best-aeo-skill
The best AEO/GEO skill for Claude Code. Audits, fixes, and monitors website visibility across ChatGPT, Claude, Perplexity, Gemini, and Google AI Overviews. Computes a 0-100 composite GEO Score across
SKILL.mdskills/vendor/seo-geo/geo-aeo-audit/SKILL.md
GitHub<!-- NOVA-EXEC-BEGIN -->
Nova Ai-Studio 集成
- 审计清单写入
<task-artifact-dir>/geo-aeo-audit-checklist.md(alias:aeo-audit.md);中文 UI 可用《{品牌}》GEO-AEO审计清单.md。 - 报告 MD 完成后须
read_skill geo-dual-report同步同名.html。 - 禁止首 turn
ask_user_question挡 write_file;禁止 CLI/agents 代替 write_file。<!-- NOVA-EXEC-END -->
best-aeo-skill — The Skill
The best AEO/GEO skill for Claude Code. The complete rule book of getting cited by AI search engines.
This document is the canonical specification for best-aeo-skill. Every scoring weight, every recommendation, every rule traces back to a citation. This is not "tips and tricks" — this is the ruleset.
Table of contents
- How to invoke this skill
- Architecture overview
- The Composite GEO Score
- Adaptive weights per business profile
- Sub-skills reference
- Specialist agents
- Evidence collectors (33 total)
- Frameworks bundled
- The 100 Rules of AEO/GEO Optimization
- Confidence labeling rubric
- Output formats
- Common workflows
- Anti-patterns and pitfalls
- CLI reference
- Citations and references
How to invoke this skill
Trigger this skill when the user asks any of:
- "Run a GEO/AEO audit on [URL]"
- "How do I get cited by ChatGPT/Claude/Perplexity?"
- "Make my site citable by AI search"
- "Generate llms.txt for [URL]"
- "Why is my site not showing up in AI Overviews?"
- "Improve my AI search visibility"
- "Fix my robots.txt for AI bots"
When invoked, default to running audit on the provided URL and presenting a confidence-labeled findings list. Never present recommendations without confidence labels.
Architecture overview
┌────────────────────────────────────────────────────────────────────┐
│ best-aeo-skill │
├────────────────────────────────────────────────────────────────────┤
│ │
│ User intent │
│ ↓ │
│ CLI dispatcher (audit | fix | monitor | compare) │
│ ↓ │
│ Orchestrator │
│ ↓ │
│ ┌─────────────┬─────────────┬─────────────┬─────────────┐ │
│ │ Technical │ Citability │ Schema │ Entity │ │
│ │ agent │ agent │ agent │ agent │ │
│ └─────┬───────┴─────┬───────┴─────┬───────┴─────┬───────┘ │
│ ↓ ↓ ↓ ↓ │
│ 33 evidence collectors (parallel execution) │
│ ↓ ↓ ↓ ↓ │
│ ┌───────────────────────────────────────────────────────┐ │
│ │ Composite Scorer (4-vector weighted sum) │ │
│ │ Profile-adaptive weights · Confidence labeling │ │
│ └────────────────────────────┬───────────────────────────┘ │
│ ↓ │
│ Ranked findings + projected score impact + auto-fix paths │
│ │
└────────────────────────────────────────────────────────────────────┘
The Composite GEO Score
A single 0–100 number computed as a weighted sum of four vectors. Default weights:
| Vector | Default weight | Maps to | Why |
|---|---|---|---|
| Technical Accessibility | 20% | robots.txt, AI bot allowance, JS rendering, CDN access | Princeton: "if crawlers can't reach you, prose doesn't matter" |
| Content Citability | 35% | Statistic density, expert quotes, citations, freshness, fluency | Largest weight. Princeton: +40% from stats, +41% from quotes, +115% from citations |
| Structured Data | 20% | FAQPage, Article, Organization, HowTo, Speakable, Product | FAQPage = highest single-signal citation rate (empirical 41M-citation study) |
| Entity & Brand Signals | 25% | Author credentials, KG links, NAP, ai.txt | Sustained citation requires entity presence, not just one-off content quality |
Score bands
| Band | Range | Meaning | Default action |
|---|---|---|---|
| Excellent | 86–100 | Cited frequently across all engines | Maintain freshness; monitor only |
| Good | 68–85 | Cited regularly; gaps to fix | Apply top-3 fixes |
| Foundation | 36–67 | Indexed but rarely cited | Run full audit + apply all fixes |
| Critical | 0–35 | Effectively invisible to AI search | Start with Technical + Schema, then content |
Interpretation rules
- A score below 36 always indicates a technical or schema problem, not a content problem. Fixing prose first is the wrong order.
