Media Mix Modelling: When Attribution Gets Complicated

Media mix modelling is a statistical approach to marketing attribution that estimates the contribution of each channel to business outcomes, typically revenue or sales, by analysing historical spend and performance data across your entire marketing programme. Unlike click-based attribution, it works at an aggregate level and can account for channels that leave no digital fingerprint, including TV, radio, out-of-home, and word of mouth.

It has been around since the 1960s and is currently experiencing a significant revival, largely because cookie deprecation and privacy regulation have made last-click attribution even less reliable than it already was. If you are trying to understand what your marketing is actually doing across a complex channel mix, media mix modelling is one of the most honest tools available, but only if you understand what it can and cannot tell you.

Why Attribution Is a Problem Worth Taking Seriously

When I was running a performance marketing operation and managing significant paid search budgets across multiple clients, the attribution conversation came up constantly. Every channel claimed credit. Paid search would show a conversion. Display would show a view-through. Email would show a click the same day. Add them up and you had attributed three times the actual revenue. Nobody wanted to hear that their channel was getting partial credit at best.

The attribution problem is not a technical problem at its core. It is a political one. Channels are managed by different teams, agencies, or platforms, and each has an incentive to report the most favourable numbers. The tools they use are designed to support that. Last-click attribution became dominant not because it was accurate but because it was simple and it gave paid search teams very good-looking reports.

If you want a broader view of how measurement frameworks fit together across a marketing programme, the Marketing Analytics hub covers the full landscape, from data infrastructure to channel-level reporting. Media mix modelling sits at the more sophisticated end of that spectrum, but the principles that make it useful or useless are the same ones that apply to any measurement approach.

Forrester has written clearly about how standard marketing measurement can undermine your understanding of the buyer experience by focusing on touchpoints rather than outcomes. That tension sits at the heart of why MMM has come back into fashion. It does not try to track individuals. It looks at the whole and works backwards.

What Media Mix Modelling Actually Does

At its most basic, a media mix model is a regression analysis. You take a time series of your business outcomes, usually weekly or monthly revenue or sales volume, and you regress it against a set of input variables: spend by channel, pricing data, seasonal indices, promotional activity, distribution changes, competitor spend where available, and any other factors that might influence the outcome.

The model then estimates the coefficient for each variable, which tells you how much each input contributed to the output. From those coefficients, you can derive return on investment by channel, which you can then use to inform budget allocation decisions.

In practice, it is more complicated than that for several reasons. First, advertising effects are not instantaneous. A TV campaign you ran in October may still be influencing purchases in December. MMM handles this through adstock transformations, which model the carry-over and decay of advertising effects over time. Different channels have different decay rates. TV tends to have longer carry-over than paid search. Getting those adstock parameters right matters a great deal.

Second, advertising effects are not linear. Spend the first pound on TV and it does almost nothing. Spend enough and you hit diminishing returns. MMM handles this through saturation curves, which model the relationship between spend and response. Those curves are what allow you to model the theoretical optimal budget allocation across channels.

Third, channels interact. A brand campaign on TV may make your paid search more efficient by increasing brand search volume. A promotional email may amplify the effect of a social campaign running at the same time. Good MMM attempts to capture these interaction effects. Simple MMM often does not, which is one reason simple models can mislead.

The Data Requirements Nobody Warns You About

I have sat through more than a few MMM presentations where the headline finding was that TV drove 60% of revenue and digital drove 15%, and the recommendation was to shift budget dramatically towards TV. In some cases that was probably right. In others, it was an artefact of the data, specifically the fact that the model had two years of TV data and six months of digital data, and the digital channels had never been given enough budget to show their full effect.

The data requirements for a credible MMM build are more demanding than most people expect. You typically need at least two years of weekly data, ideally three. You need spend data that is clean, consistent, and granulated by channel, not lumped into broad categories. You need outcome data that is similarly clean and ideally not subject to major definitional changes mid-period. And you need data on all the non-marketing factors that influence your outcome: pricing, distribution, competitor activity, macroeconomic conditions, seasonality.

