AI discoverability audit
We audit the conditions your website controls, and report each one against a captured, timestamped observation you can open yourself.
robots.txt and the sitemap,
fetches up to eight pages and runs thirteen deterministic checks. About
twenty seconds later it returns a sheet: each check it settled, its
rating, and one finding in full with the captured response. The sheet
shown is a real published scan.
Nothing is fetched before authorization is recorded, and nothing is rated before a person has read it.
01
Who authorized the audit, their relationship to the domain, the date, and a note. The tool will not fetch without them.
Permission to audit and permission to publish are recorded separately.
02
Up to two hundred pages, depth three, one second between requests, with caps on redirects, retries, response size and total runtime. Robots directives are respected. No authenticated content, no form submission, no personal data collected.
Every response is stored and hashed, and each finding cites a specific record.
03
Each deterministic finding is spot-checked against the page it describes. This step has caught twelve false findings, among them a page reported as 26 % markup when it was 90 %, and one footer badge read as twenty-two separate claims.
A finding that fails this step never reaches a report. Each correction is recorded in the repository.
04
Thirteen of the thirty-three checks are deterministic and self-rate. An operator reads the rest and writes down the reason for each rating.
Anything the evidence can't settle is recorded as inconclusive. Its points leave the denominator instead of counting as zero, and the report says so.
05
An HTML report, a machine-readable JSON appendix, and a prioritized backlog. The score shows its denominator, coverage and confidence, so a high grade on thin coverage is visible as such.
The identity, offerings and relationships your site declares, read back. Thirteen deterministic checks; the other twenty named as not covered. No account, no card.
The full rubric, the evidenced report, the JSON appendix, and a prioritized backlog.
The Implementation Audit, plus manual verification of the findings that need it, and three rechecks within 90 days that re-run the thirteen machine-settled checks and report what moved.
A monthly re-collection after an audit, scored against the same rubric, with a report of what moved.
No. This audits whether a retrieval system can fetch your pages, tell which organization they belong to, find a passage that stands on its own, and find support behind a claim. The report has no keywords, positions or traffic estimates.
Unknowable, and the report says so. The audit measures the conditions on your side. Whether an engine cites you depends on its index, ranking and prompt.
An HTML report, a JSON appendix of the same data, a backlog ranked by impact against effort, and a paste-ready remediation prompt. Every finding names its evidence records, each with the URL, observation time, response headers and a content hash.
Both. Thirteen of the 33 checks are deterministic and rate themselves. For the rest, an assisted review proposes a rating and an operator who has read the pages decides it.
Collection takes under three minutes. Review takes 38 to 55 minutes of operator time. From you: a short intake and the questions your customers ask.
Yes. The operator's own site is published unedited at citeunseen.io/sample/. It scores 58.1% (44.75 of 77 resolvable points), a C, with nine checks inconclusive.
It is the operator's own site, published as it came out, with nothing fixed first and no finding softened.
Or see the full report, the site before and after fixes, or the rubric.