Problem
Marketing teams measure performance channels (links, contextual ads, banners) by last click: purchases from that traffic divided by its budget. Everyone else counts as "organic" or "direct". TV, radio and outdoor get the largest budgets, yet nobody can say what they return, and performance channels get credit for people who first heard of the brand elsewhere. "Where did you hear about us?" surveys do not fix this: many people skip them, and nobody can say which of several touches mattered.
Example from the article: a person who arrives through a performance link most likely learned about MTS from TV, radio or street banners, yet the link gets all the credit.
Solution in brief
Fit a linear model on daily or weekly aggregates, where each channel enters through adstock (delayed effect) and diminishing-returns transforms. Sample many model variants (over 280,000 in Mosaica), keep the best by and MAPE, then keep only models that agree with business hypotheses about channels. The surviving model's coefficients split the KPI by channel and give ROI per channel.
Questions this case answers
- How to estimate the sales effect of offline channels from aggregated time series when user-level attribution is impossible
- How to select one defensible linear attribution model out of many candidate transforms of delayed and saturating effects
- How to estimate the sales effect of offline channels such as TV, radio and outdoor from aggregated time series
- How to get one defensible marketing mix model out of hundreds of thousands of candidates
Tools and methods
- Marketing mix modeling (MMM) — attributes a KPI to channels
- Linear regression with direct and cross-effects — keeps contributions additive and explainable
- Adstock and diminishing returns — transforms for delayed and saturating ad effects
- Random model sampling — searches transforms and lags
- Business calibration — filters models by hypotheses
- Robyn, LightweightMMM — existing frameworks the team compared against
Google or a one-time email code. No password.