Marketing mix modelling: how it is built, how to trust it, and when it fits

A marketing mix model (MMM, often called econometrics in the UK³) is a statistical model of aggregate data, usually weekly sales alongside spend by channel, price, distribution, seasonality and outside factors, that estimates how much each activity contributed and what would happen if the budget changed. Advertisers have used such models in various forms since the 1960s¹. The foundations piece measuring effectiveness explains where MMM sits among the other methods; this piece goes one level down.

• The decision it informs how to split next year's budget across channels and markets, and how much to spend in total.
• How it works it separates a base (what would have sold anyway) from the incremental sales each activity added, allowing for effects that linger and for diminishing returns².
• How to trust it test it on data it has not seen, check that its curves are plausible, and calibrate it with experiments^4, 12.
• Tools in 2026 Meta's Robyn and Google's Meridian are open source, and Google added a no-code planning layer to Meridian in February 2026^7, 10.
• Its limits it needs history and variation, it sees only what is in its data, and on its own it describes correlation¹.

MMM is the method for strategic, cross-channel questions: the total budget, the split between channels, the split between markets, and the shape of a channel's returns (where spending more stops paying). The IPA describes the integration of econometrics with experiments and attribution as the mark of a learning culture, not as a choice between them³. Its 2024 report Making Effectiveness Work, by Simeon Duckworth, Neil Charles and Duncan Stoddard, puts it plainly: no single method is a silver bullet, and each has a role and limits⁴.

In the triangulation used across this track, the mix model sets budgets and channel roles; experiments test its biggest assumptions; attribution optimises within channels day to day; and brand tracking shows the leading indicators the sales data cannot yet see.

Two features make an advertising model different from an ordinary regression. Carryover (often called adstock): an advertisement keeps working for weeks after it runs, so this week's sales depend partly on last month's spend. Shape, or saturation: each extra unit of spend in a channel adds less than the one before. Google researchers showed in 2017 how a Bayesian model can capture both lag effects and diminishing returns, which linear regression handles poorly, and how to compute return on ad spend (ROAS) and marginal ROAS from the results². The marginal figure is the one a budget decision needs: the average return tells you whether a channel paid back; the marginal return tells you whether the next unit will.

The same paper shows the catch. A Bayesian model starts from priors, assumptions about each channel's effect, and when the data are thin those priors weigh heavily on the answer and can bias it². That is a feature when the prior comes from a good experiment and a risk when it comes from wishful thinking. Write down every prior and where it came from.

Data is most of the work. The IAB's guidance on modernising MMM argues for standardised, multi-year, well-governed inputs, and for coverage that reaches every channel, new formats and under-represented audiences⁵. In integrated communication this matters doubly: if public relations, sponsorship, CRM messages or promotions are left out, the model will credit their effect to whatever moved at the same time. Record spend and a measure of exposure (impressions, ratings, sends) wherever you can, and keep a log of events the data cannot explain: a stock-out, a competitor's launch, a news story.

Variation is the raw material. A model learns a channel's effect from the weeks when that channel moved differently from the others. If every channel rises and falls together, no model can separate them; that is one of the core challenges Google researchers describe, along with data collection, model specification and the selection bias that targeting introduces¹. Planned variation (a market that goes dark, a burst at a different time) makes the next model better.

Google's handbook for marketing leaders offers a checklist of best practices for rigorous model design⁶, and the IPA's MESI loop (model, experiment, simulate, implement) treats the mix model as the place where hypotheses start, not where they end⁴. The strongest validation is ground truth: a randomised experiment, such as a geo test in which whole regions are assigned to see or not see a campaign¹¹. This matters because observational methods, even with rich data, often fail to reproduce what experiments measure¹². Use experiment results both as a check and, where the tool allows, as a calibration input.

