Why do two attribution models built on the same data disagree by double digits?

Why do two attribution models built on the same data disagree by double digits?

When a team switched its reporting from last-click to a Shapley-value model, paid social's credit jumped 40% overnight. Same conversions, same paths. The question: which number is right, and what does the gap mean?

The method. Shapley value comes from cooperative game theory: it asks what each channel contributes by averaging its marginal effect across every possible ordering of the channels in a journey. A channel's credit is its average added value over all the coalitions it could join — a mathematically fair division, which is why Google's data-driven attribution and many platforms adopted variants of it.

What the comparison showed. In a documented Google Analytics 4 data-driven-attribution rollout for a multichannel retailer, moving off last-click reduced credit to brand search and direct by 30-45% and raised credit to early paid social and display by a similar magnitude. Conversions counted didn't change — only their distribution across channels did.

The nuance. Shapley is fair, but fairness is not causation. It distributes observed credit elegantly; it still assumes every touch in the path mattered. If a journey would have converted without paid social, Shapley will still hand social a share. The 40% jump is a reallocation artifact, not proof social drove the sale.

Bottom line for practitioners: Shapley and last-click can't both be "correct" because neither measures causation — they measure different credit rules over the same correlational data. Pick the rule that matches your decision, then anchor at least the channels it inflates to an incrementality test.