The non-marketing factors matter because if you leave them out, the model will attribute their effects to your marketing. If you had a price cut in Q3 and did not include pricing in your model, the model will give credit for the resulting sales uplift to whatever marketing was running at the time. That is not a minor modelling issue. It can completely distort your channel ROI estimates.

When I was working with a retail client on a budget reallocation exercise, we found that the previous MMM had been built without any promotional calendar data. The brand ran deep discounts in January and August every year. The model had attributed the resulting sales spikes to the brand awareness campaign that happened to run in those windows. The recommendation had been to increase brand spend significantly. The actual driver was the promotions. Those are very different strategic conclusions.

Bayesian MMM and Why It Has Become the Standard

Traditional frequentist MMM has a significant limitation: it treats the data as the only source of information. If your data is thin, noisy, or spans a short period, the model will still produce coefficient estimates, but the confidence intervals around those estimates will be very wide, often wide enough to make the findings practically useless.

Bayesian MMM addresses this by allowing you to incorporate prior knowledge into the model. If you know from industry benchmarks or previous experiments that your TV adstock typically decays over eight to twelve weeks, you can encode that as a prior distribution rather than asking the model to estimate it purely from data. This makes the model more stable, particularly when data is limited, and it makes the uncertainty in the outputs explicit rather than hidden.

Google’s Meridian and Meta’s Robyn are both open-source Bayesian MMM frameworks that have made this approach accessible without requiring a specialist econometrics team. They have lowered the barrier to entry considerably, which is genuinely useful, but they have also made it easier to produce a model that looks sophisticated without the underlying data quality to support it. The tool is not the hard part. The data preparation and the honest interpretation of outputs are the hard parts.

How MMM Fits With Other Attribution Approaches

MMM is one of three main approaches to marketing measurement, and understanding where it sits relative to the others matters for using it correctly.

Multi-touch attribution assigns credit to individual touchpoints in a customer experience, using rules or machine learning to weight each one. It is granular, near-real-time, and works well for digital channels where tracking is available. It struggles with offline channels, with privacy restrictions that limit tracking, and with the fundamental question of whether a touchpoint caused a conversion or simply witnessed one.

Incrementality testing, usually conducted through geo-based or audience holdout experiments, measures the true causal effect of a specific channel or campaign by comparing outcomes between exposed and unexposed groups. It is the most rigorous approach but is expensive, slow, and impractical for continuous measurement across a full channel mix.

MMM sits between the two. It is not as granular as MTA and it does not establish causality the way a well-designed experiment does, but it can cover the full channel mix, including offline, and it produces budget-level strategic insights that neither of the other approaches handles well. The most honest measurement programmes use all three in combination, treating each as a different lens on the same reality rather than competing systems.

HubSpot’s writing on why marketing analytics differs from web analytics makes a related point: web analytics tells you what happened on your site, marketing analytics tells you why people came and whether it was worth the spend. MMM is firmly in the second category.

What MMM Can Tell You and What It Cannot

MMM can tell you the historical contribution of each channel to your business outcomes, expressed as a return on investment or a share of total sales. It can tell you where you are likely on the saturation curve for each channel, which informs whether you are under or over-investing relative to diminishing returns. It can tell you the relative efficiency of different channels and give you a framework for thinking about budget reallocation.

What it cannot tell you is what will happen in the future. The model is calibrated on historical data. If your market changes, if a new competitor enters, if consumer behaviour shifts, the historical relationships may no longer hold. Using MMM outputs for forward planning requires an explicit acknowledgement that you are extrapolating from the past and that the further you extrapolate, the less reliable the projection becomes.

It also cannot tell you anything about creative quality, messaging, or audience targeting within a channel. A model that tells you TV drove 40% of revenue cannot tell you whether that was because the creative was excellent or because you had strong reach. It cannot tell you which audience segments responded most strongly or whether a different execution would have performed better. For those questions, you need different tools.