- A score 68–85 is the typical post-fix state for a well-built site. Reaching 86+ requires sustained content investment.
- Scores fluctuate ±3 due to crawl variance. Treat changes within ±3 as noise.
Adaptive weights per business profile
Business profiles re-weight the four vectors:
| Profile | Technical | Citability | Schema | Entity |
|---|---|---|---|---|
| Default | 20% | 35% | 20% | 25% |
| SaaS | 18% | 32% | 25% | 25% |
| E-commerce | 18% | 25% | 32% | 25% |
| Publisher / News | 15% | 45% | 20% | 20% |
| Local business | 18% | 25% | 22% | 35% |
| Agency / consultancy | 22% | 35% | 18% | 25% |
| DevTools / API docs | 25% | 30% | 25% | 20% |
| Academic / research | 18% | 50% | 12% | 20% |
Why these weights
- SaaS weights Schema higher because product/pricing pages benefit massively from FAQPage and Product markup.
- E-commerce weights Schema highest (Product, AggregateRating) because comparison queries are the primary citation surface.
- Publishers weight Citability highest because freshness + author E-E-A-T dominate news citation.
- Local weights Entity highest because LocalBusiness, NAP consistency, and GBP linkage decide local AI Overview citations.
- Academic weights Citability extremely high because formal citation patterns (statistics, source attribution, expert quotes) compound multiplicatively in academic AI search.
Pass --profile saas (or other) to switch. Defaults to default if unspecified.
Sub-skills reference
The skill exposes 7 independently invokable sub-skills.
audit
Purpose: Compute the GEO Score. Return ranked findings with confidence labels.
Inputs:
--url <URL>(required, single page)--sitemap <URL>(alternative, batch up to 100 pages)--profile [default|saas|ecommerce|publisher|local|agency|devtools|academic]--format [json|markdown|html|sarif|junit]
Outputs:
- Composite GEO Score (0–100)
- Per-vector breakdown
- Ranked findings (severity × confidence)
- Projected score impact per fix
- Recommended next sub-skill
Example:
bestaeo audit --url https://yoursite.com --profile saas --format markdown
fix-content
Purpose: Rewrite content for citability. Idempotent.
What it adds:
- 2–4 expert quotes per ~1000 words (Princeton: +41% citation likelihood)
- 1 statistic per ~200 words target density (Princeton: +40% visibility)
- Inline citations for every numeric claim (+115% citation likelihood when sources emphasized)
- "Last updated" ISO-8601 dateModified (3.2× more citations for content <30 days)
What it preserves:
- Voice (analyzes tone before rewriting)
- Existing structure
- All factual claims (never invents)
- Original markdown / HTML formatting
Inputs:
--url <URL>or--file <path>--apply(write changes; without flag = preview only)--max-quotes-per-1000-words 4--statistic-density 0.5(per 100 words)
fix-schema
Purpose: Generate JSON-LD. Validate. Inject.
Schemas supported (priority order):
- FAQPage — highest AI citation surface; auto-detects Q&A patterns in content
- Article + author credentials + dateModified
- Organization + sameAs links to KG sources
- HowTo — for tutorials and dev docs
- Speakable — for AI voice/read-aloud (Google AI Overviews preference signal)
- Product + AggregateRating + offers
- BreadcrumbList
- Person (for author bios)
- VideoObject + Clip + SeekToAction (for video content)
- LocalBusiness + GeoCoordinates (for local profile)
Validation:
- Schema.org compliance check
- Google Rich Results structured data eligibility
- Minimum required fields verification
- Cross-schema consistency (e.g., Article author matches Person markup)
fix-llmstxt
Purpose: Generate llmstxt.org-compliant /llms.txt.
Structure produced:
# {Site Name}
> {1-3 sentence summary}
{Optional 2-paragraph context}
## Primary pages
- [Title](URL): {summary}
- ...
## Optional reference
- [Title](URL): {summary}
Heuristics:
- Pulls H1 + meta description for site-level header
- Categorizes pages by URL pattern (home, product, docs, blog, support)
- Excludes pages with
noindexor password-protected - Caps at 50 primary + 50 optional URLs
- Generates summaries via passage extraction, not LLM hallucination
fix-robotstxt
Purpose: Patch robots.txt to explicitly allow AI bots.
