Skip to content

olexan-art/ai-search-audit

v2.113.0MIT

AI Search Audit Lite: standard search and AI-search checks that rest on published rules (access, robots.txt for search and AI crawlers, the Google and Bing checklist, structured data, language versions, the Vectorscope reading), as a visual report. The scripts call no model.

English · Українська · Polski

AI Search Audit Lite

AI Search Audit Lite 2.113, an Agent Skill from GetLoopLoop for claude.ai and the Claude desktop app, Claude Code, Codex, opencode, Hermes Agent and other agents that load SKILL.md skills. Versions count in thousandths (2.101, 2.102, …), so the first number moves only on a breaking change.

Before Google, Bing or an AI assistant can show or quote a page, a crawler has to be allowed in, the page has to say what it is, and the text has to survive the machine's reading. ai-search-audit checks one public page against the standard, published rules for all three, adds three readings only an AI can make (the questions people ask, how the agent's own model describes the brand, and what it believes about the brand's reviews), and draws the result as one report: six status tiles, a map of the page built from its real document with every problem marked where it is, the biggest problems, the bots in robots.txt, the Google and Bing checklist, the language versions, where the text goes on its way to an AI retriever, and, at the bottom, how generated the text sounds. The report reads in English, Ukrainian or Polish.

Lite is a picture of the problems: each one is a short label, a severity and a place. It does not describe the problems or say how to fix them. A full version with explanations is available from GetLoopLoop.

AI Search Audit Lite on the Vectorscope page: the headline, the three cards, the findings to review first and where they are

What it checks

The scripts call no model and no embedding API. Every finding carries its layer: measured (a script counted it), derived (a fixed formula over measured values) or agent's judgement (written by the agent running the skill, into a JSON file the scripts validate, and badged in the report with the agent's model).

  • Access and indexing: the HTTP status, the redirect chain, HTTPS and HSTS; robots.txt for Googlebot, Bingbot and the AI crawlers (the Vectorscope's RFC 9309 matching); meta robots and X-Robots-Tag noindex; snippet controls (nosnippet, max-snippet:0, data-nosnippet over the main text: Google requires a page to be "indexed and eligible to be shown in Google Search with a snippet" for AI features); whether the sitemap lists the page; llms.txt, with the note that Google documents no use of it; a Markdown copy of the page; ai-catalog.json; a declared IndexNow key file.
  • The Google and Bing checklist (35 cells): title and description length, one h1, heading order, canonical, internal links, broken links (the first 30 checked, the rest counted), images (alt, width and height, lazy loading below the fold and not on the first image, modern formats, size from a HEAD of up to 20, srcset), loading priorities (blocking scripts in the head, a preloaded image that is also lazy, fetchpriority="high" on more than two images), viewport, html lang, Open Graph and Twitter tags, the favicon, the WebSite name, mixed content, a soft 404 (one path that cannot exist), the http, www and trailing-slash variants, page weight, a readable URL, hreflang pairs, FAQ and breadcrumb markup, and trust signals (an author on articles, About and Contact links, the organisation's name, logo and sameAs, a visible date that matches the markup, privacy and terms), link text that names nothing ("here", "read more"), #links with no target, nofollow on internal links, image links with no name and one alt on many images, an invalid html lang, the sitemap's lastmod dates and a Sitemap line in robots.txt, the preview image, and the HTML response's compression, Cache-Control and security headers.
  • Structured data: JSON-LD parsed, required and recommended properties as Google's pages list them for FAQPage, Product, SoftwareApplication, BreadcrumbList, Event, JobPosting, LocalBusiness, VideoObject, QAPage, WebSite, ProductGroup, Offer, Review, AggregateRating, Article, Organization and Person, and retired rich results noted as informational with the date of Google's entry (FAQ, How-to, course info, ClaimReview, special announcement, estimated salary, vehicle listing, practice problem, the sitelinks search box).
  • AI reading: the Vectorscope core (passages, the 26-passage cap, where the text goes), plus a hero lede lost with a removed <header>, headings in hidden panels (hidden="until-found" is findable text, not hidden), loading placeholders, words glued in the h1, navigation inside the text, thin sections, cut-off teasers and repeated claims. An empty single-page-app shell is detected and rendered in a local Chrome on its own, and the sections that exist only after JavaScript are drawn dashed on the map.
  • AI quotes it right: FAQ markup against the visible text; the questions people ask (8 to 12 spoken questions the agent writes along the Voicescope's four axes, each judged complete, partial or missing against one passage, with a verbatim quote the script checks); writing patterns (the PatternScope's deterministic detectors: stock phrases, repeated sentence shapes, uniform openings, unchecked claims, with positions and a density per 1,000 words; signals, not an AI detector); the target keyword (--keyword, or one read off the h1: absent from title and h1, written two ways, confined to one section, most of the text off it, a promise with no figure).
  • AI gets the brand right: the agent's model answers Brandscope's five buyer questions from its own knowledge (no browsing; "doesn't know" allowed), and the script compares them with the page: a price or a count the page states differently, a fact the page lacks, a brand the model does not know. For reviews, the model states its belief about the brand's rating on Reviewscope's 12 platforms and one invented control platform; a rating for the control is a hallucination, and a rating with no source on the page is flagged; ratings shown on the page without markup are measured. This is one model's view: the web Brandscope and Reviewscope ask three assistants.
  • How generated the text sounds (2.109, at the bottom of the report, the same block as the Vectorscope skill's): the writing patterns' Pattern score, the words language models overuse counted beside it, and the agent's swap test, a style.json with one verdict per passage: specific, or generic enough to sit on a competitor's page with only the name changed, each with a verbatim quote the script checks. Passage by passage, the findings behind every point, and where the script and the agent disagree, both. Not an authorship detector, and the block says so.
  • Languages: the hreflang set (self-reference, x-default, valid codes), each alternate fetched to see whether it links back, titles and descriptions compared across versions, html lang against hreflang, and the same facts: the numbers in each h2 section compared across the versions, aligned by section order.

