Methodology.
How CiteMaybe reads a website, makes a finding, and records the evidence behind it.
What we fetch
CiteMaybeBot requests the public HTML for an approved starting URL and up to four same-origin pages. It checks its own robots rules before fetching a page. It uses internal links and limited sitemap candidates to decide where to look.
A blocked request, timeout, or unavailable robots file is recorded as an access issue or an untested check. The scanner does not execute page JavaScript. Content added later in a browser may be absent from the captured source.
Scan limits
Each scan has a 15-minute deadline, 42 outbound requests including DNS checks, and a five-MiB total read budget. Each resource has a 15-second deadline. The limit for one HTML page is one MiB after decompression. Requests to a site are spaced apart and pages are fetched one at a time.
An unavailable robots policy can stop a page. A 404 or 410 for robots.txt is treated as an absent policy. The scanner stops when a required wait exceeds its supported crawl delay. The report records what it could inspect within those limits.
Three sets of checks
- GEO readiness
- Reachability, redirects, HTTPS, and crawler-specific access. OAI-SearchBot is used for search, GPTBot for training, and ChatGPT-User for user-triggered visits. Robots directives may not apply to user-triggered visits. These observations do not establish that a provider visited, indexed, or cited the page.
- AEO coverage
- Visible text, headings, and passages associated with the questions you supply. Keyword overlap is a heuristic. A matching passage can still leave the question unanswered.
- SEO basics
- Page titles, descriptions, language, canonical and indexing directives, JSON-LD syntax and observed types, and duplicate metadata in the scanned sample. JSON-LD syntax alone cannot establish validity for all schema requirements or eligibility for rich results.
How to read an outcome
Each finding records a page, rule identifier and version, observation time, evidence, and recommendation. Deterministic checks use directly observed values. Heuristic checks interpret limited signals. A model judgment is labeled separately when available.
- Pass
- The observed value satisfied the rule.
- Warning
- The observation deserves review in the context of the page.
- Fail
- An observed condition failed the stated rule.
- Not tested
- There was insufficient access or evidence to run the check.
- Not applicable
- The rule did not apply to this observation.
Coverage is the count of passes, warnings, and failures divided by those same outcomes plus untested checks. Inapplicable checks are excluded. Reports show counts and coverage without estimating ranking or citation probability.
Answer observations
An administrator can run optional budgeted semantic reviews or API search tests when these services are configured. Search observations record the model, settings, question, answer, and actual cited sources. Brand mentions and official-domain citations are counted separately within comparable groups of valid answers.
Neither a sampled answer nor a technical audit represents what every consumer sees in an AI product.