<b>Should you trust an attribution model you can't inspect?</b>

<b>Should you trust an attribution model you can't inspect?</b>

The question: 'data-driven attribution' (the default in most ad platforms now) promises to learn credit from your actual conversion patterns rather than imposing a rule. But the model is usually proprietary and opaque. How do you audit a black box that's grading the homework of the same channels its owner sells you?

What's under the hood: most platform DDA is some flavor of a probabilistic model — frequently a variant of the Shapley-value or a logistic path model — comparing converting and non-converting paths to estimate each touch's contribution. That's a legitimate methodology. The problem isn't the math; it's the incentive and the boundary. The platform only sees <i>its own</i> touches. Its DDA cannot credit a channel it doesn't observe, which structurally inflates the platforms with the most coverage.

What the research warns: any attribution confined to a single walled garden is a within-platform credit split, not a cross-channel truth. Independent analyses repeatedly find that summing the 'conversions' claimed by each platform's own model exceeds actual conversions — sometimes by 2x or more — because each garden claims overlapping credit.

The nuance: DDA is still better than last-click <i>within</i> its domain. The failure is treating an in-platform number as an account-wide one.

What to actually do: never sum platform-reported conversions. Pull touch-level data into a neutral environment and run your own model, or anchor everything to a holdout. Audit the boundary, not just the algorithm.

<b>Bottom line for practitioners:</b> the danger of data-driven attribution isn't that it's a black box — it's that the box only contains the seller's inventory.