How generated the text sounds, at the bottom of the report:

How generated the text sounds: the Pattern score, the marker words, the swap test and every passage

The same report in Ukrainian, in Polish, and with the dark theme switched on:

AI Search Audit Lite in Ukrainian

AI Search Audit Lite in Polish

AI Search Audit Lite, dark theme

MDN's HTTP caching guide fills all 26 passages; its last sections are merged into one zone on the map, and the links past the check's cap are counted. This run has only the swap test as a judgement file, so its brand card reads not checked:

AI Search Audit Lite on MDN's HTTP caching guide

Every technical detail sits in one closed section, All checks and technical details: the status by area, the bots in robots.txt, the 35 Google and Bing checks, the language versions, the severity chart, where the text goes, the <head> and robots.txt rows, every finding, and what the check cannot see:

AI Search Audit Lite with All checks and technical details opened

Examples

Real runs in examples/, each as HTML, Markdown and JSON. The three getlooploop.com pages were read on 17 October 2026 (the same readings as the Full skill's example) and audited again on 18 October with 2.109; MDN and example.com were read on 18 October. The judgement files are in examples/ai-search-audit-judgement/, written by Claude Code running Claude Opus 5.5, which did not know GetLoopLoop from its own knowledge and said so. The reports open in English; add #lang=uk or #lang=pl to the address, or press UK or PL at the top.

PageProblemsBlocking · high · medium · lowHow generated it soundsAmong them
getlooploop.com/vectorscope190 · 0 · 7 · 12score 1, no marker words, 4 of 17 passages generictwo links to 404 pages, two questions no passage answers, the brand unknown to the model, How-to markup with no rich result any more
getlooploop.com/uk/vectorscope210 · 0 · 8 · 13score 2, no marker words, 4 of 17 genericthe same, in Ukrainian; the keyword read off the h1 reaches 2 of 15 sections; the same facts in all three versions
getlooploop.com/es/vectorscope180 · 0 · 7 · 11score 0, no marker words, 5 of 18 genericthe same, in Spanish
MDN: HTTP caching100 · 0 · 3 · 7score 0, 0.2 marker words per 1,000, 5 of 26 generic (the intro, the footer, link lists)6 passages start mid-sentence, images without width and height, no Organization markup (only the swap test as judgement)
example.com240 · 0 · 6 · 18no score (under 60 words), 1 of 2 genericno h1; http and www variants that do not redirect; three questions with no answer; RFC 2606 and IANA, which the model knows and the page does not say

Install

The skill needs Node 20.18 or later (what cheerio declares) and no npm install: dist/read.mjs and dist/audit.mjs are the scripts bundled with everything they need, the Vectorscope's code and cheerio included, so plain node runs them, including in sandboxes that block package installs. The readable source is in scripts/; dist/ is generated from it (see Working on the skill).

