Two runs of the same prompt can diverge because one is anonymous while the other inherits account history, location, or a selected search/grounding mode. If a dashboard blends those observations, the average looks precise but is not reproducible.

A practical split:

1. Controlled baseline fixed prompt version, market/language, engine and mode; a fresh session; timestamp; and the cited URLs.

2. Personalized diagnostic explicitly tagged as a returning session or profile, with only scenario-level interpretation.

A headline crawler-access score should disappear when the evidence cannot support it.

I would suppress the composite in four cases:

1. A critical public URL fails. A broken homepage, primary service page, or documentation entry point should not be averaged away by healthy low-priority samples.

2. The robots.txt artifact or policy inventory is missing. Without the exact policy input and expected assertions, a conformance percentage is not reproducible.

Teams sometimes say AI crawlers can access our site as if that closes the loop. It does not.

I find four evidence layers useful:

1. Policy what robots.txt declares.

2. Delivery the status, redirects, content type, and challenge behavior a real request receives.

Not every change in an AI-generated answer deserves the same response.

A missing citation on one exploratory prompt is different from an inaccurate claim repeated across high-intent questions. I think an AI-visibility severity model should combine four factors:

1. Scope one prompt, one topic cluster, or the full benchmark?

2. Claim risk cosmetic wording, commercial misinformation, or a factual/compliance problem?

A visibility score is easy to export. An audit trail is harder and much more useful.

For every metric, I think a portable AI-visibility export should include:

the exact prompt and benchmark version

provider, model, locale, parameters, and acquisition time

Most visibility discussions focus on whether a brand is mentioned. A more useful question is whether the underlying claim is safe to reuse.

A citation-ready claim has five parts:

a named entity

a specific claim

A citation can look different across AI answers while resolving to the same document: tracking parameters, redirects, HTTP/HTTPS variants, trailing slashes, mobile routes, and declared canonicals all create noise.

I think an auditable visibility report should keep two layers:

1. The raw URL exactly as the answer engine returned it.

2. A normalized record containing the resolved URL, declared canonical, registrable domain, and every transformation applied.

A visibility percentage is not very useful unless the report can explain where it came from.

Before trusting a dashboard, I think a team should be able to inspect:

the exact prompt text and version

the answer surface, locale, date, and sampling conditions

Traditional rankings do not fully describe what happens when an answer engine synthesizes a response. We are building Corank around four practical signals: how often a brand appears for important questions, whether the description is accurate, which sources are cited, and how visibility compares with competitors.

Corank (https://corank.ai) turns that evidence into AEO/GEO and technical SEO work instead of stopping at a dashboard. For teams measuring AI search today: which signal has been most useful, and what is still missing from your reporting?