geo-audit（地理审计）

GEO Audit + Autofix — Claude Code Skill

Audit any website for AI search visibility, then fix the problems.

Commands

Audit Flow

Phase 1: Discovery (Sequential)

1. Fetch the homepage using the fetch_page.py script: python3 ${CLAUDE_SKILL_DIR}/scripts/fetch_page.py <url> full This returns JSON with page data (including language, viewport, OG tags, Twitter Cards, meta robots, X-Robots-Tag, hreflang links, response time, detected platform), robots.txt, llms.txt status, RSL 1.0 licensing status, and sitemap pages (with lastmod data). The script auto-detects 15 platforms: WordPress, WordPress.com, Squarespace, Wix, Shopify, Webflow, Ghost, Framer, HubSpot, Carrd, Weebly, Drupal, Next.js, Vercel, Netlify. The detected_platform field will be {"name": "Squarespace", "slug": "squarespace"} or null. When a platform is detected, tailor all fix recommendations to that platform. For example:

- WordPress: "Install Yoast SEO plugin, go to Yoast SEO > Tools > File Editor" - Squarespace: "Go to Settings > Website > SEO > robots.txt" - Wix: "Go to Dashboard > Marketing & SEO > SEO Tools > robots.txt Editor" - Shopify: "Go to Online Store > Themes > Edit Code > robots.txt.liquid"

1. Detect business type from the fetched data:

- SaaS — pricing page, "sign up", "free trial", API docs, /app, /dashboard - E-commerce — product pages, cart, "add to cart", price elements - Local Service — phone number, address, "near me", Google Maps embed - Publisher — blog, articles, bylines, publication dates - Agency — portfolio, case studies, "our services", client logos - Other — default, apply general GEO best practices

1. Identify key pages from sitemap (up to 50 pages). Prioritize:

- Homepage - About/team pages - Top content/blog pages - Product/service pages - Pricing page (if SaaS)

Phase 2: Parallel Analysis (12 Subagents)

Launch these 12 analyses simultaneously using the Agent tool:

Agent 1: AI Citability + Content Quality

• Run citability scorer on up to 10 key pages: python3 ${CLAUDE_SKILL_DIR}/scripts/citability.py <url>
• Score passages for AI citation readiness (answer block quality, self-containment, statistical density, structural readability, uniqueness signals)
• Detects comparison patterns, how-to/instructional content, cause-effect language
• Enhanced self-containment: flags dangling pronouns, context-reference phrases
• Page-level findings: data table detection (>500 word pages), heading structure validation (>50% blocks without headings)
• FAQ/featured snippet detection: Q&A headings, definition patterns ("What is..."), numbered step lists, direct answer quality
• Conversational language scoring: second-person pronouns, questions, contractions, natural connectors
• Entity recognition: named entity density, defined terms, bold/emphasis terms
• Semantic topic coverage: heading diversity, subtopic breadth, repetitive heading detection
• LLM chunk-size scoring: checks if passages match ~500-token retrieval window sizes (optimal 225-415 words), flags oversized blocks (>600 words)
• Keyword density check: extracts primary keyword from title/H1, flags stuffing (>3%) or underuse (<0.5%)
• Flesch readability scoring: pure-Python Flesch Reading Ease calculator. Content in the 60-75 range receives up to 31% more AI citations. Flags very difficult (<30) and moderate (30-50) readability.
• Multi-modal content detection: checks for video embeds, interactive elements (canvas, SVG, calculators), audio, data visualizations. Multi-modal content sees up to 156% higher AI selection rates.

Agent 2: AI Crawler Access + llms.txt Quality

• Parse robots.txt results from Phase 1 fetch data
• Check status of these AI crawlers: GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, anthropic-ai, PerplexityBot, CCBot, Bytespider, cohere-ai, cohere-training-data-crawler, Google-Extended, GoogleOther, GoogleOther-Image, GoogleOther-Video, Applebot-Extended, FacebookBot, Amazonbot, Meta-ExternalAgent, Meta-ExternalFetcher, YouBot, AI2Bot, Ai2Bot-Dolma, Diffbot, ImagesiftBot, aiHitBot, DuckAssistBot, img2dataset, MyCentralAIScraperBot, omgili, omgilibot, Quora-Bot, TikTokSpider
• Run llms.txt quality scorer: python3 ${CLAUDE_SKILL_DIR}/scripts/llms_txt.py <url>
• Scores existence, content length, markdown structure (# title, > description, ## sections), key pages section, about section, markdown link format, llms-full.txt bonus
• Checks for RSL 1.0 (Really Simple Licensing) at /.well-known/rsl.json — a machine-readable AI licensing standard backed by Reddit, Yahoo, Medium, Quora, Cloudflare, Akamai, and Creative Commons
• If llms.txt is missing, draft one based on site structure

