Problem
Marketing teams spend on many ad channels and need to know which ones bring users. The usual answer is multi-touch attribution (MTA): track users with cookies, pick a model such as last click or first click, and split the budget by it. This breaks down. Not everything is tracked, GDPR and the platforms restrict tracking more every year, every ad platform claims it brought the user, and offline channels such as billboards and TV do not fit at all.
Example from the article: Capital.com operates mostly in Europe, where cookies are blocked and user tracking is limited. It uses MMM as one more input when marketers doubt that some source works.
Solution in brief
Marketing mix modeling (MMM) skips user-level tracking. It regresses the KPI on spend per channel, transformed for delayed effect (adstock) and diminishing returns (saturation), plus holidays, seasonality and trend. The fitted model shows each channel's contribution and lets you search for a better budget split. In 2025 the tools are mostly Bayesian, led by PyMC-Marketing.
Questions this case answers
- How to estimate the effect of each marketing channel on sales when only aggregate spend and outcome series are available and user-level tracking is restricted
- How to allocate a spend budget across channels when attribution is unreliable and some channels are offline
- How to estimate the effect of each marketing channel on sales or installs without A/B tests or user-level tracking
- How to allocate a marketing budget across channels when attribution is unreliable or offline channels are involved
Tools and methods
- Marketing mix modeling (MMM) — regression of a KPI on channel spend and calendar factors.
- Adstock and saturation — delayed effect of spend and flattening response.
- PyMC-Marketing — the most flexible Bayesian MMM library.
- Robyn — frequentist, for a quick start, written in R.
- Meridian — Google's Bayesian library, runs on GPU.
- Time-varying parameters — a multiplier on the saturation curve.
Google or a one-time email code. No password.