As a plugin (Claude Code, Google Antigravity), recommended

Claude Code:

/plugin marketplace add Olexan-art/ai-search-audit
/plugin install ai-search-audit@ai-search-audit

Then turn on auto-update: in /plugin, open Marketplaces, select ai-search-audit and choose Enable auto-update. Until you do, Claude Code keeps the version you installed; /plugin marketplace update ai-search-audit fetches a newer one by hand.

Google Antigravity:

agy plugin install https://github.com/Olexan-art/ai-search-audit

To update, run the same command again.

The repository carries plugin.json and .agents/plugins/marketplace.json for Antigravity, gemini-extension.json for Gemini CLI and .claude-plugin/ for Claude Code, with the skill in skills/.

Update check and usage count

Once a day, the audit script (audit.mjs) asks getlooploop.com for the newest version of the skill. The request carries the skill's name, its version, the operating system (win, mac, linux or other), a random install id kept in ~/.getlooploop-skills/install-id, and how many times the skill ran since the last check. That is all: nothing about the pages you audit, no URL, no domain, no user or machine name. The server derives a country from the connection and does not store the IP address. If getlooploop.com does not answer, the skill reads the version from this repository's plugin.json on GitHub instead. A newer version is one line in the terminal and a notice in the report; the agent offers the update command and runs it only if you agree. Switch it off with GETLOOPLOOP_NO_UPDATE_CHECK=1 or --no-update-check (it is also off when CI is set): then nothing is sent and no id is created. Delete the ~/.getlooploop-skills folder to reset the id.

One command (Claude Code, Codex, opencode and other agents)

npx skills add Olexan-art/ai-search-audit --skill ai-search-audit

The skills CLI asks which agents to install it for (add -g for every project, -a claude-code to pick one). Then ask your agent for "an AI search audit of ". Later, npx skills update ai-search-audit brings it up to date.

claude.ai and the Claude desktop app

  1. Download ai-search-audit.zip from this repository's Releases. Do not unpack it: the upload takes the zip, with the ai-search-audit folder at its top level.
  2. Turn on code execution: Settings > Capabilities > Code execution and file creation. Skills need it. On Team and Enterprise plans an owner controls it in Organization settings > Capabilities.
  3. Upload the skill: Customize > Skills, the + button, Create skill, Upload a skill, then choose ai-search-audit.zip. Turn the skill on in the list. In the desktop app it is the same Customize > Skills page.
  4. Let the sandbox reach the site you want audited. A run fetches the page, its robots.txt, its sitemap and a few more URLs on the same site (see Network below), and code execution only reaches what its network setting allows. On Free, Pro and Max plans that setting is Allow network egress in Settings > Capabilities. On Team and Enterprise plans an owner chooses in Organization settings > Capabilities: off, package managers only (the default), package managers and specific domains, or all domains. With specific domains, add the site's domain. If your plan's setting cannot allow the site, the run stops with REFUSED unreachable and says so; use Claude Code or another agent on your own machine for that page.
  5. In a chat, ask: "an AI search audit of https://example.com/page". Claude runs the scripts in the sandbox and hands you report.html as a file to download and open in your browser, with a short summary in the chat.

Pasting this repository's link into a chat is not installing it. Claude rightly will not download and run code from a link someone pasted, so it will at best imitate the steps by hand, and those findings are not the tool's. Upload the zip as above.

Sources: Use skills in Claude · Create custom skills (the zip's structure, the name and description rules) · Create and edit files with Claude (code execution and network egress) · Agent Skills overview (network access by surface). Checked on 2026-10-16.