Agent 3: Schema / Structured Data

• Run schema checker on homepage and key pages: python3 ${CLAUDE_SKILL_DIR}/scripts/schema_check.py <url>
• Detect existing JSON-LD, microdata, RDFa
• Validate against schema.org specs including @context validation (missing or non-standard @context)
• Identify missing schema types based on business type
• Generate correct JSON-LD for missing schemas using templates in templates/schema/

Agent 4: Technical SEO (13 checks)

• Run the technical SEO analyzer: python3 ${CLAUDE_SKILL_DIR}/scripts/technical_seo.py <url> Or pipe Phase 1 data: python3 ${CLAUDE_SKILL_DIR}/scripts/fetch_page.py <url> full | python3 ${CLAUDE_SKILL_DIR}/scripts/technical_seo.py
• 18 weighted checks: HTTPS, redirect chain, canonical tag, H1 count, meta description length, word count (300+), security headers, SSR detection, viewport meta, language attribute, Open Graph tags, response time (<3s), image alt text coverage (80%+), Twitter Cards, URL structure (length, clean characters), image optimization (width/height for CLS, lazy loading, srcset), hreflang validation, structured data in JS detection
• Optionally check Core Web Vitals via PageSpeed Insights: https://www.googleapis.com/pagespeedonline/v5/runPagespeed?url=<url>&strategy=mobile

Agent 5: E-E-A-T Analysis (9 signals)

• Run the E-E-A-T analyzer: python3 ${CLAUDE_SKILL_DIR}/scripts/eeat.py <url> Or pipe Phase 1 data: python3 ${CLAUDE_SKILL_DIR}/scripts/fetch_page.py <url> full | python3 ${CLAUDE_SKILL_DIR}/scripts/eeat.py
• 9 weighted signals: author bylines, publication dates, citation language, credentials, contact information, about page link, trust indicators, first-party expertise, schema.org authorship
• Each signal has a weight; score = (earned weight / total weight) * 100

Agent 6: Brand Presence

• Run the brand presence analyzer: python3 ${CLAUDE_SKILL_DIR}/scripts/brand_presence.py <url> Or pipe Phase 1 data: python3 ${CLAUDE_SKILL_DIR}/scripts/fetch_page.py <url> full | python3 ${CLAUDE_SKILL_DIR}/scripts/brand_presence.py
• Social media links (10 platforms: Twitter/X, LinkedIn, GitHub, YouTube, Wikipedia, Instagram, Facebook, TikTok, Reddit, Crunchbase)
• Review site links (5 platforms: G2, Capterra, Trustpilot, Product Hunt, Yelp)
• Brand pages (/about, /press, /media, /newsroom, /news)
• Schema.org sameAs links in structured data

Agent 7: Platform Readiness (composite)

• Run after all other dimension scores are available: echo '{"schema":72,"eeat":65,"citability":80,"llms_txt":45,"technical_seo":90,"brand_presence":55}' \ | python3 ${CLAUDE_SKILL_DIR}/scripts/platform_readiness.py
• Computes per-platform readiness:

- Google AI Overviews: schema 50% + eeat 50% - ChatGPT: citability 50% + llms_txt 50% - Claude: citability 40% + llms_txt 30% + eeat 30% - Perplexity: technical_seo 50% + citability 50% - Gemini: schema 40% + brand_presence 30% + eeat 30%

• Overall score = average of all 5 platform scores

Agent 8: Brand Mentions (Web Search)

• Search for the brand/site name across:

- Reddit (site:reddit.com "") - YouTube (site:youtube.com "") - Wikipedia mentions - LinkedIn presence - GitHub presence (if relevant) - Industry directories

Agent 9: Search Intent Classification

• Run the search intent classifier: python3 ${CLAUDE_SKILL_DIR}/scripts/search_intent.py <url> Or pipe Phase 1 data: python3 ${CLAUDE_SKILL_DIR}/scripts/fetch_page.py <url> full | python3 ${CLAUDE_SKILL_DIR}/scripts/search_intent.py
• Classifies page into intent categories: informational, commercial, transactional, navigational
• Detects page type: article, product, pricing, comparison, service, homepage, about, support, general
• Checks intent-page type alignment — mismatches hurt AI citation
• Intent-specific recommendations (e.g., question headings for informational, comparison tables for commercial)
• Signals detected: URL structure, heading patterns, content patterns, CTA language, price elements

