Marketing mix modelling (MMM) estimates how marketing activity and other factors contribute to an aggregate business outcome, such as sales. It can help answer questions about channel contribution, diminishing returns and budget allocation. It does not recreate individual customer journeys or automatically tell an advertiser which person, interest group or creative asset to target.
Substantively reviewed and updated: 22 October 2026. This primer replaces the original article’s broad claims about a “cookieless” future and automatic optimisation with a clearer account of what the method can support.
What goes into an MMM?
A model typically combines an outcome measured over time with channel-level media activity, spending and relevant non-media information. Depending on the business and implementation, those inputs may be available by geographic market as well as time period. Prices, promotions, distribution, seasonality and demand conditions can matter because advertising is rarely the only thing changing.
The model estimates relationships under a set of assumptions. Many implementations represent delayed advertising effects and diminishing returns as activity increases. Some use Bayesian priors to incorporate existing information and express uncertainty. Google’s Meridian documentation provides one concrete example, describing channel contribution, response curves and future budget allocation as separate business questions.
What it can help measure
Historical contribution: How much of an observed outcome is estimated to be associated with a channel’s incremental effect, conditional on the model and its causal assumptions? This is different from counting every conversion a platform attributes to itself.
Response to spending: How might the outcome change as activity changes? A response curve can help distinguish the average return on past spending from the expected return on an additional increment. Those quantities can lead to different budget decisions.
Budget scenarios: Given the estimated curves and practical constraints, what allocation is projected to improve the chosen outcome? The projection remains conditional on the model and the future resembling its assumptions closely enough. It is not a guarantee of maximum return.
What it cannot establish automatically
A model cannot reliably distinguish effects that the data do not identify. If television and online video spending always change together, a precise split of their contributions may depend more on assumptions than on independent evidence. Adding more detailed creative or audience columns can make that problem worse when the observations do not contain enough variation.
MMM can sometimes analyse tactic-level differences when the data and design support them. It does not inherently recover person-level interests, select individual ad recipients or prove the best creative treatment. Controlled creative or audience experiments may answer those questions more directly.
Nor does ordinary regression automatically establish causality. Advertising budgets may increase when demand is already expected to rise; failing to account for that can overstate effectiveness. Controls, assumptions and calibration against relevant experiments matter. Accurate prediction is useful, but is a different achievement from identifying a causal effect.
A hypothetical allocation decision
Imagine a retailer comparing two channels. Channel A has the higher historical average return, but the model suggests it is near saturation at its current spend. Channel B has a lower historical average return but a higher estimated marginal return over a modest additional spend range.
That result could justify testing a small reallocation toward B. Before acting, inspect uncertainty, check the channels’ data and consider whether the proposed spend lies within a range the model has observed. Capacity, stock availability and creative changes may also affect the outcome. The example illustrates the distinction between average and marginal return; it does not establish which channel any real retailer should prefer.
Where tracking restrictions fit
MMM can be useful without assembling cross-site, person-level journeys, which makes it relevant when those observations are unavailable or incomplete. That does not mean all cookies have disappeared, or that MMM replaces every attribution or targeting function.
First-party and third-party cookies serve different purposes, and browser policies differ. In its 22 April 2025 announcement, Google said Chrome would maintain its then-current approach to third-party-cookie choice. That historical decision is a reason to avoid treating a universal cookie shutdown as an established premise; it is not a statement of all current browser policies.
When I would use it
I would consider MMM when the decision concerns aggregate channel budgets, the business has consistent outcome and media data, and there is enough relevant variation to support the intended level of detail. I would be cautious about demanding a fine-grained answer from a short or inconsistent history, or using an explanatory-looking chart as proof of causality.
For a practical workflow covering data checks, model challenges and bounded budget tests, see planning an MMM project. For experiments and an explicit financial return calculation, see measuring social media ROI. These methods are most useful when their roles and limitations are stated clearly.