Media Mix Modeling:
Where to Spend Next

Media mix modeling (MMM) estimates how advertising across channels contributes to sales, revenue, or another business outcome using aggregated historical data. It supports budget decisions by estimating channel returns and the effects of changing spend. As explained in Google’s causal inference guide, those estimates depend on assumptions about the factors influencing both advertising and demand.

What media mix modeling measures#

Media mix modeling examines advertising within the broader marketing mix, which can also include pricing and promotions. Meta’s analyst guide describes statistical analysis of marketing and non-marketing contributions to a business outcome, including channel contribution and budget allocation.

MMM works with totals across periods and, where available, geographic areas. It does not require linking each advertising touchpoint to an individual purchase. This use of aggregate observational data makes cross-channel measurement possible without reconstructing individual customer journeys.

The outcome, market, period, and channel definitions should be explicit before modeling begins. A budget question about total revenue requires a different interpretation from a question about conversion counts. Define the decision in those terms so the resulting estimates have a clear business meaning.

How a media mix model works#

An MMM relates a business outcome to media activity and other relevant factors through a statistical model. Google’s Meridian model specification includes geographic effects, trend and seasonality, controls, non-media treatments, transformed media inputs, and an error term. A simplified description of that structure is:

Outcome = time and geographic effects + other modeled factors + media effects + error

The model estimates the relationships among these terms. Its media contribution estimates represent effects under the chosen model and assumptions. The model specification represents seasonal demand, non-media treatments such as price changes, and advertising through distinct terms.

Two media transformations are particularly important for interpreting the results.

Adstock: effects that continue after exposure#

Adstock represents the delayed effect of advertising. In Meridian’s adstock formulation, current and earlier media activity are combined using weights that describe how the effect changes over time. Advertising in an earlier period can therefore contribute to an outcome in a later period. The decay parameters are estimated, while the maximum lag duration is a modeling choice.

Saturation: changing returns as activity increases#

Saturation represents diminishing marginal returns as media activity grows. Meridian uses a Hill function to model saturation. This allows the estimated benefit of additional exposure to depend on the current level of activity. A channel’s response need not remain proportional to spending throughout its operating range.

What data does MMM need?#

The Meridian data specification describes the following inputs:

Align the inputs to the same periods and markets. Meridian’s outcome and exposure requirements call for a summable outcome, such as total sales, rather than a rate such as click-through rate. Spend can serve as an exposure proxy, but changes in media prices can make spending a misleading measure of advertising volume.

For its own framework, Google’s data collection guidance gives a rough minimum of two years of weekly data for geographic models and three years for national models. It recommends at least three years when only monthly data is available. These are planning guidelines for Meridian, rather than universal guarantees of model quality.

History length is only part of readiness. The data sufficiency guidance also considers model complexity, spend variation, and the width of estimated credible intervals. More channels and controls require more information. Geographic data can help, but observations across places and periods are not equivalent to independent experimental observations.

Highly correlated channel activity makes separate effects harder to estimate, as Meta’s discussion of multicollinearity explains. Define channels around budget decisions and check for separate variation. Combine channels or obtain additional evidence when estimates remain too uncertain.

Why controls matter for incremental effects#

Confounding occurs when a factor affects both media execution and the business outcome. Controls are intended to account for those shared influences. Google’s control variable guidance distinguishes confounders from variables that merely predict the outcome and from mediators that sit between advertising and its effect.

The distinction changes what the model estimates. Including a mediator as an ordinary control can bias the estimated total media effect because it adjusts for part of the pathway through which advertising works. Collecting more variables therefore does not automatically improve the answer.

Start control selection with the information used to set budgets, allocate channels, and choose high-spend periods. Document how those factors relate to the outcome. The causal rationale for including or excluding a variable should be understandable alongside its statistical treatment.

Reading MMM results: contribution, ROI, and marginal ROI#

Google’s definitions of incremental outcome and returns distinguish four outputs:

For a revenue outcome, historical ROI expresses incremental revenue per media dollar. A profitability decision also needs the value remaining after relevant business costs; state the financial objective separately from the modeled revenue metric.

Marginal ROI is particularly relevant to additional spending. A strong historical average can coexist with a lower marginal return as a channel approaches saturation. Read both the current position on the response curve and the estimated benefit of the proposed change.

Using MMM to decide where to spend next#

Separate reallocating a budget from expanding it#

A fixed-budget decision moves money among channels while holding the total constant. A flexible-budget decision changes the total subject to a return target. Meridian’s budget optimization scenarios support both: fixed-budget allocation can maximize modeled incremental outcome, while flexible-budget allocation can use a minimum overall ROI or a channel-level marginal ROI target.

