Definition
Marketing mix modelling, also spelled media mix modeling, is a regression-based method that explains variation in an outcome, usually sales or revenue over weeks or months, in terms of marketing inputs and control variables. The registry's one-line summary is that it uses aggregated data to estimate the contribution of each channel to sales. Each channel's effect is modelled with two transformations that reflect how advertising behaves: adstock, the carry-over of an exposure's effect into later periods, and saturation, the diminishing return as spend in a period rises. Controls such as price, distribution, competitor activity, weather and holidays absorb the variation that marketing did not cause, and a baseline captures the sales that would occur with no marketing at all.
How it works
The model is fitted to a history of the outcome and all the inputs at the same granularity, typically weekly over a period long enough to contain several seasons and real variation in spend. The fitting estimates, for each channel, a coefficient together with its adstock and saturation parameters, which together describe a response curve: incremental outcome as a function of spend. From the curves come the outputs decision-makers use: a decomposition of past sales into baseline and channel contributions, the return on investment (ROIProfit relative to cost.) of each channel at its current level and at the margin, and an optimised allocation of a fixed budget that equalises marginal returns across channels. Modern practice uses Bayesian methods, which allow prior knowledge and experimental results to constrain the estimates, and recalibrates the model periodically as new data arrives. The marketing mix modelling basics guide covers inputs, adstock and saturation, outputs and data requirements in depth.
| Question | Attribution | Incrementality test | Marketing mix model |
|---|---|---|---|
| Unit of data | Individual touchpoints and conversions | Randomised groups or regions | Aggregated time series |
| Channels covered | Trackable digital only | Whatever can be withheld | Online, offline and walled gardens |
| Time to answer | Continuous | Weeks per test | Months to build, then periodic refresh |
| Causal strength | Low | High for the tested activity | Moderate; higher when calibrated with tests |
| Best use | Day-to-day optimisation within a channel | Validating a channel's true effect | Allocating budget across channels |
In practice
A model is commissioned or built when the budget is spread across channels that attribution cannot compare, which is most businesses with meaningful offline, television, audio or walled-garden spend. The work is mostly data assembly: a clean weekly history of spend and delivery for every channel, the outcome, and the controls, with enough variation in spend for the model to separate effects. The outputs feed the annual and quarterly allocation decisions covered in the budget allocation guide, and the guide to choosing an attribution approach places modelling against attribution and experiments by company size, data and decision. Incrementality tests (IncrementalityConversions caused by marketing.) are run on the channels where the model is least certain, and their results are fed back as calibration.
A model that moved budget between two channels
A consumer brand spends $12 million a year across television, paid social, paid search, radio and retail promotions, with weekly sales data for three years. The fitted model attributes 55 percent of sales to the baseline, 18 percent to promotions and 27 percent to media. Television's response curve shows it is well past its saturation point at $5 million a year, with a marginal return of $0.80 per dollar, while paid social at $1.5 million returns $2.10 at the margin. The optimiser recommends moving $1.2 million from television to social. A geo holdout on social confirms a lift consistent with the model's estimate, and the reallocation is made for the next year, with the model refreshed each quarter to check the curves have not shifted.
What it is not
Marketing mix modelling is not attribution (AttributionAssigning conversion credit to touchpoints.): it never sees an individual's path and cannot say which ad a person saw, only how much a channel's weight moved the aggregate outcome. It is not an experiment: without calibration its estimates are correlational, and channels whose spend always rises with demand, such as search or seasonal television, can be credited with sales they only accompanied. It is not fast or granular; it answers allocation questions by channel and month, not creative questions by week. And it is not objective by default: choices of adstock shape, saturation function, priors and controls change the answer, which is why the assumptions should be documented and the results validated against tests before budget moves on them.
Related terms
Attribution (AttributionAssigning conversion credit to touchpoints.) covers the touchpoint-level question the model cannot answer, and incrementality (IncrementalityConversions caused by marketing.) provides the experimental anchor the model needs to be trusted for large reallocations. Return on investment (ROIProfit relative to cost.) is the unit in which the model's channel outputs are usually expressed, at the average and at the margin, and the marginal figure is the one allocation decisions should use. Terms specific to the method, including adstock, saturation, response curves, baseline and calibration, are defined in the basics guide linked above; the model is also sometimes called econometric modelling, which is the same method under the name used in the research literature, and open-source implementations now make it available to teams that could not previously afford one.
Common mistakes
- Trusting a model with too short or too flat a history, where spend never varied enough to estimate effects, so the curves are guesses dressed as estimates.
- Leaving out controls such as price and distribution, so their effect is credited to media and the channel returns are overstated.
- Reading channel contributions as causal without calibration against experiments, especially for channels whose spend tracks demand, such as paid search.
- Using average return instead of marginal return for allocation decisions, which overfunds saturated channels and starves the ones with room to grow.
- Building the model once and never refreshing it as the market and the mix change, so the allocation it recommends drifts away from reality and nobody notices until results slip.
