ERC-8501 Generative Engine Measurement Framework (Generative Engine Optimization Framework)

Draft · Other · created 2026-11-14

An onchain framework for recording how generative engines (AI search, agents) represented a subject, and letting different parties exchange and verify those measurements.

Why it is needed

Measurement metrics for generative engines differ in numerator, denominator, and unit across tools, so figures with the same name cannot be compared. Moreover, when the recipient is an agent, being cited turns into being invoked or settled, leaving an onchain trace with economic meaning, yet there is no common language linking that trace to the original exposure event. Engine operators, subject owners, and third-party panels each observe only a different part of the same process, so none of them alone can serve as the sole authority, and their statements must coexist while disclosing which frame, source, and method they rest on. Also, the same propagation act can count as success or failure depending on which intent is used as the benchmark, so claims about attainment or attribution must be handled separately from observed facts.

What to implement

In depth

Eight content-addressed record types

GEOF defines eight record types: Object, Frame, Source, Method, Intent, Metric Definition, Observation, and Profile. Every record is a JSON document arranged via RFC 8785 JSON canonicalization, and the keccak256 hash of its canonical UTF-8 bytes is its content hash. Records are identified onchain by this content hash and offchain by a URI pointing to the document. Each record type has a JSON schema as an asset, and that schema defines which fields are REQUIRED.

Object and holderClaims

An Object has a subjectId (the keccak256 of a canonical subject string), a kind (content, entity, offer, agent, skill, task, registry, claim, or an extension x:publisher/name), identifiers, and optionally relations and holderClaims. Anyone can declare an Object for any subjectId, and multiple declarations can coexist for the same subjectId. holderClaims holds a claimant and a basis (registry-record, domain-acknowledgment, identity-binding, signature, or declared), but an Object declaration by itself implies no control over the subject and no consent or veto over observations and profiles.

Frame's seven dimensions and three clocks

A Frame sets observation scope across seven dimensions: time, engines, queries, locales, objects, observers, and repetition. Each dimension's scope is one of unrestricted (not limited, though this does not mean everything was covered), enumerated (a list of intended values), or not-applicable (the dimension is meaningless), and enumerated dimensions separately track coverageObserved, the values actually covered. A Frame must specify observationGranularity and disclosureGranularity (individual, cohort, aggregate, commitment-only) and must name all three clocks: eventEnd (when the representation is finalized), processClose (when the opener freezes the evidence version), and observationCutoff (after which downstream reactions are no longer attributed to the frozen version).

Event's Origin, Course, and Conclusion structure

Event observations split into three parts. Origin captures the trigger (recipient, engine, schedule, third-party), whether it was pre-authorized, whether there was a query and its intent (e.g. informational), the engine identifier, the recipient type, and the delivery channel. Course assigns a value or state to each of five stages, retrieved, selected, synthesized, attributed, and delivered, with any stage the observer could not see explicitly marked UNOBSERVED. Conclusion has one entry per Object addressed, recording kind (OMITTED, RETRIEVED_NOT_USED, MENTIONED, CITED, RECOMMENDED, INVOKED), the actor, prominence, stance, fidelityFlags, and downstream reaction. If a single run addresses multiple Objects, each gets its own separate conclusion entry.

Five value states

Every observation field and statement value is either an actual value or one of four states. ABSENT means it was checked and was clearly not present (a fully read answer had no mention of the subject). UNOBSERVED means there was no data available to check at all (no search log or recipient signal). NOT_APPLICABLE means the field is meaningless in this context (no query in a push-type delivery). UNDETERMINABLE means data existed but the declared method could not reach a conclusion (integrity could not be confirmed, or evaluators disagreed). Reporting UNOBSERVED as zero is prohibited, and the absence of a statement is not itself a statement.

Metric identifier syntax and ten families

A Metric Definition is identified by geof:family/metric/vN, or in extension form geof:x/publisher/metric/vN. It is immutable, and a new version is required whenever the unit, numerator, denominator, eligibility conditions, deduplication, or aggregation method changes. Family is one of ten: exposure, presentation, reception, fidelity, seo, access, dynamics, economics, attainment, or attribution, and a Metric Definition must not carry any directionality about which value is desirable. attainment and attribution statements must carry an intentRef, and a Process without an Intent supports only the other eight families.

Intent, and the separation of attainment from attribution

Intent is an optional record capturing who declared it and when, whether it is ex-ante (declared beforehand) or ex-post (reconstructed afterward), the target recipient group, related subjectIds, and the expected change (received, aware, understood, remembered, relayed, adopted, invoked, acted, ceased, maintained, etc.). Because the same propagation can count as success or failure depending on which Intent it is benchmarked against, attainment statements and attribution statements are always treated as separate, distinct statements. The fact that seventy out of a hundred people completed some action is a separate question from whether they would have done so regardless.

Registry state machine

The registry must implement ERC-165 and ERC-6372, and every timestamp is the registry's clock() value. The id for declare and declareObject is keccak256(abi.encode(kind, declarant, contentHash)), and if the same account declares an identical record again, it reverts. processId is predictable as keccak256(abi.encode(opener, frameId, nonce)). commitObservations can only be called while the current version is open; closeProcess and openAddendum can only be called by the Opener; and publishProfile can be called by anyone, but only for a frozen version. The registry does not interpret or verify recorded content, it only stores hashes, URIs, attribution, and ordering. This proposal is currently in Draft status, still under review, and its details may change or be discontinued.

Things to watch out for

Original ERC · ← ERCs