ERC-8501: Generative Engine Optimization Framework

, , , ,

Summary

GEOF is a descriptive framework for recording how generative engines — answer engines, AI search surfaces, and agent runtimes with discovery layers — represent an object, how that representation is measured, and how independent parties exchange and verify the measurements.

It lets independent parties declare, for the same object or dissemination process, what they observed; and, when a dissemination intent has been declared, how much of it was attained and how much can be attributed to the dissemination.

It assigns no desirability to any value. No thresholds, weights, ranking algorithms, engine policies, anti-abuse rules or licensing mechanics, and no definition of “good GEO”.

ERC: ERC-8501: Generative Engine Optimization Framework (Draft)
Repository: GitHub - garyyang-finchip/geo-framework: AI Generative Engine Optimization Framework · GitHub
ERC pull request: Add ERC: Generative Engine Optimization Framework by garyyang-finchip · Pull Request #2176 · ethereum/ERCs · GitHub (merged 2026-11-17)

Why

  • Generative engines retrieve, select, synthesize and attribute. Classic search metrics (impressions, clicks, position) describe part of that; the metrics that grew up around generative engines (mention rate, citation rate, share of voice, position-adjusted word count) are published under incompatible units and denominators. Two parties holding the same number cannot tell whether they measured the same thing.
  • When the recipient is an agent and the engine is an agent runtime, “being cited” becomes “being invoked” and “being settled”. These outcomes already leave on-chain traces (skill-token calls, task-tender settlements) but no shared vocabulary links them to the discovery event that caused them.
  • The engine operator, the object holder and third-party panels each observe different parts of the same process. None can be the canonical source; their statements must coexist, each attributable and each bound to the frame, sources and methods it used.
  • The same dissemination is a success under one intent and a failure under another. Statements about attainment or contribution are only meaningful relative to a declared intent, and must stay separate from statements about what was observed.

What it defines

Eight content-addressed record types (JCS canonicalization + keccak256): Object, Frame, Source, Method, Intent, Metric Definition, Observation, Profile.

Process: an observation container bound to one Frame, one or more Objects, zero or more Intents, and the Sources and Methods the opener announces. Any account commits observation batches (Merkle root + count, attributable to the sender) and publishes Profiles; only the opener freezes a version or opens an addendum. Profiles bind to a frozen version. Nothing designates an official profile.

Five value states: ABSENT (checked, none), UNOBSERVED (no material), NOT_APPLICABLE, UNDETERMINABLE (material, but the declared method reached no determination), or a value. UNOBSERVED is never zero.

Metric grammar geof:<family>/<metric>/vN (extensions geof:x/<publisher>/<metric>/vN) over ten families: exposure, presentation, reception, fidelity, seo, access, dynamics, economics, attainment, attribution. A metric definition fixes statistical unit, numerator, denominator, eligibility, de-duplication and aggregation, and carries no preferred direction. Every statement carries its frame, sources, method, counts, coverage, uncertainty and a human-readable rationale.

Three rules that keep the statements honest

  1. Fidelity statements must name a reference set, a comparator and the proposition checked. “Did it happen”, “was it relayed faithfully”, “are the source’s claims supported” and “was the purpose achieved” are four questions with four answers.
  2. Attainment and attribution statements must reference an Intent. Base observation stands without one; claims of effectiveness do not.
  3. Representation ends at INVOKED (with the actor named: engine or recipient). A purchase, a settled task or a payment is a follow-up observation with a stated link basis, never something the engine “did”.

A minimal registry (IGEOFRegistry, ERC-165 + ERC-6372): declare, declareObject, openProcess, commitObservations, closeProcess, openAddendum, publishProfile, plus views. Hashes, URIs, attribution and ordering only.

