Media mix modeling is one of the oldest techniques in marketing measurement, Procter & Gamble used early versions of it in the 1960s to understand how television advertising drove sales, and one of the most relevant techniques right now. As third-party cookies deprecate, iOS privacy restrictions tighten, and platform attribution becomes less reliable, MMM is having a significant resurgence. Not because it is new, but because it never required user-level tracking in the first place.
This guide covers what media mix modeling actually measures, how the main statistical approaches differ, what the open-source and vendor landscape looks like, and when MMM is worth the investment.
MMM takes historical spend and revenue data as inputs and returns channel contributions as outputs
What MMM Actually Measures
At its core, media mix modeling is a regression problem. You are trying to explain a business outcome, weekly sales, revenue, app installs, leads, as a function of a set of inputs: your media spend by channel, promotional activity, pricing, seasonality, and external factors like competitive activity or macroeconomic conditions.
The model fits coefficients to each variable: how much does an additional $10,000 in television spend drive in revenue, controlling for everything else? How much of the seasonal uplift in November is attributable to your promotions versus organic demand? The output is a decomposition of your business outcomes into components, what drove each dollar of revenue.
Baseline vs. Media-Driven Sales
One of the most important outputs of any MMM is the sales decomposition: how much of your revenue would have happened without any advertising (the baseline), and how much your various channels contributed on top of that?
Most companies are surprised by this number. For established brands, 50-70% of sales typically come from baseline, from organic demand, brand equity built over years, word of mouth, and existing customer relationships. The advertising is driving incremental demand on top of a substantial floor.
This does not mean advertising is not working. It means you are operating in a market where demand already exists, and your advertising is capturing and expanding that demand. The MMM tells you how efficiently each channel is doing that work.
Adstock and Carryover
Advertising effects do not end the moment your campaign stops running. A TV ad someone saw on Monday might influence a purchase on the following Saturday. A brand campaign running in Q1 might contribute to Q2 sales as brand awareness builds. MMM captures this through adstock (sometimes called carryover): a decay function that models how advertising effects diminish over time after the campaign runs.
Each channel has its own adstock rate. Television typically has a carryover effect of 4-8 weeks, meaning roughly half the effect from a TV ad is felt within the first week, but some portion continues to influence behavior for weeks afterward. Paid search has near-zero carryover, when someone clicks a paid search ad, the effect is immediate; if they do not convert that session, the ad is unlikely to influence them weeks later.
This is one reason MMM and attribution produce different channel ROI estimates. Attribution sees the final click or touchpoint before conversion. MMM captures the long-run effect of brand investment that attribution systematically misses.
Saturation Curves
Every channel has diminishing marginal returns: the first dollar spent on a channel tends to drive more incremental revenue than the thousandth dollar. As spend increases, the audience becomes saturated, you are showing ads to people who have already seen them, or the remaining reachable audience has lower purchase intent.
MMM models this through response curves (also called saturation curves): the relationship between spend and return at different investment levels. The curve tells you where you are on the diminishing returns slope, whether you are in the efficient zone, or whether additional dollars in that channel are delivering less and less.
The marginal ROI (mROI), the return on the next dollar spent, is more useful for optimization decisions than the average ROI across all historical spend. A channel might show 3x average ROI because you spent wisely at the start, but now be at 0.8x mROI because you have pushed into heavily diminishing returns territory.
Frequentist vs. Bayesian MMM
Traditional media mix modeling used ordinary least squares (OLS) regression, the standard statistical approach that minimizes the sum of squared errors between predicted and actual outcomes.
OLS MMM has a significant problem: it needs a lot of data. You typically need 2-3 years of weekly observations to get statistically reliable coefficients. If you have gaps in your data (channels that were paused for months, promotions that only ran once), the model struggles. And OLS produces point estimates, single best-guess coefficients, without any uncertainty quantification.
Modern Bayesian MMM addresses these limitations by incorporating prior knowledge into the model. Bayesian statistics allows you to specify what you already believe about a parameter before seeing the data, then update those beliefs based on the observed evidence. The output is a posterior distribution, a range of plausible values for each coefficient, rather than a single point estimate.
This matters practically. You can tell a Bayesian MMM: "Based on industry research, television advertising for this category has a carryover rate of roughly 60-80% per week." The model will use that prior as a starting point and update based on your specific data. If your data is consistent with the prior, the estimate will be precise. If your data conflicts with the prior, the model will flag the inconsistency.
Bayesian MMM is more robust with limited data, produces uncertainty estimates that are crucial for decision-making, and is more interpretable when the priors are well-specified. This is why both Google Meridian and Meta Robyn, the leading open-source MMM tools, are Bayesian.
Open-Source vs. Vendor MMM
The MMM landscape has bifurcated into two distinct options: open-source tools that are free but require significant technical expertise, and vendor solutions that cost significantly more but come with interpretation, consulting, and support.
Open-Source Tools
Google Meridian (released 2024) is Google's open-source Bayesian MMM built in Python using JAX and TensorFlow Probability. Its most distinctive feature is a native calibration framework, a structured approach to validating and adjusting MMM results using geo experiments. This calibration step is important because MMM coefficients can be biased if the model's assumptions do not match reality; Meridian makes it easier to check and correct for that.
