{"id":2737,"date":"2026-08-20T12:41:48","date_gmt":"2026-08-20T10:41:48","guid":{"rendered":"https:\/\/mairateam.com\/?post_type=glossary&p=2737"},"modified":"2026-08-20T15:55:54","modified_gmt":"2026-08-20T13:55:54","slug":"marketing-mix-modeling","status":"publish","type":"glossary","link":"https:\/\/mairateam.com\/en\/glossary\/marketing-mix-modeling\/","title":{"rendered":"Marketing mix modeling"},"content":{"rendered":"

Marketing mix modeling<\/strong>, MMM for short, is a statistical model that estimates from historical time series how individual channels, seasonality, pricing and other factors contribute to revenue. It works with no cookies and no individual tracking. It only needs aggregated data: spend per channel, revenue, seasonality. That is exactly why it is returning as the main tool for strategic budget allocation now that cookie-based measurement has decayed.<\/p>\n

In short<\/h2>\n\n\n\n\n\n\n
What it is<\/strong><\/td>\nA statistical estimate of channel contribution to revenue from historical data<\/td>\n<\/tr>\n
What it needs<\/strong><\/td>\nAggregated time series: spend, revenue, seasonality, promotions, ideally 2 to 3 years<\/td>\n<\/tr>\n
What it does not need<\/strong><\/td>\nCookies, pixels, consent, individual tracking<\/td>\n<\/tr>\n
What it is for<\/strong><\/td>\nStrategic budget split between channels, not day-to-day optimization<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

How it works<\/h2>\n

Inputs and the model equation<\/h3>\n

The base is a regression on time series. The variable being explained is revenue (or orders) per week; the explanatory variables are spend or impressions of every channel plus control variables: seasonality, holidays, promotions, price changes, weather, competitor activity, possibly macroeconomic indicators. The model searches for the coefficients with which the sum of contributions best explains actual revenue. The part of revenue no channel explains is the baseline: organic demand, brand strength, repeat purchases. Simplified: revenue in week t = baseline + contribution of channel 1 + contribution of channel 2 + ... + effect of control variables + noise.<\/p>\n

Adstock and saturation<\/h3>\n

Two transformations set MMM apart from plain regression. Adstock<\/strong> captures carryover: advertising from one week keeps working in the following weeks, so spend is spread over time before it enters the model. Worked example: with a carryover rate of 0.5, a spend of \u20ac10,000 in week one counts as \u20ac10,000, carries \u20ac5,000 into week two, \u20ac2,500 into week three and keeps declining. Channels with long carryover (TV, video, brand) tend to have a higher rate than search. Saturation<\/strong> captures diminishing returns: the response curve (typically an S-curve or a Hill function) says the first euros in a channel bring more than the last ones. The model therefore estimates not one average ROAS per channel but the whole curve, from which the marginal return is read: what the next euro brings at the current spend level.<\/p>\n

Calibration and outputs<\/h3>\n

Regression alone cannot tell cause from correlation: channels whose spend rises together with the season cannot be told apart by the model. That is why the model is calibrated: the coefficients get prior bounds (priors) from geo lift tests and holdout tests, for example that a channel's incremental ROAS came out between 1.5 and 2.5. The model is further checked on held-out weeks (out of sample), where it must not know the result. The outputs are a decomposition of revenue by channel and baseline, a response curve for every channel, marginal ROAS and, derived from it, a proposed reallocation of budget at the same total spend or for a target revenue volume. The estimates come with an uncertainty range, and that range belongs in every presentation of results; otherwise the estimate turns into false precision.<\/p>\n

When MMM makes sense and when not<\/h2>\n

MMM answers the question pixel attribution cannot: what would happen to revenue if budget moved between channels, including the ones no pixel sees, such as TV, radio or brand campaigns. The price is coarseness: the model works in weeks and channels, not campaigns and ad sets. It also needs enough budget and data variability: if spend never changed, the model has nothing to estimate from. For smaller accounts with one or two channels MMM is needlessly heavy machinery and incrementality tests serve better.<\/p>\n

From our own practice<\/h2>\n

To measure the real contribution of channels we use our own platform DiagnostIQ, which rests on the same principle as MMM: incremental impact instead of conversion crediting. A concrete outcome: an incrementality test of TikTok for an e-commerce client showed 15 percent additional revenue that would not have existed without the channel, even though pixel attribution undervalued it. That type of contradiction is exactly why strategic budgets should not be steered by attribution alone.<\/p>\n

Common mistakes<\/h2>\n