No internet in the agent's shell

Cloud sandboxes often run the shell offline or keep the skills folder read-only. The skill does not need a skills folder: unpack ai-search-audit.zip anywhere the agent can write and point it at that folder (or attach the zip to the chat). With no network, save the page in your browser as "Web page, HTML only" and, if you can, its robots.txt, and give both files to the agent: the skill reads the saved copy and marks every check that needs the network as not checked.

Use it with your agent

The skill is one folder, skills/ai-search-audit, with a SKILL.md and Node scripts. Agents that support Agent Skills load it from their skills directory; any other agent can follow SKILL.md as instructions. Either way the agent runs the same scripts from the skill folder, and the agent's own model writes the judgement files. No install step: the agent runs node dist/read.mjs and node dist/audit.mjs as they are.

git clone https://github.com/Olexan-art/ai-search-audit.git

Claude Code loads personal skills from ~/.claude/skills/ and project skills from .claude/skills/.

mkdir -p ~/.claude/skills && cp -r ai-search-audit/skills/ai-search-audit ~/.claude/skills/ai-search-audit   # personal
mkdir -p .claude/skills && cp -r ai-search-audit/skills/ai-search-audit .claude/skills/ai-search-audit       # or: this project only
New-Item -ItemType Directory -Force "$HOME\.claude\skills" | Out-Null
Copy-Item -Recurse ai-search-audit\skills\ai-search-audit "$HOME\.claude\skills\ai-search-audit"

OpenAI Codex (CLI, IDE extension, app) scans $HOME/.agents/skills (user) and .agents/skills in every directory from where you launch it up to the repository root (repo).

mkdir -p ~/.agents/skills && cp -r ai-search-audit/skills/ai-search-audit ~/.agents/skills/ai-search-audit
New-Item -ItemType Directory -Force "$HOME\.agents\skills" | Out-Null
Copy-Item -Recurse ai-search-audit\skills\ai-search-audit "$HOME\.agents\skills\ai-search-audit"

opencode loads global skills from ~/.config/opencode/skills/, ~/.claude/skills/ or ~/.agents/skills/, and project skills from .opencode/skills/, .claude/skills/ or .agents/skills/ (walking up to the git worktree). A copy made for Claude Code or Codex above is picked up too; otherwise:

mkdir -p ~/.config/opencode/skills && cp -r ai-search-audit/skills/ai-search-audit ~/.config/opencode/skills/ai-search-audit

Hermes Agent (Nous Research) keeps skills in ~/.hermes/skills/. It also reads <project>/.hermes/skills/ and <project>/.agents/skills/ once you trust the repository with hermes skills trust, and any folder listed under skills.external_dirs in ~/.hermes/config.yaml (for example ~/.agents/skills).

mkdir -p ~/.hermes/skills && cp -r ai-search-audit/skills/ai-search-audit ~/.hermes/skills/ai-search-audit

GLM (Zhipu / Z.ai) is a model family, not an agent. Z.ai's GLM Coding Plan is used inside coding agents; its documentation names Claude Code, Codex and OpenCode among them. Install the skill for the agent you run GLM in, as above: the skill then follows that agent's instructions, and GLM writes the judgement files.

Any other agent: point its instructions file (for example AGENTS.md) at the skill, so it reads the steps before it starts:

For an SEO or AI search audit of a web page, follow skills/ai-search-audit/SKILL.md and run its scripts from that folder.

Sources: Claude Code skills · Codex skills (where developers.openai.com/codex/skills now redirects) · opencode skills · Hermes Agent skills · Z.ai GLM Coding Plan. Checked on 2026-10-16; skill folders move between releases, so check your agent's current docs if a path does not work.

A run, in short

node dist/read.mjs https://example.com/page --out reading.json
# the agent writes questions.json, style.json, brand.json and reviews.json (schemas in references/judgement-files.md)
node dist/audit.mjs reading.json --questions questions.json --style style.json --brand brand.json --reviews reviews.json --check
node 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

Network: the page, its site's /robots.txt, its sitemap (up to 5 child sitemaps), /llms.txt, up to 10 hreflang alternates, a HEAD (or a bodiless GET) of up to 30 links and up to 20 images, and one request each for the favicon, a path that does not exist, up to 3 URL variants, a Markdown copy, ai-catalog.json and a declared IndexNow key, plus one DNS lookup of the host and one RDAP request to rdap.org with the site's IP address; every cap is reported, never silent. Outputs: a self-contained report.html (EN | UK | PL, light by default with a dark theme on the toggle, no network requests, every piece of page text escaped), report.md, and export.json (getlooploop/ai-search-audit@1), which carries every finding with its layer, its evidence and the rules it rests on.

