The economic thesis on the sister page only holds if the moment of decision-forming context can be accounted for. The trap is to recover accountability by re-introducing behavioral tracking — which would collapse the thesis. The way out is to measure the protocol's own artifacts, not the user: fetches, eligibility states, pre-render verifications, rendered-label proofs, and outcome windows that do not require identifying who the user was.
Measurement in this protocol does not mean reconstructing the user's path across sites, or attaching conversions to identifiers, or building any kind of profile-keyed report. It means making the protocol's own actions observable, at the points where actions occurred.
Specifically, measurement records what was fetched, what was verified, what was eligible, what was rendered — and whether outcomes happened, in time windows defined without per-user identification.
The unit being measured is the protocol event, not the user.
The intention-economy thesis (sister page) rests on a specific signal shape: decision-forming context that does not require knowing who the user is. If accountability for that signal requires knowing who the user was, the thesis fails — the system has to track the user to verify the bid, which puts behavioral tracking back into the architecture.
This is a structural failure mode, not a policy preference. A protocol that prices moments cannot be measured by tracking users without losing the property that made the moments priceable in the first place.
The accountability surface has to live on the protocol's own artifacts. Otherwise the protocol-level guarantees of /privacy-preserving-matching and /editorial-firewall are undone at the measurement layer.
Five categories of events, all readable from public artifacts and runtime observability:
| Event | What is recorded | Source |
|---|---|---|
| Manifest fetch | An agent fetched a sponsor's /.well-known/agent-ad.json at a timestamp | Origin logs (sponsor side) and crawler logs |
| Feed eligibility | A candidate ad was eligible in the broadcast feed at the moment of an agent's poll | Broadcast feed (/.well-known/abf.json + /.well-known/abf-index.json) state |
| Pre-render verification | The four-check verification flow returned pass/fail for a candidate, before render | Agent runtime's verification layer |
| Rendered-label proof | A specific ad was actually rendered, with the required label, on a specific surface | A render-time attestation, when the publisher or runtime publishes one |
| Outcome window | Counts of outcome events (clicks, completions) within a non-user-identifying time window | Aggregated runtime logs, bucketed without identifiers, using conventions documented in the integration guide |
Each of these is measurable without knowing who the user was. The events have timestamps, paths, candidate IDs, and aggregate counts — not user IDs.
The protocol's measurement surface deliberately excludes:
Measurement that requires any of these is not in this protocol's accountability surface.
A consumer of the measurement layer — a brand-safety auditor, an external reviewer, a sponsor's own analytics team — can verify the protocol's accountability by:
No step requires access to per-user data. Every step is verifiable from public artifacts or runtime-published aggregates.
OpenAI Ads is the category signal for the broader shift the sister page describes — an ad system that ranks against conversation and context, not against behavioral profiles. OpenAI's own measurement reporting on its platform is its own product surface.
Pre-render verification, manifest publication, and non-user-identifying outcome windows generalize the accountability shape: any agent runtime — not just OpenAI's — can be measured against the same observability surface, with the same artifacts in the same places. The measurement layer is portable in the way the protocol intends.
OpenAI Ads can operate alongside the protocol's measurement layer. Agent runtimes and sponsors that want a portable accountability surface benefit from the protocol-level definition.