Set the planning period, total budget, and channel spending bounds before optimizing. Minimum commitments and practical spending limits belong in these bounds. Compare only allocations that can actually be implemented. The objective and constraints define what an optimal allocation means.

Compare the proposed change with the current allocation#

Meridian’s optimization reporting compares current and recommended channel budgets, modeled returns, and incremental outcomes. Its response-curve charts also show the spending bounds used in the optimization.

For a reviewable budget proposal, record the current allocation, proposed allocation, expected outcome difference, and constraints. Evaluate the entire transfer: increasing one channel while reducing another changes both contributions. A historical channel ranking alone does not describe that trade-off.

Check assumptions about future buying conditions#

The same budget can buy different amounts of media when advertising prices change. Google’s future-budget guidance identifies cost per media unit, revenue per outcome unit, and the distribution of activity across time and geography as relevant planning inputs. A future budget estimate should state which of these conditions are expected to change.

Changing those inputs alters the planning calculation, but does not itself refit Meridian’s learned parameter distribution. New assumptions about future costs are therefore distinct from new evidence about media effectiveness.

Keep extrapolation visible#

The fitted media response depends on the observed range of activity. Google’s saturation guidance cautions about extrapolation beyond that range. Mark spending proposals that extend beyond historical levels and seek additional evidence before making large changes. A smooth response curve does not make an unobserved budget level a measured result.

MMM versus multi-touch attribution and experiments#

These methods use different evidence and answer different questions.

Multi-touch attribution assigns credit across the marketing events observed before conversion. The Google and Booking.com attribution paper explains that this credit is not inherently defined as incremental advertising impact. Path reports also depend on the events they observe; missing exposure records limit the journey they can describe.

Use MMM for aggregate allocation questions, attribution for recorded-path questions, and experiments to test specified incremental effects. Google’s experiment relevance criteria emphasize checking the outcome definition, population, period, and spending alternative. Compare those definitions before interpreting differences between reported results.

How to assess whether an MMM is credible#

Google’s model assessment guidance treats out-of-sample prediction as a preliminary check, while emphasizing causal assumptions and plausible results. Close agreement between predicted and observed sales is insufficient evidence that each channel’s incremental contribution is correct.

A useful review covers:

• Prediction: Check performance on held-out observations and investigate unexplained departures from the outcome history.
• Plausibility: Investigate an implausible baseline or a channel contribution that dominates without a defensible explanation.
• Uncertainty: Report credible intervals with estimates so the apparent difference between channels is not mistaken for certainty.
• Sensitivity: Examine whether reasonable changes in assumptions materially alter the spending recommendation.

Treat these checks as evidence about model behavior. They do not turn observational data into a randomized comparison.

Use experiments for calibration and validation#

Experiment results can provide independent information for an MMM. Meridian’s calibration guidance describes using incrementality experiments and other domain knowledge to inform channel priors: the information supplied before fitting the model to its data.

Check the experimental period, duration, population, channel definition, and spending baseline before using a result. A reduced-spend test measures a different comparison from a model estimate against zero spending. Calibration must account for those differences and translation uncertainty.

For validation, seek an experiment that measures the same causal quantity as the MMM. Calibration supplies information to the model; validation evaluates its claims against suitable evidence. Keep that distinction explicit when reporting confidence in a budget recommendation.

When to use MMM, and how to get started#

MMM is most useful for a decision that requires aggregate channel contribution or a comparison of media budgets. The readiness test is whether the available history can support those estimates at the required level. Google’s data sufficiency guidance recommends reducing scope or adding information when the data cannot support model complexity, and generally positions Meridian as a channel-level tool rather than a campaign-level measurement system.

Choose an experiment for a focused intervention that can be tested directly. Choose path attribution for a question about observed journeys. Repair the data or narrow the model before seeking precise channel returns from incomplete or weakly varying histories.

Open-source implementations include Google’s Meridian and Meta’s Robyn. Meridian’s documented workflow moves from data collection and cleaning to model fitting, result assessment, and budget analysis.

A practical starting sequence is:

1. Define the outcome, markets, planning period, and budget changes under consideration.
2. Assemble aligned outcome, spend, exposure, and control histories.
3. Document assumptions about demand, delayed effects, and saturation.
4. Fit the model and review uncertainty, plausibility, and relevant experiment results.
5. Compare feasible allocations using response curves and marginal returns.
6. Record the chosen allocation and expected outcome, then review new evidence as it becomes available.

A useful MMM budget report states how much money would move, between which channels, over which period, and the expected change in the business outcome. Include uncertainty and a comparison with the current allocation.