citation-check

Reads: search/questions.md; ../improve-search-visibility/references/citation-protocol.md; the last run record's per-question snapshot (../improve-search-visibility/references/run-memory.md); the canonical brand facts from the brain (for accuracy checks on brand questions); the tracked competitor list.

Does:

1. Free pass (step 1): ChatGPT, AI Overview and AI Mode in the host's own browser (private window, signed out; stop on any CAPTCHA or refusal and mark the surface), and the host's web search for who ranks, labelled "web search". Paid top-up (step 3, only after the one yes): Google's exact page one through gooseworks:data_call_provider (provider scrapecreators, path /v1/google/search, the query and region), inside the ceiling (the spend-gate skill (../spend-gate/SKILL.md) proxy branch), stopping at it. No browser → the minimum paste set, offered in the first message.
2. Run — ChatGPT once per question and three times for the key questions; AI Overview and AI Mode once each when Google allows; Google page one once per question in the top-up; pasted: 5 questions (up to 4 key questions + at least one brand question), ChatGPT twice, AI Overview and AI Mode once each. New chat per question, memory off, never the brand name in non-brand prompts. Wait for each answer to finish (protocol's wait rule); unfinished answers are retried once, never scored as misses.
3. Only this run's observations count. An earlier run's answers are the baseline from the last run record (../improve-search-visibility/references/run-memory.md): shown as "last checked <date>" beside today's reading and compared only like with like (same question version, same route and location, the noise rule there), never counted in today's numbers or tables. A surface not checked today says "not checked today".
4. Record every run with the full field list; save verbatim answers. Note whether ChatGPT searched (sources shown) — answers without search come from training memory.
5. Classify every cited URL by source type. For brand questions, list each stated fact and compare it to the brain: correct / wrong (quote both) / not on file.
6. Score each row (named k of N, won/partial/lost/no-brand answer, prominence, stable or unstable), then mention rate per surface and stage, own-citation rate, third-party coverage and share of voice against the brands the engines named. Won = ≥ ⅔ of runs; unstable = some but not all; one run is never stable.

Writes: the rows table used by the report and the verbatim runs file (search/visibility/<date>-runs.md, not imported); every charge is logged by the spend gate, and the report states credits charged against the ceiling in the gate's own words. Never: reports one run as fact; passes an organic or API read off as what buyers see in the AI apps; ranks the brand "#2 in ChatGPT".