Media mix modelling estimates each channel’s contribution to a business outcome using aggregated time-series data, and it is best used for budget allocation, scenario planning and forecasting. It works across long time horizons, captures both immediate and delayed effects, and depends on validation steps, including confidence intervals and multicollinearity checks, to be trustworthy.
Media mix modelling, often shortened to MMM, is a statistical approach that uses historical, aggregated data, typically weekly or daily sales, spend and external factors, to estimate how much each marketing channel contributes to a business outcome. It treats the business as a system: media, pricing, distribution and seasonality all move together, and the model’s job is to separate their individual effects. The approach uses aggregated time-series data to estimate the contribution of different media channels to sales or other business outcomes, which makes it fundamentally different from methods that track individual users.
Multi-touch attribution (MTA) works at the user level, stitching together click and impression paths to credit specific touchpoints. It is granular and fast, but it struggles with walled gardens, cookie deprecation and channels that influence buyers without a trackable click, such as television or out-of-home. We have covered this contrast in more depth in our piece on multi-touch attribution for paid media, and the same tension applies to organic channels, which we address in our attribution guide for SEO. Controlled experiments, such as geo holdouts, isolate causality cleanly but are expensive to run at scale and cannot cover every channel simultaneously.
MMM sits above both. It answers questions that MTA and experiments cannot easily answer alone:
Each method has a role. The strongest measurement programmes use MMM for strategic allocation, MTA for tactical optimisation within digital channels, and experiments to validate both.
Three concepts do most of the analytical work inside a media mix model, and understanding them changes how you read any output a data team hands you.
1. Adstock, or carry-over effect. Advertising rarely converts the moment someone sees it. A television advert seen on Monday might still be influencing a purchase decision the following week. Adstock models this by decaying each week’s media impact over subsequent weeks rather than assuming it disappears instantly. A channel with a high decay rate, such as search, has its effect concentrated in the week of spend; a channel with a slow decay rate, such as brand television, keeps contributing for several weeks afterwards.
2. Saturation, or diminishing returns. Spend does not convert at a constant rate. The first £10,000 on a channel typically buys more incremental outcome than the next £10,000, because audiences are finite and repeated exposure yields less each time. Saturation curves, often S-shaped, capture this so that a model does not naively assume returns scale linearly with budget.
3. Base versus incremental volume. Not every sale is caused by media. Base volume is what would have happened anyway, driven by brand equity, distribution, pricing and habit; incremental volume is what media activity adds on top. Getting this split right is the entire point of the exercise, because an inflated base estimate understates every channel’s contribution, while an inflated incremental estimate overstates it.
Lag structures and interaction terms refine this further. Lag structures let a channel’s effect peak a week or two after spend rather than immediately, which matters for channels with longer consideration cycles. Interaction terms capture cases where two channels reinforce each other, such as search activity converting better when brand television is also running. Adstock models carry-over effects and saturation models capture diminishing returns, and both are standard components in the modelling libraries used across the industry.
When you review a model, ask what is being reported alongside the headline ROI figures:
Pro Tip: Ask any analyst presenting MMM results to show you the saturation curve for your two largest channels. It tells you instantly whether there is room to spend more before returns flatten.
A media mix model is only as reliable as the data feeding it, and most practical failures trace back to this stage rather than the mathematics.
The essential inputs are the outcome series (sales, leads or revenue), media spend by channel, impressions or GRPs where available, pricing and promotional activity, distribution or availability, and seasonality markers. Missing any of these risks the model attributing someone else’s effect to media.
Granularity matters as much as coverage. Weekly data is the common default because it balances enough observations with meaningful variation in spend. Daily data can add noise without adding insight unless spend patterns genuinely shift day to day. Geo-level data, where sales and spend are tracked by region rather than pooled nationally, is often preferable because it introduces far more variation to learn from: regions that increased spend can be compared against regions that did not, which sharpens the model’s ability to separate channel effects from background trend. High-quality model inputs, especially geo-level data, and standardised datasets are major differentiators for reliable MMM estimates.
Cleaning and alignment are unglamorous but decisive. Timestamps across platforms need reconciling to the same week-ending convention, spend needs mapping to a consistent channel taxonomy so that, for example, paid social is not split inconsistently between platforms month to month, and missing data needs a documented strategy rather than silent interpolation. Skipping this step is the single most common reason two analysts produce different ROI figures from the same raw exports.
You do not need to build a model yourself to ask the right questions about one. A handful of specification choices explain most of the variation you will see between vendors and in-house teams.
Ordinary least squares (OLS) regression is the simplest and most transparent approach, well suited to a single market with a reasonably long history. Hierarchical, or mixed-effects, models extend this by sharing information across geographies or business units, which helps when any one region has too little data on its own. Bayesian specifications go further still, incorporating prior beliefs, such as results from past experiments, and producing credible intervals rather than single-point estimates.
