Specification

The session-event-outcome model for content attribution in AI agent interactions. Defines how retrieval, citation, engagement, and conversion events are structured, transported, and correlated across multiple observers.

Core event types

Commerce extension. checkout_completed - purchase confirmed, ties an outcome back to content - is not part of the seven core event types. It is defined in the ACP commerce extension, alongside the content_attribution object embedded in Agentic Commerce Protocol checkouts. Session outcomes (conversion, browse, abandonment) are the core-schema way to record a result.

Conformance levels

An emitter conforms at one of three cumulative levels - each includes the ones below. The level is declared in the manifest's telemetry.conformance_level; it's informational, not a constraint on an endpoint or a requirement on anyone else (spec 5.7).

Retrieval needs no agent cooperation - a content owner's CDN or origin reports it. Grounding and above need the agent to participate. See agent integration for the priority tiers.

Privacy levels

An agent declares a privacy_level on each conversation turn, controlling which fields it may carry. No PII at any level - the levels gate query-side signal (what the user asked, what the model classified it as), not content attribution. It's enforced: fields above the advertised level MUST NOT appear on a turn (spec 5.5).

Token counts (query_tokens, response_tokens) stay available everywhere - counting models need them and they reveal nothing about intent. Content-URL arrays (content_urls_retrieved, content_urls_cited) too: individual content events already expose content_url at every level, so minimal protects the query, not the existence of the attribution link.

Key concepts

• Source roles - origin, edge, index, agent. Multiple observers can report the same retrieval from different vantage points.
• Cross-observer correlation - the Content-Telemetry-ID HTTP header links events from different reporters into one corroborated retrieval.
• Multi-session attribution - prior_session_ids links research, comparison, and purchase sessions into a full journey.

Protocol bindings

Reference implementation & tooling

• cloudflare-worker - the open source edge worker that detects AI bot requests and reports content_retrieved events. Apache 2.0, single file, no build step. See the quickstart.
• conformance test suite - fixtures and a validator for checking emitter output against the spec (JSON Schema plus the application-layer rules the schema can't express).
• @openattribution/telemetry - TypeScript/JavaScript SDK (npm install @openattribution/telemetry). Session tracking, event builders, and an MCP session tracker. Source: telemetry-js.

A Python SDK (openattribution-telemetry) is in progress and not yet published. The JSON Schemas below are enough to validate and code-generate against in the meantime, and the wire format is plain JSON over HTTPS.

Content owners, agents, and platforms publish a manifest declaring identity and telemetry endpoints. Served at /.well-known/openattribution.json, one shape across all three roles. Trust derives from TLS plus DNS control of the domain - no central registry, no cross-consumer state.

What the manifest declares

The manifest URL appears as manifest_ref on telemetry sessions, linking events back to the publishing participant. Manifests are unsigned in v0.1; per-event signing is informational.