Audit website AEO/GEO skill

Audit Website AEO/GEO Skill#

You audit a live website the way an AI agent would — crawling its pages, parsing structure, and judging whether the content is citation-worthy for ChatGPT, Claude, Perplexity, and Google AI Overviews.

The audit has two halves:

• Foundational (50%) — 16 deterministic pass/fail checks run by a script. Reproducible, no judgment.
• Intelligence (50%) — 6 content-quality dimensions you score by reading the pages, using the rubric below.

Final score = 0.5 × foundational + 0.5 × intelligence, mapped to an A-F grade.

This skill produces a diagnosis. To then fix a codebase, hand off to the improve-aeo-geo skill.

Workflow#

Follow this sequence exactly.

Step 1: Get inputs#

Ask the user for:

1. Website URL (required) — the live site to audit.
2. Crawl depth (optional) — how many pages to crawl. Default 10, max 30.
3. Output location (optional) — where to save the report. Default: current directory, or workspace/<customer-name>/ if working a customer project.

If the user already gave a URL when invoking the skill, don't re-ask — just confirm crawl depth and proceed.

Step 2: Run the deterministic audit#

Run the bundled script from this skill's scripts/ directory. It requires only Node 18+ — no npm install.

node <skill-path>/scripts/aeo-audit.mjs <url> --max-pages=10 --out=<output-dir>/aeo-audit.json

The script crawls (sitemap + robots.txt + internal links), runs the 16 checks per page, aggregates site-wide, and writes a JSON report. It also prints a human-readable summary. Tell the user the foundational score and the failed checks.

If the script errors (site unreachable, 0 pages crawled), report the error and stop — don't fabricate a score.

Step 3: Read the JSON report#

Read the aeo-audit.json file. The key fields:

• scoring.foundationalScore — the deterministic score (0-100). This is final — do not change it.
• checks — the 16 site-wide checks with pass/fail and details.
• pagesForReview — up to 5 representative pages (home + richest content pages), each with an aiView object containing title, metaDescription, h1, headings, schemaTypes, jsonLdSummary, textExcerpt, internalLinkCount, author, publishedDate, modifiedDate. Use these for Step 4.
• prioritizedFixes, worstPages, coverage, heuristicIntelligenceSignals — supporting context. The heuristic signals are a deterministic prior — a sanity check, not the real evaluation.

Step 4: Score the 6 intelligence dimensions#

You are an AI agent that just found this site via web search. A user asked you a question and you landed here. Decide: would you cite this site in your answer?

Read the textExcerpt, headings, and metadata of each page in pagesForReview. Then score all 6 dimensions below, each 0-5, using only what you actually observed (no assumptions about pages you didn't see). Write the rationale before the score.

Rubric (0-5 each)#

Answer Readiness — If a user asked a question about this site's topic, could you find a direct answer here? The #1 factor — content answering questions in the first paragraph gets 4.8x more citations.

• 0 = No answers; purely promotional or navigational
• 1 = Vague content that talks around topics but never directly answers
• 2 = Some answers exist but buried deep, not in opening paragraphs
• 3 = Several questions answerable; some definition-first or FAQ-style content
• 4 = Most common questions answerable; answers lead sections
• 5 = Exceptional (dedicated FAQ blocks, definition-first paragraphs, Q&A format throughout)

Quotability — Can you extract a clean, self-contained 40-60 word passage to quote? Comparison tables get 2.8x citations; FAQ blocks +156%.

• 0 = No extractable content (interactive-only, single dense block)
• 1 = Content requires full-page context; no passage stands alone
• 2 = A few passages extractable but most need surrounding context
• 3 = Several self-contained paragraphs; some lists or structured blocks
• 4 = Good quotability (tables, lists, FAQ sections, clear answer blocks)
• 5 = Highly quotable (comparison tables, step-by-step blocks, definition paragraphs throughout)

Evidence Density — Statistics, data points, named sources, in-text citations? Adding in-text citations = +115% visibility; statistics = +40% citation rate.

• 0 = No evidence; only marketing copy and vague claims
• 1 = Vague claims only ("best in class", "industry leading")
• 2 = Mostly generalities; rare specific data points
• 3 = Some statistics and named sources; cites a few external sources
• 4 = High density (numbers, dates, named sources, links to references)
• 5 = Exceptional (statistics every 150-200 words, in-text citations throughout, verifiable metrics)

Content Depth — Enough substance to thoroughly answer questions on the topic? Long-form (2000+ words) gets 3x more citations.

• 0 = Empty or placeholder content only
• 1 = Minimal (a few sentences, no real substance)
• 2 = Thin (surface-level, missing key details a user would need)
• 3 = Adequate (covers main points but lacks sub-topics or examples)
• 4 = Rich (comprehensive coverage, multiple sub-topics, examples, data)
• 5 = Exceptional (authoritative depth, multi-faceted, a go-to reference)

Freshness — Current enough to cite confidently? 76% of ChatGPT's most-cited pages were updated in the last 30 days.

