What is attribution in advertising?

What attribution actually claims

The question attribution answers sounds factual — which ad caused this sale? — and is not. Outside a controlled experiment, nobody observes causation. What systems observe is a sequence: this user saw these things, then converted. Attribution is the rule you apply to that sequence to divide credit.

That is worth stating plainly because almost every argument about attribution is really an argument about which rule to use, conducted as though it were an argument about facts. Last click is a rule. Multi-touch is a rule. Neither is the truth; they are different opinions about the same log.

How attribution is performed

1. An interaction is recorded — a click, an impression, an agent call — with a campaign identifier attached, usually via a tracker or tracking parameter.
2. A conversion occurs and is reported to the measuring system, often by postback.
3. The two are matched, within an attribution window that bounds how long a prior interaction stays eligible.
4. Credit is assigned according to the attribution model — all to one touchpoint, or divided across several.

Why it degraded, and what replaced it

The mechanics above depend on recognising the same user at both ends. Identifier deprecation removed that on mobile and in browsers, so the deterministic match became partial. Apple's SKAdNetwork and AdAttributionKit replaced user-level matching with aggregated, delayed, deliberately coarse reporting.

The response has been to stop relying on any single instrument. Serious measurement now triangulates: platform-reported attribution for operational decisions, incrementality testing for whether spend causes anything, and media mix modelling for how budget should be split at the top. Each is weak alone and the three disagree usefully.

What breaks on AI surfaces

AI assistants remove the remaining assumption: that the user personally traverses the path.

A user can act on a recommendation without clicking it — reading a brand name and searching for it later, or in a different session on a different device. There is no interaction to record, so the conversion arrives unattributed and lands in direct or organic. This is not a tracking gap that better plumbing closes; the click genuinely did not happen.

Where an agent completes the purchase, there is no session at all — see agentic conversion — and credit has to be traced through the agent chain instead, which is what A2A attribution attempts. The practical consequence for a buyer today is that last-click will systematically undercount AI media, and a holdout is the only instrument that reads it correctly. We have written this up in measuring AI assistant ad performance honestly.