Data-driven attribution vs. rule-based: when the black box actually beats the dropdown

Data-driven attribution vs. rule-based: when the black box actually beats the dropdown

Google and Adobe push "data-driven attribution" (DDA) as the upgrade from last-click. Is the upgrade real, or repackaging?

What DDA actually does
Data-driven attribution fits a model — typically a Shapley-style or logistic approach — to your own conversion paths, learning each touchpoint's contribution from converters-vs-non-converters in your data. Unlike a fixed rule, the weights are estimated, and they shift as your mix changes.

Where it genuinely wins
DDA escapes the hand-coded bias of last-click and can surface assist channels that rule-based models flatten. With enough conversions (Google historically required hundreds per month), the weights are more stable than a coin-flip heuristic.

Where the upgrade is oversold
DDA still learns from observational path data, so it captures correlation, not causation. If branded search always appears right before conversion, DDA will credit it heavily — exactly the last-click failure mode, now laundered through a model you cannot inspect. It also can't see touches outside the platform's tracking, so a Google DDA model is blind to everything non-Google.

The diagnostic question
Ask: would this model's credit survive a holdout test? If branded search gets 30% credit but a search-pause experiment shows near-zero incremental lift, the model is confidently wrong.

Bottom line for practitioners: DDA beats rule-based models for in-platform optimization and is worth switching to. But it is not incrementality — it redistributes observed correlations more cleverly. Calibrate its credits against experiments before reallocating real budget, and never let a single-platform DDA arbitrate cross-channel decisions.