Multi-touch attribution: get definitions aligned before you chase the algorithm

Attribution arguments often sound technical on the surface, but underneath they are usually disagreements about language. If two teams use the same word “conversion” for different events, no model—rules-based or ML—will produce peace. Alignment work is unglamorous; it is also the main thing that separates dashboards people trust from dashboards people ignore.

A lightweight spec beats an implicit oral tradition: which touchpoints count, how lookback windows work, how cross-device identity is handled, and what happens when data arrives late. When those choices are explicit, disagreements move from “your SQL is wrong” to “we should change rule 3 for channel X,” which is a manageable conversation.

A healthy attribution system highlights uncertainty instead of hiding it. Trending unattributed volume, reconciliation checks against finance or source-of-truth systems, and simple anomaly alerts turn attribution into something you can improve over time—rather than a fragile score nobody wants to defend in a planning meeting.