Bots allowed by default (27):
| Provider | User-agent |
|---|---|
| Googlebot, Google-Extended, GoogleOther | |
| OpenAI | GPTBot, ChatGPT-User, OAI-SearchBot |
| Anthropic | ClaudeBot, anthropic-ai, Claude-Web, Claude-User, Claude-SearchBot |
| Perplexity | PerplexityBot, Perplexity-User |
| Apple | Applebot, Applebot-Extended |
| Meta | FacebookBot, Meta-ExternalAgent |
| You.com | YouBot |
| Cohere | cohere-ai |
| Mistral | MistralAI-User |
| Common Crawl | CCBot |
| ByteDance | Bytespider |
| Diffbot | Diffbot |
| Amazon | Amazonbot |
| DuckDuckGo | DuckDuckBot |
| Yandex | YandexBot |
| Bing | Bingbot |
Diff-aware: never overwrites existing rules. Adds explicit Allow: lines only where missing.
compare
Purpose: GEO Score head-to-head vs 2–5 competitors.
Outputs:
- Per-vector delta table
- Per-finding "they have, you don't" list
- Per-finding "you have, they don't" list
- Recommended action sequence to overtake
Example:
bestaeo compare --you https://yoursite.com --them https://competitor1.com,https://competitor2.com
monitor
Purpose: Track GEO Score over time.
Features:
- Stores history in
.bestaeo/history.jsonor remote (S3, GCS, Postgres) - Webhooks on score drop > N points
- Slack/Discord/email alerts
- Content decay detection (flags articles >30 days old with declining score)
- CI/CD gate: fails build if score drops below threshold
Specialist agents
The skill delegates to 5 specialist agents in parallel during full audits.
Technical agent
Owns: Technical Accessibility vector (20%) Evidence collectors used: robots_check, ai_bot_access, js_render, cdn_blocking, response_codes, sitemap_check, http2_check, mobile_render, lazyload_check Output: Technical findings ranked by severity
Citability agent
Owns: Content Citability vector (35%) Evidence collectors used: statistic_density, quote_extractor, citation_check, freshness_check, readability, passage_score, fluency_check, hedge_density, claim_verifier Output: Citability score + content rewrite recommendations
Schema agent
Owns: Structured Data vector (20%) Evidence collectors used: schema_validate, faq_check, article_check, jsonld_lint, speakable_check, product_check, breadcrumb_check, video_check Output: Schema gaps + ready-to-paste JSON-LD blocks
Entity agent
Owns: Entity & Brand Signals vector (25%) Evidence collectors used: entity_extractor, author_check, knowledge_graph, nap_consistency, brand_signal, sameas_links, expertise_signals Output: Entity gap analysis + author markup recommendations
Monitor agent
Owns: Longitudinal tracking Evidence collectors used: all (re-runs full audit at intervals) Output: Regression alerts, decay flags, threshold gates
Evidence collectors (33 total)
Every finding traces back to one or more of these. Output is what determines confidence label.
Technical (9)
- robots_check — parses robots.txt; verifies User-agent rules
- ai_bot_access — tests fetch as GPTBot, ClaudeBot, PerplexityBot
- js_render — compares HTML before/after JS execution
- cdn_blocking — detects Cloudflare/Akamai bot challenges
- response_codes — 200/301/404/500 distribution
- sitemap_check — XML sitemap presence + freshness
- http2_check — HTTP/2 or HTTP/3 support
- mobile_render — viewport, touch targets, mobile-first
- lazyload_check — confirms above-fold content not lazy-loaded
Citability (10)
- statistic_density — counts numeric claims per 100 words
- quote_extractor — counts and validates quoted passages
- citation_check — counts inline links to external sources
- freshness_check — extracts dateModified, computes content age
- readability — Flesch-Kincaid + sentence variance
- passage_score — passage-level extractability (chunk-readiness for RAG)
- fluency_check — detects awkward AI-rewrite patterns
- hedge_density — counts hedging language ("may", "might", "could")
- claim_verifier — flags factual claims without citation
- rag_chunk_score — 250-word chunks ready for retrieval
Schema (7)
- schema_validate — Schema.org compliance
- faq_check — FAQPage presence + structure
- article_check — Article required fields
- jsonld_lint — JSON-LD syntax validity
- speakable_check — Speakable Spec markup
- product_check — Product + AggregateRating + offers
- breadcrumb_check — BreadcrumbList navigation
Entity (7)
- entity_extractor — NER for Person, Org, Place, Product
- author_check — author markup + credentials
- knowledge_graph — sameAs links to Wikidata, Wikipedia
- nap_consistency — Name/Address/Phone matching across surfaces
- brand_signal — brand mention frequency in content
- sameas_links — verifies all sameAs URLs resolve
- expertise_signals — credentials, affiliations, bylines
Frameworks bundled
Reference frameworks shipped inside the skill.
