Last-touch is not a philosophy. It is what happens when a reporting shortcut needs a single winner before Friday. In Product Funnel Analytics rooms we see the same pattern: a mix model exists in a slide appendix, while the live dashboard still credits the last paid click that touched a person who was already going to convert.
The leak is rarely the model. It is the join. If your stage contract allows recoding from “trial” to “paid” on the same day as a branded search click, last-touch looks brilliant. Tighten the recode rule — paid requires a completed charge older than the click by a cooling window — and the channel table redistributes. Marketing teams in GB often experience this as an attack. It is not. It is a denominator becoming visible.
In Reading the Drop, module 3 asks you to publish two tables: the live dashboard, and a mix with ranges. The assessed part is not picking a model. It is writing the sentence that explains what would move a channel out of the “we would still fund this” band. Uncertainty is the point. A point estimate that always matches last-touch is a smell.
If you cannot export assist paths, say so in the memo. Pretending you have a data-driven model while the warehouse only stores last click is how leaky attribution survives another quarter. Honesty is cheaper than a new vendor.