AI Search Audit (Lite) skill

Version 2.110. The first number changes only on a breaking change.

AI Search Audit (Lite)

Purpose

The skill checks 1 public page against the published rules for crawling, indexing and quoting by Google, Bing and AI assistants, and writes 1 report. You add 4 judgement files; the report marks those findings as your judgement, with your model's name. The report lists problems only, with no explanations and no fixes.

Requirements

• Node 20.18 or later (node --version). If it is missing or older, tell the user and stop.
• No install: dist/read.mjs and dist/audit.mjs (with dist/chunks/) are bundled.
• $SKILL is the folder of this SKILL.md: ${CLAUDE_SKILL_DIR} in Claude Code; on claude.ai, the sandbox path you read this file from.
• Work in a writable directory of your own.
• Network: the page and a capped set of its site's files, links and images; first 1 DNS lookup and 1 RDAP request to rdap.org (sends the site's IP address only). Nothing is uploaded: no server of ours, no account.

Rules

• R1. Answer the user in the language they wrote in.
• R2. Never re-create the audit by hand. If the scripts cannot run, tell the user what would make them run, and stop. Do not fetch or check the page another way.
• R3. Write every JSON file with the file-writing tool, never a shell heredoc (heredocs break on apostrophes).
• R4. Never write Russian.
• R5. The scripts call no model. Do not add a score of your own.
• R6. Use a saved copy (--html) only when the user gives you one. Never make one with a browsing or fetch tool.
• R7. After a refusal, do not retry another URL unless the user asks.
• R8. Do not volunteer fixes. If the user asks, answer from your own knowledge and say the advice is yours, not the report's. The full version is from GetLoopLoop (https://www.getlooploop.com).

Procedure

1. Read the page (script). node "$SKILL/dist/read.mjs" <url> --out reading.json
   • Options: --render compares with a JavaScript render in a local Chrome or Chromium (an empty app shell is rendered without it); --save-html page.html keeps the fetched copy.
   • Saved copy: --html saved.html --final-url <url> (and --robots saved-robots.txt). Nothing is fetched; network checks are reported as not measured.
   • Result: reading.json and reviews.template.json; prints every cap hit.
   • On failure: REFUSED <code> (exit code 2). See Refusal codes.
2. Write questions.json and style.json (judgement). Read the text of every passage in reading.json (vs.passages) in full. Write 8 to 12 spoken questions and judge each against the passages. Mark every passage specific or generic.
3. Write brand.json (judgement). Name the site's own brand. Answer the 5 fixed questions from your own knowledge only.
4. Write reviews.json (judgement). Fill in reviews.template.json and write it as reviews.json.
5. Check the files (script). node "$SKILL/dist/audit.mjs" reading.json --questions questions.json --style style.json --brand brand.json --reviews reviews.json --check
   • Result: validates the 4 files; writes nothing. A warning line is advice, not a refusal.
   • On failure: REFUSED <file>: <reason>. Fix that file as the message says and run step 5 again.
6. Write the report (script). node "$SKILL/dist/audit.mjs" reading.json --questions questions.json --style style.json --brand brand.json --reviews reviews.json --html report.html --md report.md --json export.json
   • Add --keyword "phrase" when the user gave a target keyword; otherwise the keyword comes from the h1.
   • Result: the 3 files; prints the problems numbered as on the map, judgement findings marked (judgement).
7. Show the report. Required; see Show the report.

Judgement files

Read references/judgement-files.md before you write them: schemas, examples, full rules.

• All 4: agent (for example "Claude Code", "claude.ai", "Codex", "opencode", "Hermes") and model, at most 60 characters each. If you do not know one, leave it out. Never guess either.
• Letters: the script refuses a letter Russian uses and the page's language does not (on a Ukrainian page, the 4 Russian-only letters; Bulgarian and Belarusian keep their own). Quotes are checked against the passage.
• questions.json: 8 to 12 spoken, first-person questions in the page's language.
  • axes: 1 to 3 of urgency, device, object, obstacle; each axis used at least once.
  • Most without the page's keyword. No wake words. None that is another with words added. At least 1 the page probably does not answer.
  • Judge from the passages only, never from what you know. passage: the index of a passage in vs.passages (0-based), or null. verdict: complete, partial or missing.
  • quote: verbatim from the passage, at most 400 characters, none for missing.
• style.json: 1 verdict per passage: generic (fits a competitor's page with the brand name changed) or specific. evidence: verbatim, at most 200 characters. Judge the passage, not who wrote it: this is not an AI detector.
• brand.json: basis: "own_knowledge"; answer what_is, pricing, integrations, weaknesses, alternatives in English or the page's language.
  • Not sure (for example a price range): "knows": true, and write "I am not sure" in the answer.
  • Do not know: "knows": false, "answer": "".
  • The script sets knows to false when the first 400 characters disclaim knowledge ("I don't have information", "I'm not familiar", "I don't know of"). "I am not sure" does not.
• reviews.json: basis: "own_knowledge"; every platform of the template once, none added or removed; "knows": true with a rating on its scale, or "knows": false. Judge each platform separately, from your own knowledge of the brand on that platform.
• Honesty (brand and reviews): answer from what you already knew. No browsing, search, fetch or other tool; not from the page. "I don't know" is a correct answer. Do not present a guess as a fact or copy the page into an answer.