CORE-EEAT (80 items)
Content quality publish-gate. Adopted from aaron-he-zhu/seo-geo-claude-skills.
- Experience (20): firsthand examples, real screenshots, original photos, dated experiences, personal voice, specific use cases, real client examples, original tests, hands-on insights, primary research
- Expertise (20): author credentials, named contributors, technical depth, specific terminology, comparison to alternatives, methodology disclosure, source code, primary research links, technical accuracy, edge case coverage
- Authority (20): citing primary sources, linking to authoritative domains, credential markup, organization markup, awards/recognition, press mentions, peer-reviewed citations, industry reports, government/academic sources, sameAs links to KG
- Trust (20): HTTPS, contact info, privacy policy, terms, transparent author bio, last-updated dates, fact-check links, error corrections published, disclosure of affiliate links, no AI-generated boilerplate
CITE (40 items)
Domain authority benchmark.
- Backlinks (10): editorial inbound links, anchor diversity, follow ratio, dofollow share, referring domain growth, link velocity, contextual placement, partner cross-links, .edu/.gov links, brand mention links
- Authority (10): topical authority depth, content cluster coverage, internal linking density, hub-and-spoke architecture, freshness average, archival URL stability, canonical consistency, topic exclusivity, expert contribution, podcast/video presence
- Brand (10): branded search volume, social mention velocity, Wikipedia entry, founder authority, NPS proxy signals, review aggregator presence, podcast guest appearances, conference speaker activity, owned media (newsletter), retention community
- Trust (10): historical domain age, ownership transparency, organizational footprint, regulatory compliance pages, third-party audits referenced, certifications, customer logos, case studies with metrics, press kit, investor information
Princeton GEO Tactics (9 methods, +40% visibility)
From Aggarwal et al. KDD 2024.
- Cite sources — emphasize external citations (+115% citation likelihood)
- Add statistics — numeric claims with sources (+40% visibility)
- Add expert quotes — attributed quotations with quotation marks (+41%)
- Improve fluency — natural language, not formulaic
- Authority signaling — credential markup, named contributors
- Easy-to-understand — Flesch-Kincaid optimization (target grade 8-10)
- Keyword stuffing — DO NOT (negative impact in modern engines)
- Topic relevance — focused on single primary topic per page
- Length optimization — content appropriate to topic (not arbitrary minimum)
Confidence Rubric (3 labels)
Anti-hallucination labeling.
- Confirmed — directly observed by an evidence collector. Example:
parse_html.pyreturned no<title>tag → "Missing title tag" is Confirmed. - Likely — strong inference from multiple signals. Example: schema_validate returned no FAQPage AND quote_extractor found Q&A patterns → "FAQPage schema would help" is Likely.
- Hypothesis — needs human confirmation. Example: LLM judgment on tone match, or speculative future-proofing recommendation.
Every finding presented to the user MUST carry exactly one label.
The 100 Rules of AEO/GEO Optimization
Organized by category. Each rule is research-validated, traceable to a citation, and actionable.
Category A: AI Crawler Access (Rules 1–10)
Rule 1: Allow GPTBot in robots.txt explicitly. Implicit Allow is not enough; some CDN configurations block by default. Confidence: Confirmed.
Rule 2: Allow ClaudeBot, Claude-Web, anthropic-ai, Claude-User, Claude-SearchBot — Anthropic uses multiple user-agents for different surfaces.
Rule 3: Allow PerplexityBot AND Perplexity-User. Perplexity-User is for real-time user fetches; PerplexityBot is for indexing.
Rule 4: Allow Google-Extended (separate from Googlebot). Without it, Bard/Gemini training and grounding signals are blocked while traditional search still indexes.
Rule 5: Allow CCBot (Common Crawl). Many LLMs train on Common Crawl; blocking CCBot reduces presence in training data.
