What Is Mixed Media Modeling (MMM)?
Mixed media modeling (MMM) is a statistical technique that measures the contribution of each marketing channel to sales or conversions. Using historical data, it quantifies how much revenue each dollar of spend generates across TV, paid search, social, radio, display, and offline media simultaneously. Unlike single-channel attribution tools, MMM captures the full picture, including channels that leave no digital fingerprint.
Marketers use MMM to answer one question above all others: where should the next dollar go? The model produces channel-level return on ad spend (ROAS) estimates and saturation curves that show when additional investment in a given channel stops paying off.
How Mixed Media Modeling Works
MMM uses ordinary least squares (OLS) regression or Bayesian regression to decompose sales into contributions from media spend, baseline factors, and external variables. The core equation takes the following form:
Sales = Baseline + ∑(Media Contributioni) + Seasonality + Price + Promotions + Error
Each media term is transformed before entering the model to reflect two real-world phenomena:
- Adstock (carryover effect): Advertising builds memory. A TV spot run on Monday still influences purchases on Friday. Adstock decays the effect over time using a decay rate parameter, typically between 0.3 and 0.9 depending on the channel.
- Saturation (diminishing returns): The first $100,000 spent on a channel generates more incremental sales than the next $100,000. A Hill transformation or S-curve function models this diminishing return.
The transformed media variable for a single channel might look like this:
Adstocked Spendt = Spendt + (decay rate × Adstocked Spendt-1)
The model then passes that adstocked value through a saturation function before entering the regression, producing an “effective media” variable that more accurately mirrors consumer behavior than raw spend figures.
Bayesian MMM vs. Frequentist MMM
Traditional MMM used frequentist OLS regression, which works well when data is abundant and channels are relatively uncorrelated. Modern implementations increasingly use Bayesian inference, which incorporates prior knowledge (expert beliefs about likely ROAS ranges) and returns full probability distributions rather than point estimates.
Meta’s open-source Robyn library and Google’s Meridian, both released to the public in recent years, use Bayesian frameworks and have accelerated adoption by making MMM accessible without a dedicated data science team. Robyn, for instance, uses Ridge regression combined with multi-objective optimization to produce Pareto-optimal budget allocation recommendations across thousands of model iterations.
| Approach | Output | Suitable When |
|---|---|---|
| Frequentist OLS | Point estimates | Long data history (3+ years), few channels |
| Bayesian MMM | Probability distributions | Short data windows, many channels, prior knowledge available |
| Hierarchical Bayesian | Region/product-level distributions | Multi-market or multi-SKU advertisers |
Reading the Outputs: ROAS, Contribution, and Saturation Curves
An MMM run typically produces three core outputs that inform budget decisions.
Channel ROAS
ROAS from MMM reflects the incremental revenue generated per dollar spent, holding all other variables constant. A paid search channel might show a ROAS of 4.2, meaning each dollar generated $4.20 in revenue. That number differs from the ROAS reported in Google Ads because MMM accounts for baseline sales that would have occurred anyway and removes the double-counting common in multi-touch attribution models.
Decomposition
Sales decomposition separates total revenue into its component sources. For a mid-sized consumer packaged goods brand running $50M in annual media, a typical decomposition might show that 55% of sales come from the baseline (pricing, distribution, brand equity), 20% from TV, 12% from paid search, 8% from social, and 5% from promotions. That baseline figure often surprises marketers who overestimate how much their paid media is driving.
Saturation Curves
Saturation curves plot incremental return against spend level for each channel. They show where a channel is currently operating on its curve, and what marginal ROAS would be at higher or lower investment. If paid social is operating past its saturation point, the model will flag that reallocating budget to a less-saturated channel would improve total portfolio ROAS.
Data Requirements
MMM requires weekly or daily time-series data across a meaningful window, typically 2 to 5 years, to capture seasonality cycles and enough spend variation to separate channel effects. Required inputs include:
- Dependent variable: total sales, revenue, or conversions by time period
- Media spend by channel and week
- Impressions or GRPs for awareness channels (TV, radio, out-of-home)
- External variables: price index, promotions, competitor activity, economic indicators
- Seasonality flags: holidays, product launches, major events
Sparse data or channels with limited spend variation produce high standard errors, making those channel estimates unreliable. This is a known limitation when modeling new or experimental channels.
