Heuristic
Last-click
100% of credit to the final touch. Still how most budget meetings start.
METHOD
Six models in parallel. A rubric that cannot change the numbers.
Heuristics bound the story. Markov and Shapley argue from paths. An evaluator ranks them. A human still has to press send.
Heuristic
First-click
100% to the opener. Useful as a bound, not a plan.
Heuristic
Linear
Equal share across every touch in the converting path.
Heuristic
Time-decay
Later touches get more credit. Half-life is a tenant setting.
Path model
Markov removal
Removal effect: how many conversions vanish if a channel is taken out of the path.
Path model
Shapley values
Coalition credit with a 90% interval. Monte Carlo by default; exact if ≤ 8 channels.
Scoring rubric
Numeric first, language model second. The model narrative may only reference numbers present in the payload.
Coherence
Shares sum to 1.0. Removal effects are non-negative. No channel is credited more conversions than exist.
Stability
Week-over-week rank correlation. A channel whose share moves more than the threshold is flagged, not silently trusted.
Actionability
Does the ranking change a budget decision a Head of Growth could take this week? If not, it does not lead the PDF.
Known bias
Consent-mode undercount, Direct inflation, view-through gaps. Written into the appendix every week.
What we refuse to claim
- We will not invent shares. The LLM drafts commentary from the payload; a verifier rejects unseen figures.
- We will not sell MMM until there are 18–24 months of clean weekly spend. Path credit and incrementality are different numbers.
- We will not auto-publish GTM containers, bid, or pause campaigns.
- We will not claim a model is ground truth. The PDF is a second opinion with uncertainty.
- We will not run path models off the GA4 Data API. Journeys come from BigQuery or they do not come.