Rule 6: Verify each Allow with a fetch-as-bot test. robots.txt rules don't guarantee CDN compliance.
Rule 7: Cloudflare bot management can block AI bots even with robots.txt Allow. Verify in Cloudflare dashboard "AI Bots" rule.
Rule 8: Akamai bot manager has separate AI bot category since Q3 2024. Verify allowance there.
Rule 9: Render JavaScript server-side or use SSR/static generation. AI bots have inconsistent JS execution; pure SPA = invisible.
Rule 10: Test indexability with Google Rich Results, Schema validator, AND a custom fetch-as-GPTBot to confirm.
Category B: Content Citability (Rules 11–35)
Rule 11: Statistic density target: 1 numeric claim per 200 words. Below threshold, citation likelihood drops linearly. Source: Princeton KDD 2024.
Rule 12: Every numeric claim must have an inline citation to a primary source. Uncited statistics are filtered out by Perplexity.
Rule 13: Expert quotes target: 2–4 attributed quotations per 1000 words. Use quotation marks. Princeton: +41% citation likelihood.
Rule 14: Quote attribution must include speaker name AND credential or affiliation. "Anonymous expert" patterns reduce citation rate.
Rule 15: Source emphasis (bold, citation, link) increases citation likelihood by +115%. Strongest single Princeton finding.
Rule 16: Freshness: content <30 days old receives 3.2× more citations than older content. Update timestamps actively, not just superficially.
Rule 17: "Last updated" date must be in machine-readable format: ISO-8601 dateModified in JSON-LD AND visible in page.
Rule 18: Content decay flags: declining citation despite stable rankings. Audit articles older than 90 days for refresh opportunity.
Rule 19: Readability target: Flesch-Kincaid grade 8–10. Higher (academic) loses general AI citations; lower (oversimplified) loses authority.
Rule 20: Sentence variance: mix short (<15 words) and medium (15–25 words). Monotonic length signals AI-generation patterns.
Rule 21: Avoid AI-rewrite tells: "It's important to note", "In conclusion", "Furthermore", "delve into". These signals reduce trust.
Rule 22: Hedging density: keep "may/might/could" under 1% of words. Excessive hedging reduces authoritative scoring.
Rule 23: First paragraph must contain the primary claim. Top-of-document passages dominate snippet extraction.
Rule 24: Use H2 + H3 hierarchy. Each heading is a passage anchor for AI extraction.
Rule 25: Maximum 80 words per paragraph. Longer paragraphs lose passage extractability.
Rule 26: Lists (bulleted, numbered) increase AI citation rate by +14% (empirical).
Rule 27: Tables with headers increase factual citation rate. Avoid pivot tables in AI-cited content.
Rule 28: Code blocks: use <pre><code> with language hint. Critical for dev-tool documentation visibility.
Rule 29: Avoid ALL CAPS — interpreted as low-quality signal.
Rule 30: Avoid emoji-heavy content for B2B/research surfaces. AI engines reduce weight for emoji-dense content in factual citations.
Rule 31: Use specific examples over general abstractions. "ChatGPT cited 47% of pages" beats "many users see results".
Rule 32: Quantify outcomes when possible: "47% increase" beats "significant increase".
Rule 33: Date your examples. "In 2025, X happened" beats "recently, X happened".
Rule 34: Original research, original screenshots, original tests outperform aggregation by 3–5×. Princeton: experience signals.
Rule 35: Topic exclusivity per page. Don't cover 5 topics on one page; create 5 pages each focused.
Category C: Structured Data (Rules 36–55)
Rule 36: FAQPage schema produces highest AI citation rate of any single signal. Add to every page with Q&A patterns.
Rule 37: FAQPage Q&A pairs: 5–10 ideal. <3 reduces signal weight; >15 dilutes.
Rule 38: FAQ Q must be a real question (with ?). "Pricing" as Q name fails validation.
Rule 39: FAQ A must be 30–200 words. Too short: low value; too long: not extractable.
Rule 40: Article schema required fields: headline, datePublished, dateModified, author (with Person markup), publisher.
Rule 41: Article author MUST have credentials, sameAs to LinkedIn/Wikipedia/etc. Anonymous authorship reduces citation rate by 60%+.
Rule 42: Organization schema required fields: name, url, logo, sameAs (Wikidata, Wikipedia, Crunchbase, LinkedIn).
Rule 43: Organization sameAs MUST resolve. Broken sameAs links reduce schema trust score.