Two free tools from the large ad platforms have changed the economics of MMM. As of October 2026:

• Robyn is published on CRAN as a semi-automated MMM package from Meta Marketing Science, under the MIT licence. Its description lists ridge regression, evolutionary algorithms, budget allocation outputs and optional ground-truth calibration; the latest version we could confirm was 3.12.1, published on 2 July 2025⁷.
• Meridian is Google's open-source, Bayesian MMM^8, 10. In February 2026 Google added Scenario Planner, a no-code interface intended to let marketing teams compare budget scenarios without writing code¹⁰. Its developer reference describes a module that builds Looker Studio dashboards from Meridian's outputs⁸, and, as last updated in March 2026, accepts model fit, marketing analysis and budget optimisation results⁹. The model still has to be trained on historical data first.

Free software lowers the licence cost, not the work. Whichever tool you use, the questions in the checklist above still apply, and someone on your side must be able to answer them.

It fits a business with meaningful spend across several channels, sales data by week (and ideally by region), and decisions big enough to repay the effort. Brands that sell through retailers or offline, where user-level tracking cannot follow the customer, often have no better option. The IAB frames MMM's return in exactly those terms: privacy limits on identifiers and fragmented media delivery have made aggregate methods relevant again⁵.

Its limits are structural. It is backward-looking, so a new channel without history is a guess until tested. It works at the level of channels and weeks, so it cannot tell which creative or keyword won (that is attribution's and testing's job). Long-term brand effects lie beyond its usual window unless the model is built to look for them; the brand-activation split and the brand tracker help to keep the long term in view. And without calibration, it describes association, not cause^1, 12.

Illustrative, with no real data. A regional food brand sells through supermarkets and spends on television, online video, out-of-home, paid search and in-store promotion. The finance director asks whether to move money from television to online video. The team gathers three years of weekly sales by region, spend and impressions by channel, prices, promotions and distribution, and adds the PR calendar. The first model credits paid search with an implausibly large share of sales. The team recognises the symptom: search demand rises because of television, so the two move together. It adds a geo test, holding television back in a few matched regions for six weeks, and uses the result to calibrate television. The recalibrated model shows television's returns flattening at the current weight while online video still has headroom. The recommendation is a modest shift, tested again next year, not a wholesale move.

1. Buying a model before framing the question The level of detail (national or regional, weekly or monthly) follows from the decision.
2. Leaving disciplines out PR, sponsorship, CRM and promotions missing from the data means their effects are credited elsewhere.
3. Reading average returns as marginal A channel can have a high average return and little room to grow.
4. Treating a new model as truth Compare it with the last version and with experiments before moving money.
5. Letting the seller grade its own channel A model built by a party with a stake in one channel needs independent checking.
6. Never refreshing Markets move; a model left alone for two years is history, not guidance.

On Monday:

1. Write the two or three budget decisions a model must inform this year, and who will make them; the measurement framework helps to tie them to objectives.
2. Inventory your data: weekly sales by region, spend and exposure for every discipline, price, distribution and an events log. Note the gaps.
3. Check variation: if channels always move together, plan a deliberate change this quarter.
4. Decide build or buy, and in either case ask for the validation checklist above in writing.
5. Plan one experiment to calibrate the channel the model will find hardest to separate.
6. Agree a refresh cadence and a rule: no budget move on a single version of the model.
7. Bring the results to one review with finance; see making the business case and budgeting for IMC.

• Google Research on the challenges of media mix modelling¹ and on carryover and shape effects².
• The IPA on econometrics³ and Making Effectiveness Work⁴.
• The IAB's guidance on modernising MMM⁵ and Google's CMO handbook⁶.
• Robyn on CRAN⁷; Meridian's Scenario Planner reference^8, 9 and launch coverage¹⁰.
• Vaver and Koehler on geo experiments¹¹; Gordon and colleagues on experiments against observational methods¹².
• On IMC Report: measuring effectiveness, the measurement objective, the CMO and marketing operations hubs, and the glossary entries for MMM and incrementality.