Attribution model

A customer touched four things before buying. How much did each contribute? Nothing in the data answers that, so a model decides.

The common ones and what each rewards. Last click gives everything to the final touchpoint and rewards whatever sits closest to the purchase. First click gives everything to the first and rewards discovery. Linear splits credit evenly and rewards being present. Time decay weights recent touchpoints more heavily. Position-based loads the first and last and thins the middle. Data-driven models fit weights from observed patterns rather than a fixed rule, which sounds like an escape from the choice and is really a different one — the training data and the objective still encode a view.

The thing worth internalising: none of these is more correct than the others, because there is no observable fact they are approximating. A conversion is one event with several things preceding it. Dividing it into fractions is an accounting convention, and switching models does not reveal a truer number — it applies a different convention and produces a different number from the same events.

That is not an argument for indifference. It is an argument for picking a model deliberately, saying which one, and not comparing figures across models — which happens constantly and silently, because different tools default differently and nobody prints the model on the report.

Where it stops. Every model here divides credit among touchpoints the system observed. None of them asks whether the conversion would have happened anyway — that is incrementality, and no attribution model can answer it.