How marketing mix models work

Explore adstock, saturation, baselines and confounding with visual explanations and interactive model mechanics.

A marketing mix model combines a baseline with transformed media inputs and other drivers. Each transformation expresses a claim: advertising can carry over, additional spend may have diminishing returns, and sales would still exist without the measured campaigns.

Begin with Adstock and Saturation curves. Then connect the analysis to the MMM course exercises, where you practise making a decision from model outputs.

Before acting on a recommendation, connect this topic to budget decisions: the result needs both credible measurement and a commercially worthwhile action.

A practical route through the topic

1 Raw weekly spend
2 Carryover and saturation
3 Controls and baseline
4 Contribution and uncertainty

With geometric adstock, a one-off spend of 100 and a decay of 0.5 gives an exposure sequence of 100, 50, 25 and 12.5. That is transformed exposure, not extra money spent. A saturation curve then maps exposure to a diminishing response. Changing the order or normalisation changes the interpretation.

Try it before reading further

Change the inputs below and explain what changed. These teaching tools illustrate assumptions; they do not estimate your business’s return from the information on this page.

Weekly retention 50%

Reporting window 4 weeks

Half-life

1.0 wk

Counted inside window

94%

Arrives after report

What to check before using the result

Flexible curves can explain historical data in several different ways. Check whether decay, saturation and channel returns change when you alter priors, controls or the training window. A parameter at its bound deserves investigation.

Transformation units and ordering documented
Controls chosen using a causal argument, not correlation alone
Residual patterns and time holdouts inspected
Channel estimates checked for sensitivity to plausible assumptions

Your first reading path

For an overview, visit marketing mix modelling training. If you use AI to assist with analysis, the prompt engineering learning page explains the complementary skill.

Step 1 Adstock What adstock is, the geometric adstock formula (two conventions), half-life, delayed and Weibull adstock, how to estimate decay, plus a free Excel calculator. Step 2 Saturation curves What a saturation curve is, how the Hill function works (half-saturation and shape), logistic and exponential alternatives, and marginal vs average return. Step 3 Base sales What base sales are, how baseline vs incremental sales are split in marketing mix modeling and retail promotion analysis, and why base is not organic. Step 4 Confounding variables Confounding variables in marketing mix models: omitted variable bias, endogeneity, and how to choose controls by causal role, with an interactive causal graph. Step 5 Bayesian MMM Bayesian MMM explained: priors, likelihood and posterior for channel ROI, MCMC, credible intervals, prior predictive checks, experiment priors and geo models.

Foundations from Mike’s learning briefs

Homoscedasticity and Heteroscedasticity Residual variance is one diagnostic; inspect the regression assumptions before interpreting coefficient uncertainty. Response Curves Response curves translate exposure into predicted outcomes; their slopes drive marginal budget decisions. Diminishing Returns Diminishing marginal response explains why a channel with strong average returns may not deserve the next pound. Seasonality Recurring demand and one-off events should be separated before assigning their sales movements to media. Mean Absolute Percentage Error Percentage error can become unstable near zero, and prediction error alone cannot validate a channel decomposition. Bias-Variance Tradeoff Regularisation changes the bias-variance balance; compare the resulting predictions on future periods. Linear Regression Regression estimates conditional relationships; diagnostics and a credible causal design determine their interpretation. The Curse of Dimensionality More predictors need more independent information; regularisation can stabilise estimates but cannot manufacture evidence. Kitchen Sink Model Adding every available predictor can introduce bias or unstable effects rather than remove uncertainty. Model Accuracy A holdout checks prediction on unseen periods; the chosen error metric still needs to match the outcome scale. Multicollinearity Channels that move together can predict sales while remaining difficult to separate causally. Bayesian MCMC Sampling a posterior is useful only when you also challenge convergence and predictive behaviour. Partial Dependence A response plot should distinguish a model sensitivity calculation from an experimentally established effect. The Definitive Guide to the Log Log Model A log-log coefficient expresses elasticity, which differs from the percentage interpretation of a log-linear model. The Definitive Guide to the Log Linear Model The choice of which variable to log changes the coefficient interpretation and the business calculation. Dummy Variables Indicator variables can represent events, but including an event still requires a causal argument.

Turn understanding into practice

Your next step: make a better marketing decision.

Work through realistic business problems with guided lessons, practical templates and simulations.

Choose your learning plan Explore the MMM course See the curriculum before choosing a plan.

