Web AI Vitals

What this skill does

• Scores 7 categories each 0–100 (crawl access, raw-HTML readability, structure & semantics, structured data, direct answer/GEO, AI guide files, trust & freshness)
• Explains how AI visibility actually happens: via live retrieval (RAG) — and checks the prerequisites for it (crawlable + indexed)
• Checks crawl access per bot, separated into search vs. training (OAI-SearchBot/ChatGPT-User/GPTBot, Claude-SearchBot/Claude-User/ClaudeBot, PerplexityBot, Google-Extended, Bingbot)
• Detects CDN/WAF blocks (Cloudflare, 403/429 challenges) — the most common cause of “AI can’t see my site”
• Rates raw-HTML readability: is the content there WITHOUT JavaScript? (AI bots mostly don’t render JS — client-only SPAs are often empty shells)
• Checks semantic HTML, heading hierarchy and chunkable sections for better RAG retrieval
• Checks structured data (Schema.org/JSON-LD: Organization, Article, FAQ, Product, sameAs) and its consistency with the visible content
• Rates the direct-answer/GEO format: citable 40–60-word blocks, question headings, FAQ, original data
• Checks AI guide files (llms.txt, ai.txt, sitemap with lastmod, canonical, knowledge.json) — with an honest take on llms.txt’s real-world value
• Fetches data itself where possible: raw HTML like an AI bot (curl with GPTBot user-agent), robots.txt, headers, guide files; otherwise analysis from supplied data + exact commands
• Gives a concrete one-sentence fix per finding + recommendation strength (mandatory/recommended/situational/optional) and can generate ready robots.txt/llms.txt/JSON-LD snippets
• Treats allow vs. block neutrally (lock training out, keep search open) — measured vs. estimated is flagged

Description

Web AI Vitals is a skill for AI visibility (GEO) — it checks across seven categories from 0–100 how well your site can be crawled, read, and cited by ChatGPT, Claude, and Perplexity, and names a concrete fix for each finding. A comprehensive AI-readability and GEO (Generative Engine Optimization) audit that checks whether AI answer engines can crawl, read in the raw HTML, structurally understand and cite a page — across seven categories scored 0–100. AI crawl access checks robots.txt per relevant bot (OpenAI: OAI-SearchBot/ChatGPT-User/GPTBot; Anthropic: Claude-SearchBot/Claude-User/ClaudeBot; Perplexity; Google: Googlebot/Google-Extended; Microsoft: Bingbot — ChatGPT retrieval leans heavily on the Bing index) plus CDN/WAF blocks (Cloudflare, 403/429 challenges) and accidental blocks. Raw-HTML readability checks whether the content is in the initial HTML WITHOUT JavaScript — the major AI crawlers mostly don’t render JS to keep costs down (Google is the exception). Structure & semantics rates a clean h1/h2/h3 hierarchy, semantic landmarks, lists/tables and chunkable, single-topic sections (RAG splits pages into chunks). Structured data checks Schema.org/JSON-LD (Organization, Article, FAQPage, Product, Person, Breadcrumb, sameAs) and its consistency with the visible content. Direct-answer format (GEO) rates citable 40–60-word blocks, question headings, FAQ, TL;DR and original data. AI guide files checks llms.txt (honestly framed: widely adopted but barely fetched by the major bots right now — low leverage), ai.txt, a sitemap with lastmod, canonical and supplementary knowledge files. Trust & freshness rates author/publisher identity, date/freshness signals (Perplexity weights recency), evidence, HTTPS and the indexing prerequisite (no live retrieval without an index). Visibility in AI answers comes via live retrieval (RAG); the skill checks exactly its prerequisites — crawlable AND indexed. Measured vs. estimated is always flagged, never invented. Allow vs. block (lock training out, keep search open) is a neutral strategic choice for the owner. Output: score table, bot-access matrix, findings per category with severity/fix/recommendation strength, prioritized actions and a realistic expectation. Authorized sites only, no scraping, no bypassing bot protection. Synergy: performance/security/classic SEO via Web Vitals, content briefings via the SEO Briefing, rendering code fixes via Code Reviewer Pro.

Examples

What it does not do

• Guarantees no citations or rankings — it audits a site’s AI readability, it is not a ranking tool
• Invents no visibility figures; without a real source clearly flagged as an estimate
• Scrapes no foreign content and bypasses no bot protection/paywalls; authorized sites only
• Pushes no allow/block strategy — locking training crawlers out is legitimate
• Replaces no real AI-visibility trackers, no Bing/Google Webmaster Tools and no legal review

Compatibility & tech

Tested (internal)

5 scenarios

Recommended runtime

Strongest available model; most effective with web/shell access for live fetching of the page, robots.txt and headers.