Agent 10: Content Freshness

• Run the content freshness scorer: python3 ${CLAUDE_SKILL_DIR}/scripts/content_freshness.py <url> Or pipe Phase 1 data: python3 ${CLAUDE_SKILL_DIR}/scripts/fetch_page.py <url> full | python3 ${CLAUDE_SKILL_DIR}/scripts/content_freshness.py
• Parses dates from 5 sources: meta tags (article:published_time, article:modified_time), JSON-LD (datePublished, dateModified), <time> elements, visible text ("Published on...", "Updated..."), HTTP Last-Modified header
• Calculates content age in days from the most recent date
• Scoring: <30 days = 100, <90 days = 85, <180 days = 70, <365 days = 55, <2 years = 35, older = 15
• Flags: no dates found, stale content (>1 year), published date without modified date, no visible date on page

Agent 11: Internal Link Structure

• Run the internal link analyzer: python3 ${CLAUDE_SKILL_DIR}/scripts/internal_links.py <url> Or pipe Phase 1 data (includes sitemap for coverage check): python3 ${CLAUDE_SKILL_DIR}/scripts/fetch_page.py <url> full | python3 ${CLAUDE_SKILL_DIR}/scripts/internal_links.py
• Link density: internal links per 1000 words (target: 5-15)
• Anchor text quality: flags generic anchors ("click here", "read more", empty text)
• Link diversity: ratio of unique destinations to total links
• Hub pattern detection: checks if page links to related content in the same section (hub-and-spoke topology)
• Sitemap coverage: how many sitemap pages are reachable via internal links from this page
• Sitemap validation: URL count limits (50k max), lastmod presence and accuracy, stale sitemap entries (>2 years)

Agent 12: Hreflang / International SEO

• Run the hreflang validator: python3 ${CLAUDE_SKILL_DIR}/scripts/hreflang.py <url> Or pipe Phase 1 data: python3 ${CLAUDE_SKILL_DIR}/scripts/fetch_page.py <url> full | python3 ${CLAUDE_SKILL_DIR}/scripts/hreflang.py
• If no hreflang tags: neutral pass (not every site needs them)
• If hreflang tags exist, validates 5 checks: self-referencing tag, x-default presence, valid ISO 639-1 language codes, consistent protocol (http/https), canonical URL alignment
• Weighted scoring like technical_seo.py

Phase 3: Report

Present findings grouped by category. For each category:

• Current state (what exists now)
• Issues found (with severity: Critical / High / Medium / Low)
• Confidence label for each finding:

- Confirmed — directly verified (e.g., missing HTTPS, no H1 tag, absent robots.txt entry) - Likely — based on heuristic analysis (e.g., SSR detection, regex-based E-E-A-T signals, composite platform scores) - Hypothesis — suggested improvement, may not apply to every site (e.g., missing data tables, FAQ content suggestions)

• Specific findings with evidence

Do NOT compute a weighted composite score. Report each category independently. Let the user interpret priorities.

Categories to report:

1. AI Citability
2. AI Crawler Access
3. llms.txt & RSL Status
4. Structured Data / Schema
5. Technical SEO
6. Content Quality & E-E-A-T
7. Brand Presence
8. Platform Readiness
9. Search Intent
10. Content Freshness
11. Internal Link Structure
12. Hreflang / International SEO

Phase 4: Autofix

Auto-generate (write immediately to geo-fixes/ in current working directory):

• geo-fixes/llms.txt — ready-to-deploy llms.txt file
• geo-fixes/llms-full.txt — detailed version
• geo-fixes/schema/ — JSON-LD files for each missing schema type
• geo-fixes/robots-txt-additions.txt — recommended robots.txt additions for AI crawlers
• geo-fixes/meta-tags.html — optimized meta tag snippets
• geo-fixes/FINDINGS.md — full audit report in markdown

Recommend for approval (present to user, do NOT implement until approved):

• Heading structure changes (H1/H2/H3 hierarchy fixes)
• Meta description rewrites
• Content restructuring for citability (passage length, answer blocks)
• Alt text additions for images
• Internal linking improvements
• Any changes to existing source code files

AI-Powered Suggestions

Use your intelligence as an LLM to generate contextual, site-specific suggestions beyond what the scripts can detect. These go in the report and fix files.

