Cited by AI? Start with the misses.

Five labels that never collapse into one score

These are observation labels, not quality judgments. An answer can print a URL and still be wrong. A miss can coexist with real citations in another account, session, locale, model revision, or interface. The page does not authenticate the provider, evaluate the answer's truth, or estimate how often another user sees anything.

The published half: least flattering first

The frozen panel has 6 prompts. The public receipt set shows exactly 3: the lowest outcome rank, then prompt id. All-panel counts remain visible above, so withholding the other individual receipts cannot hide the denominator or improve the headline. The raw response prose stays in the private sealed store; this page exposes the exact prompt, mechanical label, target URL token when present, and response hash needed to audit the classification without republishing unverified model prose.

Selection rule: MISS, then name-only mention, then answer URL mention, with UNKNOWN listed separately rather than treated as a negative. Ties break by frozen prompt id. No human selected a favorable answer after seeing the responses.

How to reproduce the baseline

The privacy-safe frozen manifest below publishes all six prompts and target URLs. It does not disclose the three individual answer outcomes outside the least-flattering receipt subset.

1. Use the six exact prompts in the frozen manifest above. Do not add our name, domain, or target URL to a prompt that did not contain it.
2. Run one fresh answer per prompt against the named assistant surface. Record UTC date, displayed model label, account or API surface, locale, memory state, and whether search or grounding was enabled.
3. Retain the raw response privately and compute SHA-256 over its UTF-8 bytes. Preserve provider citation targets separately from URLs scraped out of answer prose.
4. For answer-text URL matching, lowercase only for comparison, remove terminal punctuation, and require the exact target URL or a descendant path. A same-domain link to another page is not a target match.
5. Apply exact asset-name matching only after URL matching. No fuzzy match, stemming, synonym, sentiment, or model judgment may promote a row.
6. Record missing rows, fetch errors, empty bodies, unreadable metadata, and hash failures as UNKNOWN. Publish the least-flattering half by the frozen rule before reading any aggregate conclusion.

The baseline assistant was a local loopback model with search calls fixed at zero. That makes it a control for the mechanics, not evidence about commercial assistants. The collector had no paid provider and no paid fallback; its sealed cost estimate was exactly $0. It is also a self-probe, so it is not a visitor, referral, lead, or demand event.

What the response hash proves

A matching SHA-256 proves only that the stored response bytes did not change between sealing and verification. It does not prove who produced the answer, whether the displayed model label was accurate, whether the answer was grounded, whether a URL is true, or whether another session returned the same text. The collection receipt, prompt manifest, response hash, and classification are separate facts so one cannot silently stand in for another.

The older series is reliability evidence, not dataset evidence

The pre-existing house-brand panel contains 468 rows across 36 UTC dates. 161 rows have an answer body and 307 carry a collection error. It has 14 URL-like answer-text elements across 14 rows and zero provider-authenticated citation targets. Every row came from one local assistant family. That panel asked about the company brand and domain, not the owned datasets on this page, so all of it is excluded from the new dataset-citation denominator.

The table below publishes exactly half of the older dates using its own least-flattering rule: fewest URL-like rows, then fewest answer bodies, then most UNKNOWN errors. These dates are shown because collection failure is part of the historical record. They are not called misses, and they are never credited as citations.

Legacy URL-like elements were extracted from answer text. They are not provider grounding metadata. The older panel remains useful for exposing pipeline reliability and reply-format failures, not for claiming assistant visibility.

Crawler telemetry stays outside this result

Our public surfaces maintain separate user-agent access logs and daily crawler census artifacts. Those records can show that a client identifying itself as GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, ChatGPT-User, or another token requested a URL. They do not contain the assistant answer, the user's prompt, a source target, or a referral chain. User-agent text is also self-asserted and can be copied.

For those reasons, no crawler record is joined into this page's numerator or denominator. A crawl can precede a citation, but it cannot prove one. A generic browser fetch, a bot fetch, our own health check, and this collection panel are four different event classes. None is automatically a human visit or demand signal.

The same boundary applies in reverse: a dated assistant transcript can contain a URL even when our access log did not capture a matching fetch. That may be cache, provider proxying, an old index, or missing telemetry. The answer receipt remains an answer observation; the access log remains access telemetry. This page does not invent a causal link between them.

Turn your supplied sessions into a bounded QA table

If you are a GEO or SEO agency lead who already collects assistant answers for a client but cannot grant an outside monitoring vendor account access, the existing AI Answer Repeatability Report applies the same separation to a frozen customer-supplied panel.

Boundaries of the paid report

The report organizes sessions your agency collected. It is not an assurance engagement, benchmark, ranking, share-of-voice score, legal record, advertising substantiation, provider authentication, or proof of what another user saw. Capture-context fields are operator-asserted and unverified. Full account exports, credentials, account headers, unrelated prompts, private client records, and third-party personal information are not accepted.

Read the complete fixed engagement and retention limits · Browse data services · Browse the public evidence index