GenRank Governance Policy

Every mature measurement regime, from credit ratings and financial benchmarks to media measurement and statutory audit, treats the separation between scoring and selling score improvementas load-bearing rather than optional. The AI-visibility field, as of the IAB’s August 2026 framework, has written no such rule. GenRank adopts one unilaterally.

1. Separation of measurement from advice

GenRank does not sell, and will not sell, any service to an entity it ranks that is intended to change that entity’s rank.

This follows the line drawn consistently by EU CRA Regulation 1060/2009 Annex I §B.4, SEC Rule 17g-5(c), and Reg S-X 2-01(c)(4):

Two rules that admit no exception, because disclosure cannot cure them:

• No fee may depend on the level or direction of any score (cf. CRA §B.3c).
• Personnel who set methodology may not participate in sales, nor be influenced by sales considerations (cf. 17g-5(c)(8)).

Correction is always free. Where an entity believes it is mismeasured, the remedy is a free correction request, never a paid service. No paid product may change a published number.

2. Material change

Following IOSCO Principle 12, a material change is any change that would move a material number of published scores or alter what a score means:

• adding, removing or re-weighting a model
• changing the scoring or aggregation formula
• changing sampling design (including the number of samples)
• changing the question corpus in a way that alters index composition
• changing entity resolution such that entities merge or split

For each: publish in advance, state the estimated magnitude of effect, keep the prior version documented, and stamp the methodology version on every affected row. Comments received and our responses to them are published. The running record is the changelog and corrections log.

3. Corrections and complaints

Anyone may dispute a published number. Complaints are handled by someone not responsible for the determination being challenged (IOSCO P16). Substantiated errors are corrected, and the correction is published with its cause. We keep a five-year audit trail sufficient to reconstruct any determination: the question, the model versions, the raw responses, the sample count, and the extraction that produced each entity.

4. What we publish, always

Alongside every score: the number of trials behind it, its confidence interval where estimable, the model set and versions, the grounding state, the decoding parameters, and the methodology version. Per IOSCO P9, enough with each determination for a reader to understand how it was reached, and MRC Minimum Standard B.8, which requires that the formula for the error, not merely the error, be disclosed.

5. What we do not claim

• We do not claim to measure what a consumer sees in a chat UI. We poll APIs, and UI/API divergence is large.
• We do not claim cardinal accuracy. We defend the ordering.
• We do not cite agreement with other vendors as evidence of validity. Three tools reading the same signal agree with each other and can all be wrong together.
• We do not claim Decision-Grade for per-question rows. Those are Directional; category rollups are where the aggregate is strong enough to claim more.

6. Independent oversight (committed, not yet stood up)

IOSCO P5 and EU BMR Art. 4(3) prescribe the same remedy for structural conflict: an oversight function with balanced stakeholder representation. EU BMR Art. 24(7) adds the idea worth copying, a voluntary opt-in to binding oversight with a four-year lock-in, which makes the signal expensive to fake.

Every incumbent measurement regime accepted oversight after being caught. Volunteering before the crisis is the one variant nobody has tried. This is an open commitment with no date attached yet, and it is listed here rather than omitted so that it can be held against us.