1. Meta Description Rewrites

For every page with a missing, too-short, or too-long meta description (from technical_seo.py findings):

• Read the page title, H1, and first 200 words of body content
• Write an optimized meta description (120-160 characters) that:

- Includes the primary topic/keyword naturally - Uses active voice and a clear value proposition - Is compelling enough to earn a click from both search results and AI citations

• Write the result to geo-fixes/meta-tags.html as a ready-to-paste <meta name="description"> tag with a comment noting which page it's for

2. Alt Text Generation

For every image flagged as missing alt text (from technical_seo.py):

• Examine the image src filename, surrounding text context, and page topic
• Write descriptive, specific alt text (10-20 words) that:

- Describes what the image shows, not just the topic - Includes relevant keywords naturally (no keyword stuffing) - Avoids starting with "Image of" or "Photo of"

• Present as a table: | Image src | Suggested alt text |

3. Heading Restructure

If the page has no H1, multiple H1s, or poor heading structure (from technical_seo.py and citability.py):

• Analyze the page content and propose a complete heading hierarchy:

- One H1 that captures the primary topic - H2s for each major subtopic (aim for 4-8 on content-heavy pages) - H3s for sub-sections where appropriate

• Frame at least 1-2 headings as questions (for FAQ/snippet eligibility)
• Present as: Current heading structure → Proposed heading structure

4. FAQ Section Drafting

If the page has no FAQ/featured-snippet-ready content (from citability.py citability-no-faq-content):

• Identify 3-5 questions that the page content implicitly answers
• Write each as a Q&A pair:

- Question: Natural language question a user would ask - Answer: 2-3 sentence direct answer, starting with a declarative statement

• Generate matching FAQPage JSON-LD schema and write to geo-fixes/schema/faq-page.json
• The FAQ content itself goes in the report as a recommendation (not auto-applied)

5. Content Gap Analysis

Based on the page topic, business type, and existing content:

• Identify 3-5 subtopics or questions the page does NOT cover but should
• For each gap:

- Missing topic: What's not covered - Why it matters: How this gap affects AI citability (e.g., "ChatGPT frequently gets asked about X but your page doesn't address it") - Suggested addition: A 1-2 sentence description of what to add

• Consider the business type (SaaS, e-commerce, local, publisher, agency) when identifying gaps — e.g., a SaaS page missing pricing comparisons, or a local business missing service area information

6. Robots.txt Crawler Explainer

For every blocked AI crawler found in the robots.txt analysis:

• Explain what the crawler does and who operates it
• Explain the tradeoff of allowing vs. blocking it:

- If allowed: What happens (e.g., "Your content may appear in ChatGPT responses with citation links back to your site") - If blocked: What you lose (e.g., "Your content won't be indexed by OpenAI's search product")

• Flag any crawlers that are blocked but probably shouldn't be (e.g., blocking GPTBot but wanting ChatGPT visibility)
• Flag any crawlers that are allowed but the user might want to block (e.g., training-only crawlers like cohere-training-data-crawler or img2dataset)

Present this as a table: | Crawler | Operator | Status | What it does | Recommendation |

Content rewrite suggestions — For the 3-5 weakest citability passages (lowest scores from citability.py), generate a rewritten version that:

• Starts with a clear declarative sentence (5-30 words) that directly answers the implied question
• Stays within the optimal 134-167 word passage length
• Replaces dangling pronouns with proper nouns
• Adds statistical density where appropriate (numbers, percentages, named sources)
• Structures as Q&A, definition, or how-to format when the topic allows

Present each rewrite as:

• Original (first 50 words + score)
• Suggested rewrite (full passage)
• Why (which citability factors improve)

Present code-level recommendations as a numbered list with:

• File path (if in a repo context)
• Current state
• Recommended change
• Why it helps AI visibility

Wait for user to approve specific items before making any code edits.

Script Reference

fetch_page.py

python3 ${CLAUDE_SKILL_DIR}/scripts/fetch_page.py <url> [mode]

Modes: page, robots, llms, sitemap, full

Returns JSON to stdout. Page mode includes: language attribute, viewport meta, Open Graph tags, Twitter Cards, hreflang links, meta robots, X-Robots-Tag header, response time (ms), enhanced image data (width, height, srcset, sizes, fetchpriority), and checks 32 AI crawlers. Full mode also includes RSL 1.0 status and enhanced sitemap data with lastmod.

citability.py

python3 ${CLAUDE_SKILL_DIR}/scripts/citability.py <url>

Returns JSON with per-passage citability scores, page-level summary, and structural findings (data table detection, heading structure validation, FAQ/snippet detection, conversational language, entity density, semantic topic coverage, LLM chunk-size scoring, keyword density, Flesch readability scoring, multi-modal content detection).

schema_check.py

python3 ${CLAUDE_SKILL_DIR}/scripts/schema_check.py <url>

Returns JSON with detected schemas, validation results (including @context validation), and missing schema recommendations.

