either is alone.

Attribution, whether last-click, data-driven, or something more sophisticated, works at the individual session or user level. It is good at telling you which campaigns and keywords are driving conversions within a channel. It operates in near real time. It is terrible at measuring channels that do not leave a digital footprint, like TV or out-of-home, and it is structurally biased toward lower-funnel activity because those touchpoints are closer to the conversion event.

MMM works at the aggregate level over longer time horizons. It can capture the effect of offline channels. It is better at measuring brand investment because the effect of brand advertising shows up in the baseline sales trend over time. But it is slow, expensive to run properly, and cannot tell you which specific creative or keyword is performing. It also cannot tell you what happened last Tuesday.

The practical approach is to use MMM for strategic budget allocation decisions across channels, and attribution for tactical optimization within channels. When the two disagree, that disagreement is usually informative. It often means attribution is overcrediting a channel that benefits from brand investment MMM is capturing, or that MMM is missing something about a newer channel with limited historical data. Either way, the tension is worth investigating rather than resolving by ignoring one of them.

For a clearer picture of how attribution fits into a broader measurement approach, this Moz piece on GA4 directional reporting is a useful framing for how to think about analytics as a guide rather than a verdict.

Where MMM Breaks Down

MMM has well-documented limitations that do not always make it into vendor pitches. Worth knowing them before you commission one.

The first is multicollinearity. If you run TV and digital video at the same time, every time, the model cannot separate their effects. It will attribute the combined effect to whichever variable is easier to explain statistically, which may not be the right one. Brands that have never varied their channel mix significantly will get back a model that reflects their habits more than their actual channel performance.

The second is that MMM is backward-looking by design. It tells you what worked historically, in the market conditions that existed, at the spend levels you were running. Extrapolating those findings to a future with different competitive dynamics, different creative, or significantly different spend levels requires judgment that the model cannot provide. Saturation curves are particularly prone to misuse here. The point at which a channel starts showing diminishing returns in your historical data is not a fixed law of physics. It changes with creative quality, audience targeting, and market conditions.

The third is the black-box problem. More sophisticated Bayesian MMM implementations can be harder to interrogate than simpler regression models. When a model tells you that TV drove 23% of sales, the natural question is: how does it know that? The answer involves a set of prior assumptions, variable specifications, and model choices that the vendor may not fully disclose. This does not make the output wrong, but it does mean you should be asking more questions than most clients do.

I judged the Effie Awards for several years, which means I spent time looking at effectiveness cases from brands across categories. The ones that impressed me were not the ones with the most sophisticated measurement apparatus. They were the ones where the team clearly understood what their measurement was and was not telling them, and made decisions accordingly. That kind of epistemic honesty is rarer than it should be.

Build It In-House or Commission It?

There are three realistic options for running an MMM: build it in-house, use a specialist vendor, or use one of the open-source frameworks that have become available in recent years, most notably Meta’s Robyn and Google’s Meridian.

Building in-house requires a data scientist who understands time-series regression, a clean and comprehensive data pipeline, and enough organizational patience to iterate on model specifications. It is viable for larger organizations with mature data infrastructure. The advantage is full transparency into the model and no vendor dependency. The disadvantage is that it takes longer to get right than most people expect, and the first version is usually not very good.

Specialist vendors offer speed and experience. They have run models across many categories and can bring pattern recognition that an in-house team building their first model cannot. The risk is that vendor models can be opaque, and some vendors are better at producing convincing-looking outputs than at producing accurate ones. Ask to see validation methodology. Ask how they test whether the model is actually predictive. If the answer is vague, that is a red flag.

Open-source frameworks like Robyn have democratized MMM considerably. They are free, reasonably well-documented, and have active communities. They still require someone who can run them competently and interpret the outputs critically. The code being open does not mean the model is automatically right. But they have made it possible for mid-size brands to run credible MMM work that would previously have required a six-figure vendor engagement.

For organizations investing in the data infrastructure to support this kind of analysis, exporting GA4 data to BigQuery is worth understanding as part of building a more complete picture of your marketing data.