• 0 = No date signals; content appears abandoned or timeless-generic
• 1 = Dates present but clearly outdated (2+ years, stale references)
• 2 = Moderately dated; no "last updated" indicator
• 3 = Reasonably current OR explicit "last updated" date visible
• 4 = Recent content with update timestamps and current references
• 5 = Clearly current (recent dates, active maintenance evident)

Structural Clarity — Does the HTML parse cleanly into readable text? A prerequisite — clean heading hierarchy = 3.2x more citations.

• 0 = Unreadable (no text, blocked, non-semantic markup)
• 1 = Very poor (walls of text, no headings, topic unclear)
• 2 = Weak (some structure but confusing or inconsistent headings)
• 3 = Adequate (clear headings and paragraphs, topic identifiable)
• 4 = Good (clean H1-H2-H3 hierarchy, scannable, purpose obvious)
• 5 = Excellent (perfect heading outline, semantic HTML, zero noise)

For each dimension, record: a 1-2 sentence rationale, the 0-5 score, and a one-line key finding (under 14 words).

Step 5: Compute the final score#

1. Intelligence score = average(6 dimension scores) × 20 → rounds each 0-5 to 0-100.
2. Final score = round(0.5 × foundationalScore + 0.5 × intelligenceScore).
3. Grade from the final score:

Sanity-check your intelligence score against heuristicIntelligenceSignals in the JSON. If they diverge by more than ~25 points on any dimension, re-read that page's excerpt and confirm your score is grounded in observed content.

Step 6: Write the audit report#

Write a Markdown report to <output-dir>/aeo_audit_report.md using the format in Report Format below. Then summarize for the user: the grade, the 3 highest-impact fixes, and a one-line recommendation.

Step 7: Hand off#

If the user wants to act on the findings:

• To fix a codebase → recommend the improve-aeo-geo skill, passing this report as input.
• To re-measure after fixes → re-run this skill on the same URL and compare scores.

The 16 deterministic checks#

Run by the script. For reference (id — what it verifies — points):

A site-wide check passes when 80%+ of crawled pages pass it (the script handles aggregation). Foundational score = earned points ÷ 142 × 100.

Report Format#

# AEO/GEO Audit — [domain]

**Audited:** [date] · **Pages crawled:** [N]

## Score

| | Score | |
|---|---|---|
| Foundational (16 checks) | XX/100 | |
| Intelligence (6 dimensions) | XX/100 | |
| **Final** | **XX/100** | **Grade: X** |

[One-sentence verdict on AI-citation readiness.]

## Foundational Checks

[Table of the 16 checks: ✓/✗, label, detail. Group failures at the top.]

## Intelligence Evaluation

For each of the 6 dimensions: score (X/5 → XX/100), rationale, key finding.

## Prioritized Fixes

Numbered list, highest impact first. For each: what to change, why it matters,
impact/effort. Pull from `prioritizedFixes` and your dimension findings.

## Weakest Pages

[From `worstPages` — URL and per-page %.]

## Recommendation

[2-3 sentences: biggest opportunity, and whether to run improve-aeo-geo next.]

Rules#

• Never fabricate the crawl. Always run the script. If it fails, report the failure — don't invent pages or scores.
• The foundational score is the script's output. Don't recompute or adjust it.
• Score intelligence only from observed content. Base every dimension score on textExcerpt / headings / metadata in pagesForReview. No assumptions about unseen pages.
• Rationale before score. Write why, then the number — for every dimension.
• One report file, saved to the output directory. Don't scatter partial outputs.
• This skill diagnoses; it does not edit code. Code fixes are the job of improve-aeo-geo.

Research References#

All statistics above are from verifiable primary research:

Supporting file: examples/anthropic-com-audit.md#

AEO/GEO Audit — anthropic.com#

Audited: 2026-05-16 · Pages crawled: 6

Example output produced by audit-website-aeo. Real crawl data; the intelligence scores below were assigned by reading the page excerpts.

Score#

The site is technically clean and well-written, but missing structured-data and freshness signals keep it from being an easy, confident citation source.

Foundational Checks#

Intelligence Evaluation#

Answer Readiness — 3/5 (60/100). Product and guidance pages lead with clean definitions ("Anthropic Interviewer is a research tool, powered by Claude, that…"); the homepage is promotional and navigational. Several questions answerable, but no FAQ-first structure across the site.

Quotability — 4/5 (80/100). Self-contained passages and concrete metrics ("Claude Haiku 4.5 scores 73.3% on SWE-bench Verified") are easy to extract; two pages carry FAQPage structure. Strong — clear answer blocks and lists.