Forrester’s broader point about reporting discipline applies here: the ability to produce a number does not mean that number is the right one to act on. MMM produces outputs with apparent precision, coefficients to several decimal places, ROI figures that look authoritative. That precision is partly a statistical artefact. The uncertainty in those estimates is real and should be communicated alongside the headline numbers.

Building a Model Worth Using

The practical steps to building a credible MMM are less glamorous than the modelling itself. They start with data collection and cleaning, which typically takes longer than the modelling. You need to assemble spend data by channel by week, outcome data at the same granularity, and control variables covering everything that might influence the outcome independently of marketing.

Once the data is assembled, you need to make decisions about model structure: which channels to include, how to handle channel hierarchies (do you model paid search as one variable or split it by brand and generic?), what adstock transformations to apply, and what saturation function to use. These decisions should be informed by prior knowledge and validated against the data, not made arbitrarily.

Model validation matters more than most practitioners acknowledge. A model that fits the historical data well is not necessarily a good model. Overfitting is a real risk, particularly when you have many variables relative to your data length. Standard validation approaches include holding out a portion of the data and testing how well the model predicts it, checking that the coefficients have the expected signs and magnitudes, and comparing model-implied ROIs against any available experimental benchmarks.

The output should not be presented as a single point estimate. The most useful MMM outputs show a range of plausible values for each channel’s contribution, communicate the assumptions that were made and how sensitive the results are to those assumptions, and are explicit about what the model does not include. A good analyst presents their model with appropriate humility. A bad one presents it as a definitive answer.

Tools like Google Analytics can supplement MMM by providing digital channel performance data that feeds into the model, though they are not a substitute for it. For setting up the data infrastructure that feeds your broader analytics programme, Semrush’s guide to Google Analytics setup is a reasonable starting point for the digital data layer. Similarly, Mailchimp’s overview of marketing dashboards is useful for thinking about how to surface model outputs alongside channel-level data in a format that drives decisions.

When MMM Is Worth the Investment and When It Is Not

MMM is not appropriate for every business. The minimum viable conditions are a reasonably complex channel mix (at least three or four channels with meaningful spend), sufficient data history (two years minimum, three preferred), and a business outcome that can be measured consistently over time. If you are a small business running primarily paid social and email, MMM will not tell you anything you could not learn more cheaply from proper incrementality testing.

The businesses where MMM adds genuine value are those with significant offline spend, those operating in categories where the path to purchase is long or non-linear, and those where budget decisions are made at a strategic level rather than purely through platform optimisation. FMCG, financial services, automotive, and retail with a significant offline component are the natural homes for MMM. Pure-play e-commerce with short purchase cycles often gets more value from more granular digital attribution approaches.

The cost of a credible MMM build from an external provider ranges considerably depending on complexity and the quality of the team. In-house builds using open-source tools like Robyn or Meridian can reduce that cost but require data science capability and a significant time investment in data preparation. Neither route is inherently better. The right choice depends on your internal capabilities and the strategic importance of the decision you are trying to inform.

What I have seen go wrong most often is not the modelling itself but the use of the outputs. A model gets built, it produces a recommendation to shift budget from digital to TV, and that recommendation gets implemented without any validation or monitoring. Twelve months later, revenue is down, and nobody connects it to the budget change because the model has been filed away and forgotten. MMM should be a living input to budget decisions, not a one-time exercise that produces a static answer.

If you want to explore how these measurement approaches connect to broader analytics strategy, the Marketing Analytics hub covers everything from data foundations to channel-specific measurement in more depth. MMM sits within a larger measurement ecosystem, and understanding that context makes the modelling decisions considerably more sensible.

For those exploring alternatives to standard analytics platforms as part of their data infrastructure, Moz’s overview of Google Analytics alternatives is worth reading alongside any MMM planning work, particularly if your current tracking setup has gaps that would affect model input quality. And if you are thinking about how to present MMM outputs alongside content and channel performance data, Buffer’s writing on content marketing metrics is a useful reminder that not everything worth measuring fits neatly into a regression.

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