Collect ai citability

Measure whether answer engines and AI assistants can FETCH this documentation at all, and whether there is anything in it worth quoting. Finds what no traffic-based check can see: the reader who asked an assistant, got their answer from your page, and never visited — and the page that cannot be cited however well it is written, because it renders client-side, sits behind a robots rule nobody remembers adding, or serves a challenge to a named crawler while a browser gets 200. Returns scores out of 100 on five separate axes — crawler_access, machine_surfaces, atom_citability, answer_shape, version_freshness — each with the evidence it rests on, plus ranked findings. Scores are computed from the observations, not written by the model, so two runs are comparable. Use it for 'why does ChatGPT never cite us', 'is our GEO/AEO set up', 'can AI crawlers read our docs', 'do we need llms.txt', 'are we invisible to answer engines', «почему нас не цитирует ИИ», «видят ли нас ассистенты». This is the fetchability and citability probe; it does not judge writing quality or search rankings, and it does not say which of its findings to fix first — that is docsbook_expert, which ranks them against what you are actually trying to achieve. Returns the finished result — a validated collect_ai_citability.v1 payload: an evidence map every claim cites, scores computed by code rather than written by the model, and ranked findings. Numbers that trace to no evidence fail the contract rather than shipping, so an invented figure is not a risk you have to check for. Changes nothing; safe on a read-only token. Typical wait 5–60 s: this call crawls the site itself, one fetch after another, and a slow or unreachable host is what stretches it. Nothing is returned until it finishes. Evidence, not a verdict — nothing here says what the gap MEANS. If you have not already asked docsbook_expert how to read it, ask: it says what this evidence is worth against, and what to do with it. One call, changes nothing.