Meridian is technically sophisticated and relatively new. The community is smaller than Robyn's, documentation is still developing, and the geo calibration workflow requires running actual experiments before you can use the calibration outputs. If your team knows Python and has data science depth, Meridian is worth evaluating seriously.
Meta Robyn (released 2022) is Meta's open-source MMM written in R. It uses Nevergrad, a gradient-free optimization library, to automatically search thousands of model configurations and select the best-fitting one. This automation makes it more accessible than building a custom model from scratch, you can get initial results without specifying every hyperparameter manually.
Robyn has two years of community development behind it, more tutorials, more third-party integrations, and a larger user base than Meridian. If your team is stronger in R, or if you want a more mature ecosystem with more examples to reference, Robyn is a reasonable starting point.
One important caveat applies to both: they are released by advertising platforms with significant interest in the results of marketing measurement. This is not a disqualifying problem, both tools are published, reviewable, and independently usable, but it is worth acknowledging. Neither Google nor Meta has an incentive to release a tool that systematically undervalues their platforms. Validate results with external incrementality tests.
Vendor MMM Solutions
If you do not have data scientists who can build and interpret a Bayesian model, open-source MMM is not a realistic option. The tools exist, but using them well requires skills in probabilistic programming, model validation, and time series analysis that most marketing teams do not have in-house.
Vendor solutions, Analytic Partners, Nielsen Marketing Mix, Ekimetrics, Neustar (now TransUnion), and others, provide a fully managed service: data collection, model build, interpretation, scenario planning tools, and ongoing support. The cost is substantial: typical engagements range from $150,000 to $1,000,000+ per year depending on scope, number of markets, and the depth of analysis.
This is a significant investment. It is worth it for companies spending $5M+ per year on advertising where a 5-10% improvement in media efficiency more than covers the measurement cost. It is typically not worth it at lower spend levels.
What MMM Is Good At, and Where It Fails
MMM is strong for:
Understanding brand and reach channels that attribution tools cannot see. Television, podcast, OOH, and connected TV do not generate clicks or digital touchpoints. They build awareness and influence purchase decisions over long timelines. Attribution tools essentially ignore these channels. MMM includes them in the regression and can estimate their contribution.
Long-run ROI measurement. MMM captures effects over months, not hours. This matters for any campaign with a brand-building objective, awareness, consideration, or purchase intent that converts weeks or quarters later.
Budget allocation across channels. Because MMM provides a cross-channel, comparable ROI estimate (not each platform's self-reported number), it is the best tool available for deciding how to distribute budget across fundamentally different channel types.
Privacy-safe measurement. MMM needs no user-level data, no cookies, no device IDs, no email addresses. It works on aggregate spend and outcome data. As tracking restrictions tighten, this becomes an increasingly important advantage.
MMM is weak for:
Real-time optimization. You are looking at historical data. A model trained on data through last month cannot tell you which campaign is working best this week. For campaign-level, in-flight optimization, you need attribution data or direct platform signals.
Small channels. If a channel represents less than 5-10% of your total spend, the signal-to-noise ratio in the regression is often too low to produce reliable coefficient estimates. MMM cannot reliably measure a channel you spend $10,000/month on alongside channels where you spend $500,000/month.
Low-spend businesses. The rough threshold for MMM to be technically feasible and economically worthwhile is approximately $1M/year in advertising spend. Below that, the baseline variation in your sales data tends to swamp the media signal, and the cost of building the model exceeds the expected value of better decisions.
Short-run effects. MMM is designed for weekly and monthly data. Daily optimization signals, hourly bidding decisions, and creative-level performance are outside its scope.
How to Commission or Build an MMM
If you are a marketing leader evaluating whether to invest in MMM, here is the process in practical terms.
Step 1: Assess your data readiness. Gather 2-3 years of weekly data: spend by channel broken down as granularly as possible (not just "Google" but "Google Search brand," "Google Search non-brand," "Google Display," "YouTube"), plus weekly revenue or your primary KPI, plus external variables (promotions, price changes, competitor activity if available, seasonality indices). If this data does not exist or is inconsistent, fix the data infrastructure before investing in a model.
Step 2: Decide between open-source and vendor. If you have data scientists with Bayesian modeling experience and the capacity to own the model long-term, open-source is viable. If you do not, budget for a vendor. Hybrid options exist: some vendors will build a Meridian or Robyn model and manage it for you, which is cheaper than a fully proprietary solution.
Step 3: Build in validation. Any MMM output should be validated before you act on it. The standard validation approach: hold out the last 3-4 months of data from the training set, build the model on the prior data, and see how well it predicts the held-out period. If the model cannot predict well in-sample, its prescriptions for future budget allocation are unreliable.
Step 4: Plan for ongoing retraining. An MMM is not a one-time project. Markets change, channel mixes change, creative strategy changes. Plan for quarterly or annual model retraining with fresh data. Budget accordingly.
Step 5: Calibrate with incrementality data. The single most important step to improve MMM reliability is calibrating the model's coefficients with actual incrementality test results. If your model estimates a Facebook coefficient of 2.5x ROI and a geo experiment shows Facebook's actual iROAS is 1.2x, adjust the model's priors. Google Meridian's calibration framework is designed specifically for this.