Modes

Full audit · Single category · “Why can’t ChatGPT see my site?” · File generator (robots.txt/llms.txt/JSON-LD) · Strategy advice (visible vs. block training) · Re-audit/comparison

Inputs

URL (+ optional web/shell access) · Raw HTML (view-source/curl) · robots.txt · HTTP header dump · existing llms.txt/ai.txt/sitemap.xml · optional: target AI, strategy, page type

Output format

Score table (7 categories à 0–100 + traffic light) with a bot-access matrix, findings per category (severity + concrete fix + recommendation strength), prioritized actions (impact × effort), a realistic expectation (timeline) and a stated data basis/limits

Subcategory

AI visibility & GEO

License

Proprietär

Security profile

What you get

• web-ai-vitals-1.7.0/8 files
  • .claude-plugin/marketplace.json
  • plugins/web-ai-vitals/skills/web-ai-vitals/8 files

    • SKILL.md
    • manifest.json
    • assets/1 file

      • ai-vitals-card-sample.svg
    • references/4 files

      • agent-discovery.md
      • bots-and-snippets.md
      • web-ai-vitals-audit-output.md
      • web-ai-vitals-checklist.md
    • scripts/1 file

      • render-ai-vitals-card.mjs
  • .agents/skills/web-ai-vitals/→ universal — same content (Codex, Cursor, Copilot, Gemini, Windsurf, Cline)
  • LICENSE.txt

Installation

Also works as a chat prompt

Note

Strongest with web/shell access (Claude Code, servers, CI): then the skill fetches raw HTML like an AI bot, robots.txt, headers and guide files itself. In plain chat, supplied HTML/robots.txt + a single prompt with copy-paste commands is enough. Off-site signals and foreign indexes (Bing/Google) can’t be fully checked from the page alone. Authorized sites only.

Changelog

• v1.7.001.07.2026Added the agentic-discovery layer (category 6 + a new reference): now also checks the Content-Signal robots line (ai-train/search/ai-input), Link headers + /.well-known/api-catalog (linkset to JSON feeds), an agent-skills/capability index, Markdown-for-agents (served on Accept: text/markdown), WebMCP (read-only tools via navigator.modelContext), an MCP server + server-card.json, and DNS-AID (SVCB). Honesty rule: never advertise an endpoint that resolves to nothing (a dead promise = red); serve discovery files no-cache. Weighted as Emerging with low weight — no ranking guarantee.
• v1.6.101.07.2026Fairer score-card colors: per-category color bands by realistic achievability (Direct-Answer Format, AI Guidance Files, Trust & Freshness are harder to max out) with a smooth red→amber→green blend instead of hard steps — weak scores still stay red; optional per-run `bands` override. Also: engine facts (bot names, crawler UAs, index behavior, model cutoffs) are now flagged as a dated snapshot — verify against the official provider docs and cite the retrieval date.
• v1.6.030.06.2026Operational audit upgrade (from a real test audit): every finding now carries an evidence level (Measured/Derived/Account-only/Estimated). New: an AI-engine matrix (a read per ChatGPT/Perplexity/Claude/Gemini/Copilot), a mandatory to-do backlog with priority, acceptance criterion, verification, effort & risk per task, a shareable SVG result graphic (a score card with date, the 7 categories and the top-3 — honest: weak scores stay red) plus a bundled render script, and a handoff step at the end: a copy-paste prompt for another AI tool, a developer/CMS implementation plan, or direct implementation (only with access). Plus a console-assisted mode for GSC/Bing exports.
• v1.5.025.06.2026Platform-native output: now actively uses your platform's native formatting — tables, checklists, code blocks and clickable/button options, with a numbered fallback on non-interactive platforms. Plus internal metadata that prepares the upcoming agent builder.
• v1.4.022.06.2026Initial release: AI-visibility/GEO audit across 7 categories (0–100) — AI crawl access (robots.txt per search/training bot), raw-HTML readability (AI bots mostly don’t render JS), structure & semantics, structured data (Schema.org/JSON-LD), direct-answer format (GEO), AI guide files (llms.txt/ai.txt/sitemap/knowledge) and trust/freshness (E-E-A-T, Bing index, citations). From web research (2026): AI bots mostly skip JS, major bots barely fetch llms.txt, ChatGPT retrieval via the Bing index, robots.txt rules per bot, GEO direct answers, E-E-A-T/freshness. measured-vs-estimated methodology and live-retrieval prerequisites (crawlable + indexed), auto data collection with shell access, file generator (robots.txt/llms.txt/JSON-LD); authorized sites only, no scraping.
• v1.3.018.06.2026Auf AUTHORING-v2-Standard überarbeitet: realistische Eingabe/Antwort-Beispiele (inkl. Kalibrierungsfall), Copy-Checkliste(n) und interne Evals; Cross-Verweise ergänzt.
• v1.2.018.06.2026Reworked to the AUTHORING v2 standard: realistic input/answer examples (incl. a calibration case), copy checklist(s) and internal evals; cross-references added.
• v1.1.011.06.2026Added references/bots-and-snippets.md: deeper detail reference via progressive disclosure; SKILL.md stays lean.
• v1.0.011.06.2026Added from deeper web research: AI reading order (title→meta description ≤160 chars→H1→first 2–3 sentences→H2 structure, blockquotes as summaries); Structure category expanded with a <main> landmark, Markdown cleanliness (HTML→Markdown is token-efficient) and internal linking (~4–6 topical links/page → deeper crawling); GEO expanded with semantic completeness & query fan-out (the strongest single lever for citations) and self-contained citable passages; Trust/Index expanded with classic search rank as a prerequisite (Google AIO = normal index, ChatGPT = Bing) and authority/brand mentions (~⅓ of citation likelihood); llms.txt framing refined (OpenAI/Microsoft partly fetch llms.txt/llms-full.txt, Google doesn’t per Mueller; token budget <200K).

Frequently asked questions