eeat.py

python3 ${CLAUDE_SKILL_DIR}/scripts/eeat.py <url>

Or pipe from fetch_page.py:

python3 ${CLAUDE_SKILL_DIR}/scripts/fetch_page.py <url> full | python3 ${CLAUDE_SKILL_DIR}/scripts/eeat.py

Returns JSON with 9 E-E-A-T signals, per-signal weights, overall score, grade, and findings.

brand_presence.py

python3 ${CLAUDE_SKILL_DIR}/scripts/brand_presence.py <url>

Or pipe from fetch_page.py:

python3 ${CLAUDE_SKILL_DIR}/scripts/fetch_page.py <url> full | python3 ${CLAUDE_SKILL_DIR}/scripts/brand_presence.py

Returns JSON with social platform presence (10), review platforms (5), brand pages, sameAs detection, score, and findings.

technical_seo.py

python3 ${CLAUDE_SKILL_DIR}/scripts/technical_seo.py <url>

Or pipe from fetch_page.py:

python3 ${CLAUDE_SKILL_DIR}/scripts/fetch_page.py <url> full | python3 ${CLAUDE_SKILL_DIR}/scripts/technical_seo.py

Returns JSON with 18 weighted technical checks, per-check pass/fail, overall score, grade, and findings.

llms_txt.py

python3 ${CLAUDE_SKILL_DIR}/scripts/llms_txt.py <url>

Or pipe from fetch_page.py:

python3 ${CLAUDE_SKILL_DIR}/scripts/fetch_page.py <url> full | python3 ${CLAUDE_SKILL_DIR}/scripts/llms_txt.py

Returns JSON with llms.txt quality score, spec compliance checks, and findings.

platform_readiness.py

python3 ${CLAUDE_SKILL_DIR}/scripts/platform_readiness.py <schema> <eeat> <citability> <llms_txt> <tech_seo> <brand>

Or pipe dimension scores as JSON:

echo '{"schema":72,"eeat":65,"citability":80,"llms_txt":45,"technical_seo":90,"brand_presence":55}' \
  | python3 ${CLAUDE_SKILL_DIR}/scripts/platform_readiness.py

Returns JSON with per-platform readiness scores (Google AI Overviews, ChatGPT, Claude, Perplexity, Gemini), recommendations, and overall score.

search_intent.py

python3 ${CLAUDE_SKILL_DIR}/scripts/search_intent.py <url>

Or pipe from fetch_page.py:

python3 ${CLAUDE_SKILL_DIR}/scripts/fetch_page.py <url> full | python3 ${CLAUDE_SKILL_DIR}/scripts/search_intent.py

Returns JSON with primary/secondary intent classification (informational, commercial, transactional, navigational), confidence level, detected page type, intent-page alignment check, and intent-specific findings.

content_freshness.py

python3 ${CLAUDE_SKILL_DIR}/scripts/content_freshness.py <url>

Or pipe from fetch_page.py:

python3 ${CLAUDE_SKILL_DIR}/scripts/fetch_page.py <url> full | python3 ${CLAUDE_SKILL_DIR}/scripts/content_freshness.py

Returns JSON with dates found (from meta tags, JSON-LD, <time> elements, visible text, HTTP headers), content age, freshness score, grade, and findings.

internal_links.py

python3 ${CLAUDE_SKILL_DIR}/scripts/internal_links.py <url>

Or pipe from fetch_page.py:

python3 ${CLAUDE_SKILL_DIR}/scripts/fetch_page.py <url> full | python3 ${CLAUDE_SKILL_DIR}/scripts/internal_links.py

Returns JSON with link density, anchor text quality, link diversity, hub pattern detection, sitemap coverage, sitemap validation (URL count, lastmod, stale entries), score, grade, and findings.

hreflang.py

python3 ${CLAUDE_SKILL_DIR}/scripts/hreflang.py <url>

Or pipe from fetch_page.py:

python3 ${CLAUDE_SKILL_DIR}/scripts/fetch_page.py <url> full | python3 ${CLAUDE_SKILL_DIR}/scripts/hreflang.py

Returns JSON with hreflang tag count, languages found, validation results (self-referencing, x-default, language codes, protocol consistency, canonical alignment), score, grade, and findings. Returns neutral score (100) if no hreflang tags are present.

Business Type Adaptations

Adjust recommendations based on detected type:

Quality Gates

• Max 50 pages crawled per audit
• 30-second timeout per page fetch
• 1-second delay between requests
• Always respect robots.txt
• Skip pages with >80% content similarity