Multicollinearity is the practical issue to watch for most closely. When two channels move together, such as paid search and paid social rising and falling in the same weeks, the model struggles to credit either one precisely. Ask for variance inflation factors (VIFs) for every channel; a VIF above roughly 5 to 10 signals a collinearity problem worth investigating before trusting the ROI split. Autocorrelation, where this week’s error is correlated with last week’s, is another common issue, addressed through lagged terms or autoregressive error structures rather than ignored.
Marginal ROI reporting with uncertainty: practitioners caution that ROI coefficients without confidence intervals can be misleading, and VIF checks for multicollinearity are critical to interpreting any single-channel figure correctly.
Credible intervals, the Bayesian equivalent of confidence intervals, tell you how much uncertainty surrounds each estimate. A narrow interval signals a channel with plenty of variation to learn from; a wide one signals a channel where the data cannot yet distinguish its effect confidently from noise. Scenario simulation then takes these parameter estimates and runs them forward: given the fitted saturation and adstock curves, what happens to the outcome if channel A’s budget rises by 20% and channel B’s falls by the same amount. This is where uplift calculations for planning purposes are actually constructed, not from the raw coefficients alone.
A coefficient on its own is not a budget decision. The step that matters is converting it into marginal ROI: not what the average pound returned historically, but what the next pound would return given where spend currently sits on the saturation curve. A channel can show strong average ROI while sitting near saturation, meaning the next increment of spend returns far less than the historical average suggests.
When presenting to stakeholders who are not analysts themselves, language matters as much as the numbers. “Search returns £2.10 to £4.40 per pound spent, with £3.20 as our central estimate” gives a decision-maker something honest to act on; a single point figure invites false confidence.
Scenario planning itself is a simple, repeatable exercise once the model is fitted: define two or three plausible budget shifts, run each through the model’s saturation and adstock parameters, and compare the forecast outcome for each against the status quo. This is how MMM earns its place in annual planning cycles rather than sitting as a one-off research exercise.
One factor planners underweight is what media choice signals about a brand, independent of direct response. Media choice can influence consumers’ willingness to pay, with channel selection affecting price perception in controlled experimental work. A brand that shifts entirely toward performance channels may protect short-term efficiency while quietly eroding the pricing power that premium or brand-building channels support. Allocation decisions built purely on short-term marginal ROI can miss this, which is why mature planning pairs MMM outputs with a brand-lift view; we walk through how to run that kind of study in our brand lift measurement guide.
Pro Tip: When two channels show similar marginal ROI, favour the one with the narrower credible interval. Precision is worth more than a marginally higher central estimate when the budget decision is sizeable.
A model that has not been validated against something outside itself is a hypothesis, not an answer. Treat validation as a required stage, not an optional extra.
Where experimental lift data exists, feeding it into the model as a Bayesian prior, rather than treating it as a separate check performed afterward, tends to produce estimates that hold up better against future out-of-sample data. This single step separates measurement programmes that stay credible over multiple planning cycles from those that quietly lose trust after the first disputed quarter.
Across client engagements, the recurring failure point in media mix modelling is rarely the mathematics. It is the lag between a model producing an insight and a business acting on it, often stretched across weeks of manual data wrangling and report formatting. Our workflows compress that cycle: data pipelines that once took days to reconcile across platforms now run on standardised, repeatable processes, which means the model gets refreshed on actual business cadence rather than whenever an analyst has a spare week.
Patterns drawn from engagements feed into predictive benchmarks that help calibrate a new model faster than starting from a blank sheet. The internal evidence worth collecting alongside any MMM effort includes CRM-level signals on lead quality, experimentation tags from any geo or holdout tests already run, and geo-segmentation in the underlying data wherever the business footprint allows it. Each of these sharpens the model’s ability to separate real channel effect from background noise, and each ties the measurement work back to the outcome that actually matters: customer engagement analytics, revenue, not impressions. We go into how we connect channel activity to revenue outcomes in our piece on measuring the growth engine properly.
Media mix modelling is a statistical method that uses aggregated, historical time-series data on sales, spend and external factors to estimate how much each marketing channel contributes to a business outcome. It supports long-term planning decisions such as budget allocation and forecasting rather than tracking individual customer journeys.
MMM works with aggregated, channel-level data over time and is suited to strategic budget decisions and long-term effects, while multi-touch attribution tracks individual user journeys to credit specific touchpoints within a conversion path. MTA offers more granularity for digital optimisation, but struggles with channels that lack trackable clicks, such as television, which MMM can still capture.
A typical media mix might combine television, radio, out-of-home, paid search, paid social, display and email, each playing a different role across the funnel. Some channels, such as linear television or print, are frequently undervalued in planning despite delivering outsized returns in some contexts, similar to a ‘moneyball’ approach to spotting overlooked value.
The two terms are used interchangeably in practice and refer to the same statistical approach to estimating channel contribution from aggregated data. “Marketing mix modelling” is the slightly older, more formal term, while “media mix modelling” has become common shorthand as the technique has focused increasingly on media channels specifically.
There is no universal split, but industry analysis has found an optimal budget mix sitting close to an even balance between online and offline spend, with the exact figure varying by business and category. The right balance for any given business depends on its specific saturation curves and base versus incremental split, which is precisely what a properly validated model is built to reveal.