Two worked cases, one framework

Case A — web correction notice Case B — agentic skill
Intent public searchers should understand condition Z three named agents should invoke and settle
Exposure answer-inclusion 0.74 (1066/1440) public retrieval 0.042 (9/212); targeted reach 3/3
Fidelity key proposition present in 0.23 of mentioning answers descriptor hash match 11/11
Attainment UNOBSERVED — no recipient evidence 1.0 for invoked, 1.0 for acted
Attribution not stated UNDETERMINABLE, basis declared
Reading wide, unfaithful, effect unknown narrow, faithful, attained, contribution unproven

Neither case gets a score. Both are walked on chain in the reference tests and on Sepolia, and the record bundles and derived vectors are in the repository.

Relation to adjacent work

Independent of, and mappable to, an information-flow paradigm registry (ERC-8489): a GEO process may span several information flows, and a monthly platform report supports a profile without any single flow being visible. Objects of kind skill / task / agent / registry reference records in an on-chain value registry (ERC-8444). Skill calls (ERC-8338) and task-tender settlements (ERC-8414) are on-chain sources for invocation and follow-up statements. Identity and credit (ERC-8434, ERC-8419), safety adoption (ERC-8468) and dispute resolution (ERC-8471) consume profiles through their own interfaces. The ERC text requires only ERC-165 and ERC-6372.

Status and asks

  • Merged into ethereum/ERCs as a Draft (ERC-8501). Reference registry with 50 Foundry tests (both cases end to end), JSON Schemas, validator, vectors. Sepolia evidence run complete: registry 0xe725180476760B4012D8Eb9aD1De3337B5e40478, both cases on chain in 43 transactions (details in the reply below and in docs/sepolia-evidence.md in the repository).
  • Looking for review on: (1) the five-state value vocabulary; (2) opener-only closure with open commitment — is a per-observer closure needed in the kernel, or is “open your own process on the same frame” enough; (3) whether UNDETERMINABLE should be allowed as a statement value outside the fidelity, attainment and attribution families; (4) the metric identifier grammar’s publisher namespace.

Sepolia evidence run — both worked cases on chain

The reference GEOFRegistry is deployed on Sepolia at 0xe725180476760B4012D8Eb9aD1De3337B5e40478 (deployment tx, block 12134766). Both worked cases from the draft were walked end to end: 43 transactions across three accounts, blocks 12134766–12134808, all successful.

  • Case A (web correction notice): two Objects, an Intent and a publisher-side Source declared by the company; Frame, two Sources and four Methods declared by a third-party panel that opens the Process; four EVENT batches and one AGGREGATE committed by two different accounts; version 1 frozen; two Profiles published against it by the panel and an auditor; an addendum for a late FOLLOWUP batch; version 2 frozen; the panel supersedes its own Profile while the auditor’s stands. Process 0xc648a9d15bfab71c7ce298a5e4f345f8142d07f5a1098f5ab5a36672a1835ae1.
  • Case B (agentic skill): skill Object with a value-registry reference, Intent toward three named agents, Frame with queries = not-applicable; discovery EVENT and gateway AGGREGATE from the provider, settlements as a FOLLOWUP batch from an independent indexer; version 1 frozen; two Profiles. Process 0x80277a1a8b0a74daa3d0682fd2f4744993aaddcca2435a187c387491fe330af4.

Reads at block 12134808 confirm the model’s guarantees: a frozen version’s commitment count and evidence root do not change when an addendum is opened; a Profile binds to a frozen version only; supersedes resolves to the same publisher’s earlier Profile; ERC-6372 mode=timestamp.

Full record (every tx hash, gas, ids, state snapshot, cast commands to reproduce the reads): deployments/sepolia.json and docs/sepolia-evidence.md in the repository. One operational note for implementers: Sepolia charged 1x–10x the gas of the same transactions on a local chain, so the evidence driver skips simulation and takes limits from live estimation.

ERC-8501 has been merged into ethereum/ERCs as a Draft (PR #2176): ERC-8501: Generative Engine Optimization Framework. Further revisions will come as “Update ERC-8501” PRs; the four review questions in the opening post still stand.