Rule 44: HowTo schema: required for any tutorial. Steps as ItemList with HowToStep entities.
Rule 45: HowTo step images: improve citation rate. Use real screenshots, not stock illustrations.
Rule 46: Speakable schema for AI voice surfaces (Google AI Overviews voice). Mark sentences/headings most representative of content.
Rule 47: Product schema required: name, description, image, brand, offers (price + currency + availability), aggregateRating.
Rule 48: Product reviews must have actual review markup. Hidden reviews under cookie consent fail validation.
Rule 49: BreadcrumbList on every non-home page. Improves contextual citation.
Rule 50: VideoObject + Clip + SeekToAction for video content — required for video chapter citations in AI Overviews.
Rule 51: Don't ship invalid JSON-LD. One syntax error invalidates the entire block. Run schema_validate before deploy.
Rule 52: Don't duplicate schemas of the same type on one page. Multiple FAQPage blocks confuse parsers.
Rule 53: Cross-reference schema types: Article author Person markup must match Organization founder markup.
Rule 54: Use @id properties when entities are referenced multiple times on the same page.
Rule 55: Test final output in Google Rich Results AND Schema.org validator AND a custom JSON-LD parser.
Category D: Entity & Brand (Rules 56–75)
Rule 56: Author bios must include role, organization affiliation, years of experience, and 2+ credentials minimum.
Rule 57: Author profile pages with full Person schema increase article citation by +20%.
Rule 58: sameAs links from Person → LinkedIn, Twitter/X, GitHub, ORCID (for academics), Wikipedia (if applicable).
Rule 59: Organization Wikipedia entry: massive trust boost. If unavailable, focus on Wikidata entry creation.
Rule 60: Wikidata entity (Q-number) is more achievable than Wikipedia and provides similar trust signal.
Rule 61: NAP (Name/Address/Phone) consistency across all surfaces: website footer, GBP, social profiles, citations.
Rule 62: Inconsistent NAP reduces local AI Overview citation by 40%+.
Rule 63: Brand mention frequency: aim for 1 branded reference per 500 words in own content.
Rule 64: Backlink anchor diversity: brand-name anchors ideal 50–60% of total. Pure exact-match keywords trigger spam filters.
Rule 65: Press releases, podcast appearances, conference talks: each builds entity signals visible to LLMs.
Rule 66: ai.txt file at /.well-known/ai.txt. New standard. Include permissions + author attribution.
Rule 67: llms.txt file at root. While only 0.1% AI bots fetch it currently, it's near-zero cost and Anthropic does honor for ClaudeBot.
Rule 68: RSS/Atom feeds increase Common Crawl + AI training data presence. Maintain at least one feed.
Rule 69: OpenGraph + Twitter Card metadata on every page. Used for social citation extraction.
Rule 70: Schema.org Person profile pages for top 3 authors minimum. Solo bylines reduce trust unless founder.
Rule 71: Founder/CEO entity: must exist with verified profiles on LinkedIn + at least one industry-specific surface.
Rule 72: Trust signals: physical address, contact email, phone (when applicable), regulatory disclosures.
Rule 73: Customer logos with permission. Building entity authority via association.
Rule 74: Original research / annual reports / state-of-X reports: highest entity authority single-action.
Rule 75: Speaking engagements at named conferences: each adds entity signal. Document with sameAs links.
Category E: Multi-engine optimization (Rules 76–90)
Rule 76: ChatGPT favors authoritative long-form (1500+ words), consensus-based content, with explicit attribution.
Rule 77: ChatGPT prefers content with clear "this is the answer" framing. Hedged content gets cited less.
Rule 78: Claude (Anthropic) rewards precise attribution and factual density. Claude is most likely to cite multiple sources for one claim.
Rule 79: Claude tends to cite sources with active "Last updated" timestamps within 90 days.
Rule 80: Perplexity favors academic and news sources, heavy citation density, fresh content.
Rule 81: Perplexity's citation extraction prefers content with explicit [1] [2] [3] reference numbering or footnote-style.
Rule 82: Google AI Overviews lean on traditional SEO best practices PLUS direct-answer formatting.
Rule 83: AI Overviews reward FAQPage schema heavily. If your AI Overview presence is low, add FAQ schema first.
Rule 84: AI Overviews trigger on 25.11% of all Google searches in 2026. For local queries, the rate is 38%.