Verification

• Every JSON file was written with the file-writing tool.
• Step 5 printed no REFUSED.
• report.html exists.

Show the report

Required. The HTML report is the result; a chat message alone, or the file name alone, has not shown it.

If the user or caller said not to publish or open anything, stop when the HTML file exists and give its absolute path. This overrides the bullets below.

Present one file: report.html. Keep report.md and export.json beside it; do not present, attach, open or preview them. Never open or preview export.json or report.md as the result, and never end the reply on either: the HTML is the report. If the environment shows the last file you present, it must be the HTML.

• claude.ai or the Claude desktop app: save the 3 files where your instructions say user files go, and present only report.html as a file in your reply, not even followed by the others. A file in a scratch folder the user cannot see is not shown. Do not retype the report into an artifact.
• Claude Code: publish only report.html with an HTML or artifact publishing tool, if the session has one, and give the link. Otherwise open it: start "" "<absolute path>" (Windows), open "<absolute path>" (macOS), xdg-open "<absolute path>" (Linux), and give the path.
• Codex, opencode, Hermes Agent or any other agent: give the absolute path of report.html and offer to open it with the command above or a preview tool. Preview only the HTML. Name the other 2 files only if the user asks for the data.

The report makes no network requests. It opens in English and in the light theme; add #lang=uk, #lang=pl or #theme=dark (for example report.html#lang=pl&theme=dark). Give the #lang link when the user writes in Ukrainian or Polish.

Chat summary

• First, if present: ai_search_blocked, search_engine_blocked, robots_unreachable, snippet_blocked or noindex.
• The verdict line and the counts by severity.
• Problems 1 to 5, each with its place.
• Every cap the audit printed (links past 30, images past 20, alternates past 10, child sitemaps past 5, merged sections, a render that could not run).
• Once: the checks are deterministic; the questions, brand and reviews are your judgement as 1 model (name it), not what ChatGPT, Gemini or Perplexity say (the web Brandscope and Reviewscope ask 3 assistants).
• 1 line: Pattern score and band, marker words per 1,000, passages you judged generic (derived, derived, your judgement). Never say the page was or was not written by AI.
• Never claim an assistant has indexed, ranks or will cite the page.

Refusal codes

Code	Tell the user
invalid_url	Not a web address. Ask for the full URL.
blocked_scheme	Only http and https pages can be read.
blocked_host, blocked_address	Not a public page: localhost, a private network or a non-standard port.
unreachable	The page could not be reached. Read references/network-and-limits.md.
http_error	An error status (in the reading), or more than 5 redirects.
blocked	The site turns automated readers away (HTTP 401, 403 or 429). Retrying will not help.
not_html	A file such as a PDF or an image, not an HTML page.
too_large	Over the 3 MB ceiling.
empty	No readable text after navigation and boilerplate were removed.
refused_ru_domain	A Russian domain zone (.ru, .su, .рф, .москва, .moscow); not fetched.
refused_ru_hosting	Hosted in Russia (RDAP country RU); not fetched.
refused_ru_audience	The site declares a Russian-language version (hreflang, og:locale or Content-Language).
refused_language_ru	The page is in Russian.

On a refused_ru_* or refused_language_ru code: skip steps 2 to 5, run step 6 without the judgement files, show the report, tell the user in 1 line that this tool does not analyse such pages (the code's line), and stop. Do not analyse the page another way.

References

File	Read when
references/judgement-files.md	Before steps 2 to 4.
references/checks.md	The user asks about a finding code, tile, severity, threshold or check.
references/network-and-limits.md	What is sent; a failed fetch; no network; a read-only skill folder.
references/outputs.md	The user asks for the data or what a file holds.