Explore every guide in this topic

Confounding variables Confounding variables in marketing mix models: omitted variable bias, endogeneity, and how to choose controls by causal role, with an interactive causal graph. Multicollinearity Multicollinearity in marketing mix models: why correlated channels give a great fit but unstable ROIs, how to check VIF and condition numbers, and the fixes. Ridge regression Ridge regression for marketing mix models: the objective, why standardization matters, how Robyn picks lambda, the bias-variance trade-off, and its limits. Adstock What adstock is, the geometric adstock formula (two conventions), half-life, delayed and Weibull adstock, how to estimate decay, plus a free Excel calculator. Saturation curves What a saturation curve is, how the Hill function works (half-saturation and shape), logistic and exponential alternatives, and marginal vs average return. Reach and frequency What reach and frequency mean, where effective frequency and the three-exposure rule came from, reach curves, frequency caps, and how Meridian models R&F. Base sales What base sales are, how baseline vs incremental sales are split in marketing mix modeling and retail promotion analysis, and why base is not organic. Bayesian MMM Bayesian MMM explained: priors, likelihood and posterior for channel ROI, MCMC, credible intervals, prior predictive checks, experiment priors and geo models. Model validation How to validate a marketing mix model: in-sample vs holdout error, rolling-origin cross-validation, MAPE, NRMSE, DECOMP.RSSD, R², stability and experiments. Model diagnostics MMM model diagnostics in practice: residuals over time, autocorrelation, Durbin-Watson, heteroskedasticity, structural breaks and outliers, and what each means. MMM books The best marketing mix modeling books, annotated: market response models, how advertising works, Bayesian statistics and causal inference, with a path by role. MMM research papers The marketing mix modeling papers worth reading, from Broadbent's adstock to Google's Bayesian MMM series and 2026 work, each summarized with a verified link. Jin et al. (2017) Jin et al. (2017), Google's Bayesian methods for media mix modeling with carryover and shape effects: delayed adstock, Hill curves, ROAS, priors and data needs. Chan & Perry (2017) Chan and Perry's 2017 Google paper on the challenges and opportunities in media mix modeling: thin data, selection bias, model uncertainty and the fixes. Geo-level hierarchical MMM Sun et al.'s 2017 paper on geo-level Bayesian hierarchical media mix modeling explained: why regional data helps, how partial pooling works and what it found. Lewis & Rao (2015) Lewis and Rao's 2015 QJE paper on 25 ad experiments shows why ROI is so hard to measure: sales noise swamps ad effects. Calculate the sample size you'd need. Facebook RCTs vs observational Gordon, Zettelmeyer and colleagues compared Facebook lift experiments with observational methods. Why matching and ML overstate ad lift, and what to do. The Long and the Short of It The Long and the Short of It by Binet and Field (IPA, 2013): brand vs activation, the 60:40 rule and its refinements, excess share of voice and what MMM misses.

Apply it to your marketing

Adstock half-life: convert decay into a time assumption Translate geometric decay into days or weeks, compare equivalent time scales and avoid interpreting normalised adstock as extra advertising spend. MMM channel grouping: when separate ROAS is unsupported Decide whether correlated channels should be grouped, tested or reported with wider uncertainty instead of forcing a precise channel-by-channel split. Paid search MMM: separate demand capture from demand creation Model paid search without assigning all existing demand to ads. Work through branded search, auction changes and controls that advertising may itself affect. Paid social MMM: distinguish saturation from creative fatigue Diagnose whether weaker paid social returns reflect spend saturation, creative changes, auction costs or audience composition before changing the budget. MMM data leakage: audit the path from raw data to holdout Find future information hidden in scaling, imputation, revised outcomes and model selection before interpreting marketing mix model validation results. MMM saturation curves: know when you are extrapolating Compare response curves inside and outside observed spend, and constrain budget recommendations when the data does not identify the high-spend response.

Frequently asked questions

What is the difference between adstock and saturation?

Adstock describes how an advertising effect persists over time. Saturation describes how the response changes with exposure or spend, often with diminishing marginal returns. They answer different questions.

Why can two models fit equally well but recommend different budgets?

Correlated inputs and flexible transformations can support several decompositions of the same sales history. Experiments, constraints and sensitivity checks help distinguish plausible decisions, but may not remove all uncertainty.

Connect the ideas

Marketing mix modelling Choose a measurement question, audit your data and learn how an MMM turns channel activity into a budget decision. Incrementality and experiments Design holdouts and geo tests, understand statistical power and use experiments to challenge attribution and calibrate MMM. Marketing budget decisions Compare marginal returns, contribution margins and bounded scenarios before moving marketing budget. MMM tools, code and platforms Choose between R, Python, spreadsheets and managed MMM platforms, with tutorials, datasets and validation checkpoints.