Rule 85: Gemini blends traditional Google ranking signals with AI-specific signals. Optimize for both.
Rule 86: Each engine's bot may need different signals. Test fetch-as for each: GPTBot, ClaudeBot, PerplexityBot.
Rule 87: Bing Copilot uses Bing's index. Bing's IndexNow protocol can ship updates within minutes.
Rule 88: Apple Intelligence uses Applebot-Extended. Apple's signals are still emerging in 2026.
Rule 89: ByteDance Bytespider — unclear policy. Allow if you want presence in Chinese AI surfaces.
Rule 90: Multi-engine targeting requires DIFFERENT optimizations weighted differently per engine. Use the --engine flag to weight per target.
Category F: Anti-patterns (Rules 91–100)
Rule 91: DO NOT keyword-stuff. AI engines penalize stuffing more aggressively than Google does.
Rule 92: DO NOT use AI-generated boilerplate. Detection is increasingly accurate; flagged content gets de-cited.
Rule 93: DO NOT block AI bots while expecting AI citations. Some sites block GPTBot then complain about no ChatGPT visibility.
Rule 94: DO NOT over-rely on llms.txt. As of 2026, only 0.1% of AI bot traffic fetches it. Useful but not foundational.
Rule 95: DO NOT generate FAQPage with synthetic Q&A that doesn't match real user questions. AI engines detect this and penalize.
Rule 96: DO NOT publish thin content for SEO velocity. AI engines weight quality over quantity even more than Google.
Rule 97: DO NOT remove or relocate URLs without 301s. Citation links break, citation history is lost.
Rule 98: DO NOT hide content behind cookie banners or modals at first paint. AI bots see what's rendered initially.
Rule 99: DO NOT use stock images without alt text and original captions. "Photo of person" is not a citation signal.
Rule 100: DO NOT auto-translate content with poor LLMs. Bad translations destroy citation rates in non-English markets.
Confidence labeling rubric
Every finding output by this skill MUST carry exactly one label.
Confirmed
Directly observed by ≥1 evidence collector. Example signals:
- HTTP response code present in trace
- DOM element absent in parsed HTML
- JSON-LD parser returned syntax error
- Header missing in HTTP response
- Author byline element not found
When in doubt about confidence: prefer Likely over Confirmed. Confirmed should be unambiguous.
Likely
Strong inference from ≥2 evidence collectors that agree, OR ≥1 collector with high specificity that maps to a known anti-pattern. Example signals:
- schema_validate returned no FAQPage AND quote_extractor found Q&A patterns
- statistic_density returned 0.2/100 AND citability_score ≤ 50
- Multiple authors mentioned but Person schema absent
- robots.txt allows GPTBot but CDN blocks GPTBot fetch
Hypothesis
LLM judgment, speculative recommendation, or single weak signal. Example signals:
- Tone analysis suggesting voice mismatch
- Recommendations for future-proofing emerging engines
- Recommendations relying on user-only-knowable context
NEVER present a Hypothesis as a fix path without flagging human review needed.
Output formats
The skill produces output in 5 formats simultaneously when requested.
Markdown (default)
Human-readable report with score, findings, and projected impact.
JSON
Machine-readable for automation. Schema:
{
"url": "https://example.com",
"geoScore": 64,
"scoreBand": "Foundation",
"vectors": {
"technical": 72,
"citability": 58,
"schema": 48,
"entity": 78
},
"findings": [
{
"id": "F-001",
"severity": "high",
"confidence": "Confirmed",
"category": "schema",
"rule": "Rule 36",
"message": "Missing FAQPage schema",
"evidence": ["schema_validate", "quote_extractor"],
"projectedImpact": 18,
"fixCommand": "bestaeo fix-schema --types faq"
}
]
}
HTML
Visual dashboard with interactive elements. Auto-opens in browser.
SARIF
Static Analysis Results Interchange Format. For GitHub Code Scanning integration.
JUnit
For CI/CD pipelines. Each finding as a test case; severity maps to pass/fail.