Evidence Density — 3/5 (60/100). Product pages cite hard numbers and sample sizes ("1,250 professionals", "81,000 people"); value/careers pages run on qualitative copy. No author attribution anywhere. Uneven — strong on product pages, thin elsewhere.

Content Depth — 3/5 (60/100). Every page clears the depth threshold, but the crawl surfaced only home + miscellaneous pages — no docs or blog corpus. Adequate; lacks a deep knowledge section.

Freshness — 3/5 (60/100). Some pages show visible recent dates ("Last updated Jul 10, 2025"), but these are inline text, not machine-readable article:modified_time meta tags — and there is no RSS feed. Current content, weak machine-readable freshness signals.

Structural Clarity — 3/5 (60/100). Single H1 and clean titles throughout, but heading hierarchy passes only 2/6 pages and the homepage extract is noisy with repeated nav text. Readable, but the outline does not chunk cleanly.

Prioritized Fixes#

1. Add JSON-LD structured data to all key templates (High impact / Medium effort) — only 2/6 pages carry any schema. Add Organization + WebSite site-wide, Article to content pages, Product/SoftwareApplication to model pages. Agents rely on this to classify entities.
2. Publish a valid llms.txt (Medium / Low) — none exists. A heading + curated links gives AI systems a trusted index of the site.
3. Emit machine-readable date meta tags (Medium / Low) — convert visible "Last updated" dates into article:published_time / article:modified_time. Freshness is a top citation driver.
4. Fix heading hierarchy (Medium / Low) — 4/6 pages skip levels or use one. Enforce H1 → H2 → H3 so passages chunk cleanly.
5. Add an RSS/Atom feed (Medium / Low) — helps AI systems discover new announcements and releases.

Weakest Pages#

Recommendation#

The biggest, lowest-effort win is structured data — it fails on two-thirds of pages and is a hard prerequisite for AI classification. Combined with llms.txt and machine-readable dates, these fixes would lift the foundational score into the 80s. Run improve-aeo-geo against the site's codebase to apply them, then re-run this audit to confirm the delta.

Supporting file: README.md#

Audit Website AEO/GEO#

An agent skill that audits a live website for AI-engine discoverability — the way ChatGPT, Claude, Perplexity, and Google AI Overviews actually see it.

It crawls the site, runs 16 deterministic checks, then evaluates 6 content-quality dimensions and produces a single A-F score with prioritized fixes.

This skill diagnoses. To then fix a codebase, hand the report to improve-aeo-geo (opens in a new tab).

Install#

Clone the repo, then symlink or copy audit-website-aeo/ into ~/.codex/skills/ or ~/.claude/skills/. See the root README (opens in a new tab) for examples.

Requires Node 18+ for the crawler script (uses the built-in fetch). No npm install — the script is zero-dependency.

Usage#

/audit-website-aeo

Or:

"Audit my website for AEO" "Check if AI engines can cite https://example.com (opens in a new tab)" "Run an AEO/GEO audit on [URL]"

The skill asks for a URL and crawl depth, runs the audit, and writes a report.

How it works#

1. Crawl — Discovers pages via sitemap.xml, robots.txt, and internal links (up to 30 pages), respecting robots.txt.
2. Deterministic audit — A Node script runs 16 binary checks per page and aggregates site-wide (a check passes when 80%+ of pages pass it).
3. Intelligence evaluation — The agent reads the richest pages and scores 6 dimensions 0-5 using a research-backed rubric.
4. Scoring — Final = 50% foundational + 50% intelligence, mapped to an A-F grade.
5. Report — A Markdown report with the score, failed checks, dimension findings, prioritized fixes, and weakest pages.

The 16 deterministic checks#

Title · meta description · canonical URL · single H1 · structured data (JSON-LD) · recognized schema types · Open Graph · internal linking · image alt coverage · text depth · indexability · AI-accessible meta tags · heading hierarchy · llms.txt · AI-bot access in robots.txt · RSS/Atom feed.

The 6 intelligence dimensions#

The crawler script#

scripts/aeo-audit.mjs is a standalone, zero-dependency Node crawler and checker. The skill runs it for you, but you can also run it directly:

node scripts/aeo-audit.mjs https://example.com --max-pages=20 --out=aeo-audit.json

It prints a human-readable summary and writes a JSON report containing the foundational score, all 16 checks, prioritized fixes, weakest pages, and pagesForReview (the pages the agent reads for the intelligence evaluation).

Parsing note: the script parses HTML with regex/string ops, not a full DOM, so it has no dependencies but is slightly less precise than a browser. Checks measure presence and structure, which tolerates this well.

Where it fits in the workflow#

audit-website-aeo   →   improve-aeo-geo   →   audit-website-aeo (re-run)
   diagnose              fix the codebase        measure the delta

Run it before improve-aeo-geo to get a baseline, and again afterward to confirm the score improved.

License#

MIT — see the root LICENSE (opens in a new tab).