What is biased attribution?

Where the bias comes from

It is rarely dishonesty and almost always structure. A platform that measures its own performance faces a series of judgement calls — how long the window is, whether an unclicked impression counts, how a partial match is resolved — and each one has a direction that flatters it. Making the defensible choice every time still produces a systematic tilt.

The clearest symptom is arithmetic. Add up the conversions each platform claims and the total exceeds the conversions the business actually recorded, sometimes by a factor of two. Every platform touched some of the same users and each counted them fully.

The self-attributing network problem

A self-attributing network reports on its own inventory and does not expose the underlying interaction data for independent matching. You receive a number and cannot audit how it was produced.

This is why mobile measurement partners exist: a neutral third party that sees interactions from every source and applies one consistent rule. It removes the incentive problem, though not the underlying uncertainty — an MMP still has to pick a model, and still cannot see unclicked or cross-device influence.

How to correct for it

• Reconcile against your own system of record. Your order database is the denominator; platform claims are hypotheses about it.
• Compute a de-duplication factor. Total claimed divided by actual, tracked over time — a stable ratio is manageable, a drifting one is a signal.
• Hold out. Incrementality testing is immune to the incentive, because it compares exposed against unexposed rather than asking a platform what it did.
• Verify window and model parity before any cross-platform comparison.

Bias on AI surfaces

The risk profile inverts here, and it is worth being clear about which direction. There is currently no independent counting layer for most AI inventory, so the platform's number is the only number — the classic bias condition.

But the dominant error today is under-counting rather than over-claiming, because so much influence produces no click at all. A buyer applying the usual scepticism to a self-reported AI number, and no holdout, will conclude the channel does nothing. The correction is the same instrument in both cases: measure the causal effect yourself.