Marketing Mix Modeling: What It Is, What Data It Needs, and What a Small Advertiser Can Realistically Expect

Marketing mix modeling (MMM) is a statistical method that estimates how much each marketing channel, and each non-marketing factor, contributed to an outcome such as sales or leads. Google's Meridian project describes it as measuring the impact of marketing campaigns and activities to guide budget planning, using aggregated data across channels while accounting for factors outside marketing.

The model works on totals, not individuals. You give it, for example, weekly spend by channel along with weekly sales, plus things like price changes, promotions and seasonality, and it estimates the relationship between them. Because no cookie or user-level record is needed, the approach is not affected by consent banners, ad blockers or browser tracking limits.

The name echoes the older "marketing mix" idea of balancing product, price, place and promotion, and the modern use is narrower: estimating the contribution of each paid and unpaid driver to a result. The output is usually a set of response curves showing how the outcome changes as spend on a channel rises, which is what lets a model recommend moving budget from one place to another.

At minimum, an outcome series (revenue, orders or qualified leads), spend or activity by channel over the same periods, and the other forces that move the outcome: promotions, pricing, seasonality, holidays, stock problems and anything unusual such as a competitor's launch. The data has to be consistent: the same definition of the outcome and the same way of counting spend across the whole history.

The model also needs variation. If every channel is funded at a steady level for two years, the data cannot show what happens when one goes up or down. Geographic data helps where available; the Meridian documentation says geo-level data is encouraged but national-level modelling is also supported. How many periods are enough depends on the model, the number of channels and the noise, so read the documentation of the tool you intend to use rather than relying on a rule of thumb from an article.

Data preparation is most of the work. Spend sits in ad platforms, sales in finance or ecommerce systems, promotions in a spreadsheet someone keeps, and each has its own calendar and definitions. Aligning them to a common period, correcting for late-reported conversions and documenting each source is dull and decisive, because the model will faithfully learn from errors it cannot see.

Attribution follows individual journeys through clicks and tags, so it breaks when journeys cannot be followed: consented-out visitors, cross-device behaviour, offline sales, TV, radio, outdoor and the effect of a channel that influences people who later search for you by name. MMM sees those effects in aggregate, because it does not need to follow anyone.

It does not replace attribution. Attribution is fast, granular and good for day-to-day campaign decisions; MMM is slower and answers strategic budget questions. The two disagree sometimes, and the disagreement is informative. Experiments, such as holding out a region from a channel, give a third view and can be used to calibrate the model. Robyn's documentation says it calibrates models against ground-truth methods such as geo-based and lift experiments, and Meridian provides a way to feed experiment results in as prior information.

Another sign it is time for MMM is when leadership asks questions that click data cannot answer, such as what happens to sales if the brand budget is cut by a fifth. Attribution can describe what was clicked. A model, with all its uncertainty, can at least attempt a scenario, and the stated range makes the uncertainty visible rather than hiding it.

A model reports correlation structured by assumptions. Channels that spend rises and falls together are hard to separate, and spending often rises during busy seasons, which can make a channel look more effective than it was. The output is a set of estimates with ranges, and different reasonable model setups can give different answers. Treat a single headline ROI figure with suspicion.

MMM is also slow and coarse. It will not tell you which headline performed better, and it looks backwards, so a channel you started last month has almost no history to learn from. Models need refreshing as the business changes. And the work needs someone who can judge whether the results make sense; a model that says one channel drives most of your sales is as likely to be a modelling artefact as a discovery.

A model cannot measure what it was not given. If a major driver, such as a distributor change or a supply shortage, is missing from the data, its effect will be credited to whichever channel happened to move at the same time. That is why a person who knows the business has to review the inputs and the outputs, and why unexplained results should be treated as questions to investigate, not as findings.

Two open-source projects are widely referenced. Google's Meridian is described by its project page as an open-source MMM framework based on Bayesian causal inference that lets advertisers run models in-house. It runs on Python, and the project recommends a GPU because the sampling method is compute-intensive. Meta's Robyn uses ridge regression to handle correlated variables, automated hyperparameter search and a budget allocator, and states that it needs no personal data or cookie data.

Both are tools, not answers. They are free to use, but you still need data preparation, modelling judgement and the time to validate results. Running either well is analyst work, which is why many advertisers use them through a specialist rather than alone.

Before choosing, read each project's current documentation and issue tracker. Open-source tools change quickly, requirements shift, and the maintainers' notes about known limitations are often more candid than any article, including this one. Check the licence and the supported versions against what your team can actually run.

A small advertiser with a few channels, modest spend and short history will often get a wide, unconvincing answer from a full model, because there is little variation to learn from and noise swamps the signal. Build-and-maintain effort can also exceed the value of the decisions it informs. Below a certain spend, the budget is better managed with simpler tools.

We recommend starting with cheaper evidence: a clean record of spend and outcomes by month, deliberate changes in spend by region or period that create variation, a holdout test on your largest channel, and conversion tracking that is honest about what it misses. If those habits are already in place and the budget is large enough to justify analyst time, a model becomes a sensible next step.

If you do commission a model, ask what decision it will change. A useful answer is concrete, such as whether to move a fixed share of budget between two channels for a quarter. Plan to test the recommendation with a limited change rather than moving everything at once, and read the results as one input alongside what you see in your own sales data.