This skill's own thresholds

Where no published rule gives a number: a title of 10 to 60 characters; a description of 70 to 160; a redirect chain from 2 hops; a heavy page over 500 KB of HTML or 30 scripts; a hard-to-read URL over 115 characters, with more than 2 query parameters or a session id; hidden headings counted high from 200 hidden characters; an image over 200 KB, and a srcset expected over 100 KB or from 800 px; above the fold as the hero and its first two images; data-nosnippet over half the visible text; an app shell under 150 words; stock phrases medium from 2.0 per 1,000 words; a keyword that misses half the passage text or reaches a third of the sections or fewer. Each is a named constant in skills/ai-search-audit/scripts/lib/ with its reason.

What it cannot see

  • Field data and the index. Core Web Vitals and "in Google's index" are always not measured: they need real-user data and Search Console, which a script on one page does not have.
  • What assistants answer today. The brand and review readings are one model's knowledge at its training time. They are not a measurement of any assistant and say nothing about what ChatGPT, Gemini or Perplexity answer now.
  • Pages behind a login. One public URL per audit, no cookies and no login. The guard refuses localhost, private and link-local addresses, hostnames without a dot, non-standard ports and anything that is not http(s), on every redirect hop (5 at most).
  • Language picked by location. No Accept-Language is sent, so a site that chooses a language by where the request comes from may answer in another one.
  • JavaScript, unless asked. The reading takes the HTML as served. --render, or an empty app shell, also runs the page in a local Chrome, Chromium or Edge; with no browser, the report says the render was not done.
  • Everything past a cap. 30 links and 20 images are checked, 10 language versions read, 5 child sitemaps opened; 12 seconds per request (8 for a link, an image or a site check), 3,000,000 characters of HTML. Every cap is named in the report, never applied silently.
  • How to fix it. Lite lists problems; it does not explain them or propose fixes.

Asking your agent

A short request with a clear stop condition and a reply format keeps the agent on the procedure. Copy it and put in your page:

## Task
Run the skill `ai-search-audit` on 1 page. Write its report.

## Inputs
- Page: https://example.com/page
- Working folder: ./audit
- Keyword (optional): your target phrase

## Rules
1. Read SKILL.md first. Follow it exactly.
2. Write every file into the working folder.

## Stop condition
Stop when report.html exists. Show it.

## Reply format
- The report link or path.
- The verdict line and the 3 biggest problems.

Working on the skill

scripts/ is the source; dist/ is built from it with esbuild (a development dependency only) and committed, so the skill runs without an install.

cd skills/ai-search-audit
npm install          # cheerio, and esbuild for the build
npm run build        # scripts/ → dist/read.mjs, dist/audit.mjs (not minified, readable)
npm test             # includes a check that dist/ equals a fresh build, so dist never drifts from scripts/
npm run zip          # dist-release/ai-search-audit.zip, the file to attach to a release for claude.ai

Change scripts/, run npm run build, and commit both.

The reading, the robots.txt rules, the passage checks and the language detection are the Vectorscope skill's code, vendored unchanged in skills/ai-search-audit/scripts/vendor/vectorscope/, which records the source commit and version. To update it, change it in getlooploop-skills, then copy it again from a checkout: node tools/sync-vectorscope.mjs <path to getlooploop-skills>, followed by npm run build && npm test.

Related

  • getlooploop-skills: the Vectorscope skill, which shows how an AI retriever reads a page passage by passage, and compares 2 to 5 pages. This skill's AI reading is its code.
  • GetLoopLoop and the web Vectorscope.

Licence

Free to use, change and share under the MIT licence, see LICENSE. If you publish something built with it, a link to getlooploop.com is appreciated.