MMM vs. Multi-Touch Attribution
MMM and multi-touch attribution (MTA) measure different things and complement rather than replace each other. MMM works top-down from aggregate data, captures offline channels, and is not affected by cookie deprecation or privacy regulations. MTA works bottom-up from user-level clickstream data, offers channel and creative granularity, but misses offline touchpoints and is increasingly constrained by cookie deprecation and consent limitations.
A combined approach, sometimes called “unified measurement,” uses MTA outputs as a calibration signal within the MMM framework, improving accuracy for digital channels where user-level data is still available while preserving the offline coverage MMM provides.
Real-World Applications
P&G, one of the world’s largest advertisers by spend, publicly credited MMM analysis in the mid-2010s for identifying that significant portions of its digital media were generating near-zero incremental sales. That finding prompted a widely-discussed reallocation away from programmatic display. The company reported cutting roughly $200M in digital spend and seeing sales hold or improve. That result became a reference case for the value of rigorous measurement over channel-reported metrics.
DTC brands have used Robyn and similar open-source tools to run their own MMM internally, a task that previously required expensive media measurement vendors. A brand spending $2M annually across Meta, Google, and connected TV can now build a working MMM in weeks rather than months.
Limitations to Account For
MMM produces estimates, not ground truth. Multicollinearity between correlated channels (brands that always run TV and digital simultaneously) can inflate standard errors and make individual channel coefficients unstable. The model is also backward-looking, calibrated on past spending patterns that may not hold as markets shift. Running geo-lift experiments or holdout tests alongside MMM provides calibration checkpoints that strengthen confidence in the model’s channel-level estimates.
Practitioners also note that MMM typically cannot measure effects at the creative or audience level, which limits its usefulness for tactical optimization. For that granularity, pairing MMM with brand lift studies or platform-native experiments fills the gap.
When to Use Mixed Media Modeling
MMM is best suited for advertisers with diversified media mixes, meaningful offline spend, and enough historical data to support regression analysis. It is less useful for brands running a single channel or those with fewer than 18 months of spend history. For brands spending above roughly $5M annually across three or more channels, MMM typically delivers enough insight to justify the modeling investment through improved budget allocation alone.
As cookie deprecation continues to erode the reliability of user-level attribution, MMM has regained prominence. It is a durable measurement approach that does not depend on tracking individual users across the web.
Frequently Asked Questions About Mixed Media Modeling
What is Mixed Media Modeling (MMM) in simple terms?
Mixed media modeling (MMM) is a statistical method that uses historical sales and spend data to estimate how much each advertising channel contributed to revenue. Marketers use it to calculate channel-level ROAS and decide where to shift budget for better returns. It covers both digital and offline channels, including TV and radio, which most attribution tools miss entirely.
How is Mixed Media Modeling different from multi-touch attribution?
MMM works top-down from aggregate sales data and captures all channels, including offline media that leaves no digital footprint. Multi-touch attribution tracks individual user journeys through digital touchpoints and offers creative-level detail, but misses offline media entirely. The two approaches measure different things and work best in combination.
How much data do you need to run a Mixed Media Model?
MMM typically requires two to five years of weekly or daily data to capture seasonal patterns and enough spend variation to isolate channel effects reliably. Brands with fewer than 18 months of history will get less reliable results, especially for channels with limited or inconsistent spend.
What is the difference between Bayesian MMM and traditional MMM?
Traditional MMM uses ordinary least squares (OLS) regression and returns single point estimates. Bayesian MMM incorporates prior knowledge about likely ROAS ranges and returns full probability distributions, making it more reliable when data windows are short or channels are numerous. Meta’s Robyn and Google’s Meridian both use Bayesian frameworks.
Does Mixed Media Modeling work without third-party cookies?
Yes, MMM is unaffected by cookie deprecation. Unlike multi-touch attribution, MMM works entirely from aggregate sales and spend data and does not depend on tracking individual users across websites. This makes it a durable measurement method as browser-level privacy restrictions continue to limit user-level tracking.