Minimum disclosure requirements for AI-visibility claims

Claims about AI visibility are now routine commercial and editorial currency: an entity “shows up in ChatGPT,” is “recommended by AI assistants,” “gained 40% AI visibility.” Most such claims cannot be evaluated by the people asked to act on them, because the facts needed to judge a claim — how many observations, over what period, on which system, asked how — are not stated. The public evidence is unambiguous that these facts are load-bearing: rewording a question while preserving its intent replaces much of an answer’s recommendation set [1], and re-issuing an identical question a day later replaces a large share of its cited sources and mentioned brands [2]. Against that instability, an undisclosed claim is not weak evidence; it is no evidence.

This convention states the minimum a published AI-visibility claim must disclose to be evaluable at all. It is the shortest document in this series by design. It defines no metrics (BA-C-2 does), prescribes no sampling design (BA-C-3 does), and takes no position on how a measurement is implemented. It states only what a claim must say.

Normative note. The key words MUST, MUST NOT, SHOULD, and MAY carry their established normative meanings. A published claim that violates a MUST is not a measurement under this convention. Requirements are stated so that compliance can be checked from the published claim alone.

1. Scope

This convention applies to any published claim that a named entity is, is not, became, or ceased to be visible in an AI assistant or answer engine — whether the claim appears in a research report, a dashboard, marketing or sales material, a case study, or press coverage. It applies to point-in-time claims (“mentioned in 62% of answers”) and to change claims (“visibility doubled after the campaign”) alike. It does not apply to statements framed as anecdote and presented as nothing more (“here is one answer we received”), provided no quantitative or comparative conclusion is drawn from them.

2. The minimum disclosure set

A published AI-visibility claim MUST state every item below that applies to it, either inline or in an unambiguously linked note:

A claim that omits an applicable item is not a measurement under this convention; it is an anecdote formatted as a statistic. The single most common omission is item 3: a percentage without its sample size cannot be distinguished from a coin flip reported to two significant figures. The second most common is item 6: engines change without notice, so a claim without a window describes no defined quantity.

A conforming claim fits in one sentence, mapped part by part to the ten disclosures in Figure 1. For example:

Entity X was mentioned in 62% of answers (95% CI 53–70%, Wilson; n = 120) for the information need “recommended providers of Y,” sampled across 14 phrasings on Engine Z (consumer interface, model build of 2026-06), 2026-06-01 → 2026-06-28, region DE; refusal rate 3%; mentions detected by lexicon matching (precision/recall 0.97/0.94 on a labeled sample).

The definitions of the quantities named above — Visibility Probability, Share of Voice, Discovery Depth, and their supporting measures — are those of BA-C-2, and metric-specific disclosures required there (competitor set and weighting for SoV, constraint dimensions for DD, classifier accuracy for sentiment) apply in addition to this list.

3. Change claims

A claim that visibility changed — improved, declined, jumped after an intervention — asserts more than two point-in-time claims, because engines drift on their own at every time scale [1] [2]. A published change claim MUST satisfy the Barkhausen Criterion as formalized in BA-C-3: significance established between proper interval estimates; sustainment across at least $K \geq 2$ consecutive windows, with $K$ and the window length disclosed; non-coincidence with a flagged engine change, or re-basing after it; and disclosure of how many cells were monitored, with the multiplicity treatment stated. A before-and-after pair of screenshots, or a comparison of two undated percentages, does not meet this requirement under any circumstances.

4. Compliance and citation

A claim, report, or dashboard MAY state compliance as: disclosed in conformance with BA-C-4 v1.0. Compliance refers to disclosure only; it is not an endorsement of the underlying design, whose validity is governed by BA-C-3.

Reviewers, journalists, and procurement evaluators MAY cite non-compliance neutrally: the claim does not meet the minimum disclosure requirements of BA-C-4. This phrasing makes no accusation about the claim’s truth — it records that the claim, as published, cannot be evaluated. That is the purpose of a disclosure floor: it separates claims that can be checked from claims that must be taken on faith, before any argument about who is right.

5. What this convention deliberately does not require

The requirements above are a floor, chosen so that any competent measurer can comply without surrendering legitimate confidential material. This convention does NOT require publication of raw answer data, of the specific query phrasings used, of prompt-construction or variant-engineering methods, or of any tooling. It requires counts, definitions, and conditions — the statistical design layer a reader needs to judge whether a conclusion follows — and nothing from the operational layer beneath it. A measurer who cannot state items 1–10 about a claim does not have a disclosure problem; they have a measurement problem.

Limitations

Minimum disclosure is necessary for evaluation, not sufficient for validity. A claim can disclose all ten items and still rest on a defective design — too few phrasings, uncontrolled personalization, cherry-picked windows; validity is the province of BA-C-3, and a reader of a fully disclosed claim still has to judge the design the disclosure reveals.

Disclosure can also be false. This convention makes fabricated disclosures checkable in principle — a stated $n$, window, and engine version are concrete claims a challenger can attempt to reproduce — but it cannot prevent them; no disclosure convention can.

Finally, the disclosure set reflects the engines as they exist in the stated windows of the evidence cited here. These results describe the engines as sampled during their respective windows; engines change without notice, and results should be assumed perishable. If future engine architectures make new conditions load-bearing — new personalization surfaces, new answer formats — the set will grow by a new version of this convention, not by silent reinterpretation.

@techreport{BA-C-4,
  author       = {{Barkhausen AI}},
  title        = {Minimum disclosure requirements for AI-visibility claims},
  institution  = {Barkhausen AI},
  year         = {2026},
  url          = {https://barkhausen.ai/conventions/minimum-disclosure/}
}