How to Use MMM Outputs Without Being Misled by Them

The most important thing to understand about MMM outputs is that they are directional, not definitive. A model that tells you paid social is delivering a 2.1x ROAS and TV is delivering 1.4x is not telling you to cut TV and put everything into paid social. It is telling you something about the relative efficiency of those channels in the historical period, at the spend levels you were running, in the market conditions that existed. That is useful information. It is not a budget decision.

When I was running agencies and managing significant media budgets, the value of analytical tools was never in following what they said mechanically. It was in using them to surface questions worth asking. If the model says a channel is underperforming, the right question is: why might that be, and is there something we know about this channel that the model cannot capture? Sometimes the answer confirms the model. Sometimes it reveals a model specification problem. Both outcomes are useful.

A few practical principles for using MMM outputs well. First, always look at confidence intervals, not just point estimates. A channel contribution estimate with wide confidence intervals is telling you the model is uncertain. Treating that as a precise figure and making large budget decisions on it is a mistake. Second, validate the model against holdout periods. If you can, run a controlled experiment where you reduce spend in one channel and see whether the model’s prediction of the sales impact matches what actually happens. This is the most direct test of whether the model is working. Third, update the model regularly. An MMM built on data from two years ago is not telling you about today’s market.

The broader analytics ecosystem matters here too. MMM works best when it sits alongside other measurement approaches, not as a standalone oracle. Understanding how your analytics platform captures user behavior helps you identify where aggregate models are filling gaps that session-level data cannot cover, and where they are potentially double-counting effects that attribution is also capturing.

The Organizational Challenge Nobody Talks About

Running an MMM is the easy part. The hard part is getting an organization to act on it, especially when the outputs challenge existing assumptions about which channels matter.

MMM frequently tells brands that their brand investment is more valuable than their performance marketing, and that some of their highest-volume digital channels are less efficient than they look in attribution reports. These findings are often correct. They are also politically inconvenient, because they suggest reallocating budget away from channels where performance is easy to demonstrate toward channels where it is harder to see in real time.

I have been in rooms where MMM findings were presented clearly and correctly, and then quietly set aside because the head of paid search did not want to hear that their channel was overcredited, or because the CFO wanted a metric they could check every week rather than a model they had to trust. This is not a measurement problem. It is a culture problem. And no amount of analytical sophistication solves it.

The organizations that get the most value from MMM are the ones where senior leadership genuinely understands what the model is and is not, where there is tolerance for the uncertainty that honest modeling involves, and where budget decisions are made on the basis of the best available evidence rather than the most convenient evidence. That combination is less common than you might hope.

If you want to build the kind of measurement culture where MMM findings can actually influence decisions, the broader work of connecting analytics to commercial outcomes is worth reading about. The Marketing Analytics hub covers that territory in more depth, from how to structure measurement frameworks to how to make reporting operationally useful rather than decorative.

What Good MMM Practice Actually Looks Like

Good MMM practice starts before the model runs. It starts with a clear question. Not “tell us what’s working” but something more specific: “We are considering a significant reallocation from digital to broadcast. What does the historical data suggest about the likely impact on sales?” A specific question produces a more useful model specification and a more actionable output.

It continues with data audit and preparation. Every channel in the model should have consistent, complete spend data for the full modeling period. Control variables should be identified and sourced before modeling begins, not retrofitted afterward when results look unexpected. This preparation work is unglamorous. It is also where most of the value is created or destroyed.

The modeling itself should involve iteration and validation. A first-pass model is rarely the right model. Variable selection, transformation choices, and prior specifications all affect the output. The team running the model should be able to explain why they made the choices they made, and what the sensitivity of the results is to those choices.

Outputs should be presented with appropriate uncertainty. Point estimates alone are misleading. Scenario ranges are more honest and more useful for decision-making. And the findings should be triangulated against other evidence: attribution data, incrementality tests where they exist, qualitative knowledge of the market. Where they align, confidence increases. Where they conflict, investigation is warranted.

Finally, the outputs should feed into a decision, not just a presentation. If the MMM findings do not change anything about how budget is allocated or how channels are evaluated, the exercise has produced insight without value. That is a waste of the investment and, more importantly, a missed opportunity to make better decisions.

About the Author

Keith Lacy is a marketing strategist and former agency CEO with 20+ years of experience across agency leadership, performance marketing, and commercial strategy. He writes The Marketing Juice to cut through the noise and share what works.

Frequently Asked Questions