Common workflows
Workflow 1: First-time audit
bestaeo audit --url https://yoursite.com --profile saas --format markdown
Read top 3 fixes. Apply manually or via:
bestaeo fix --url https://yoursite.com --apply
Workflow 2: CI/CD gate
bestaeo audit --url $DEPLOY_URL --format junit > geo-report.xml
bestaeo monitor --url $DEPLOY_URL --threshold 80 --fail-on-drop
Workflow 3: Competitor analysis
bestaeo compare \
--you https://yoursite.com \
--them https://competitor1.com,https://competitor2.com \
--format markdown
Workflow 4: Site-wide audit
bestaeo audit \
--sitemap https://yoursite.com/sitemap.xml \
--max-urls 100 \
--profile publisher \
--format html
Workflow 5: Apply all fixes
bestaeo fix \
--url https://yoursite.com \
--apply \
--include content,schema,llmstxt,robotstxt
Anti-patterns and pitfalls
Don't fix prose first if technical/schema score is below 60
Prose rewriting is high-effort, low-leverage when foundations are missing. Fix Technical and Schema first.
Don't trust llms.txt as a foundational signal
0.1% of AI bots fetch it as of 2026. Generate it for the 0.1%, but don't over-weight in scoring.
Don't ship synthetic FAQ content
AI engines detect mass-generated Q&A. Use real user questions from search console, support tickets, or sales calls.
Don't remove canonical tags when consolidating URLs
Changing canonicals without 301s loses citation history.
Don't auto-fix on production without backup
fix --apply should be tested on staging first, then promoted.
Don't optimize for one engine in isolation
Optimizing only for ChatGPT loses Perplexity and Claude. Use multi-engine scoring profiles.
Don't ignore citation decay
A 90-day-old article with declining citation needs refresh, not "more backlinks".
CLI reference
Global flags
--profile— business profile (default, saas, ecommerce, publisher, local, agency, devtools, academic)--format— output format (json, markdown, html, sarif, junit)--output— write to file--engine— target a specific engine (chatgpt, claude, perplexity, gemini, ai-overviews, all)--verbose— full evidence dump--no-color— for piping--ci— CI/CD mode (machine-readable, exit codes)
Sub-commands
bestaeo audit— diagnosebestaeo fix— apply fixesbestaeo fix-content— content rewrites onlybestaeo fix-schema— schema generation onlybestaeo fix-llmstxt— llms.txt onlybestaeo fix-robotstxt— robots.txt onlybestaeo compare— head-to-headbestaeo monitor— track over timebestaeo init— initialize config in current directorybestaeo --versionbestaeo --help
Citations and references
This skill is built on peer-reviewed research and empirical analysis.
Primary research
Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., Deshpande, A. (2024). GEO: Generative Engine Optimization. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD '24). arxiv:2311.09735 · DOI
AutoGEO Authors (2026). AutoGEO: Automatic Optimization for Generative Engine Citation. International Conference on Learning Representations (ICLR 2026).
C-SEO Bench Authors (2025). Conversational SEO Benchmark: Empirical Analysis of Content Manipulation Effectiveness in AI Search.
Empirical analysis
- 199-biotechnologies team (2025). 41M-citation corpus analysis. Compiled from public AI search results and citation patterns.
- OtterlyAI (2026). llms.txt Adoption Study.
- SE Ranking (2026). llms.txt Audit of 300,000 Domains.
Industry data sources
- Superlines: AI Search Statistics 2026 — engine market share, AI Overview trigger rates.
- HubSpot: Answer Engine Optimization Case Studies — practitioner outcomes.
- Position.digital: 150+ AI SEO Statistics for 2026.
- Stackmatix: AI Search Market Share Q1 2026.
Open-source projects studied (predecessors)
- AgriciDaniel/claude-seo (5.8k ⭐) — extension architecture, parallel subagents
- aaron-he-zhu/seo-geo-claude-skills (1.4k ⭐) — CORE-EEAT (80) + CITE (40) frameworks
- Bhanunamikaze/Agentic-SEO-Skill (458 ⭐) — confidence labels, evidence collectors
- Auriti-Labs/geo-optimizer-skill (328 ⭐) — Princeton-backed methods, MCP server
- 199-biotechnologies/claude-skill-seo-geo-optimizer (28 ⭐) — IndexNow, freshness, entity extraction
Standards
- llmstxt.org — llms.txt specification
- Schema.org — structured data vocabulary
- Google Search Central — AI Overviews
License
MIT — use, fork, ship. Star ⭐ on GitHub if it helps.
Contact
- Website: bestaeoskill.com
- Repo: github.com/bestaeoskill/best-aeo-skill
- Issues: github.com/bestaeoskill/best-aeo-skill/issues
best-aeo-skill v1.0.0 · Released May 2026 The best AEO/GEO skill for Claude. Open source. Forever.