Claude Agent Readiness Skill

A GEO (Generative Engine Optimization) audit skill for Claude. Where SEO optimizes for traditional search-engine rankings, GEO optimizes for how AI systems discover, parse, and cite content. This skill is built on the published GEO/AEO literature and platform documentation as of mid-2026 and treats hype and evidence as separate categories.

A Claude skill that audits websites and code repositories for how well they are discoverable, fetchable, parsable, and citable by AI search systems (ChatGPT, Claude, Perplexity, Google AI Mode, Bing Copilot) and coding agents (Codex, Cursor, Aider, Junie, etc.).

The skill produces two outputs side by side: a structured audit report with score, findings, and prioritized actions, and an optional sparring discussion on the trade-offs behind the recommendations.

What is GEO?

Generative Engine Optimization (GEO) is the practice of making content discoverable, fetchable, and citable by generative AI systems — ChatGPT, Claude, Perplexity, Google AI Mode, Bing Copilot, and the rest. The term was formalized by the Princeton/Georgia Tech/AI2/IIT Delhi research group at KDD 2024 (arXiv:2311.09735).

GEO is sometimes used interchangeably with AEO (Answer Engine Optimization) and LLMO (LLM Optimization). They all describe the same shift: search traffic increasingly flows through generative answers rather than blue-link results, and the optimization rules for being cited by a model are not the same as the rules for ranking in a SERP. SEO remains a necessary baseline for Google AI Overview citations (37.9 % of AIO citations come from the top 10, per Ahrefs March 2026), but for ChatGPT, Perplexity, and Gemini the overlap with traditional top-10 rankings is below 30 % — different mechanics, different levers.

This skill is the operational form of that distinction.

What makes this skill different

Most GEO/AEO tooling on the market sells optimization without separating evidence from hype. This skill makes that separation explicit:

• Every recommendation carries a tier tag — S (platform primary sources), A (peer-reviewed academic research), B (empirical practitioner studies).
• Every recommendation carries a confidence flag — 🟢🟢 (method-sound, multi-tested, non-vendor-interest), 🟢 (solid finding), 🟡 (plausible but methodologically limited), 🔴 (marketing claim, anecdote, or no clear primary source).
• A mandatory "What to skip" section in every audit report names overhyped practices (Schema.org for AI citations, /llms.txt as a visibility lever, blanket date-bumping, programmatic SEO at scale) with the data that disproves them.

The knowledge base behind the skill is built on the GEO literature (Princeton KDD 2024, Toronto 2025, MIT/Columbia E-GEO, EMNLP 2025), platform documentation (OpenAI, Anthropic, Google, Bing, Vercel, Cloudflare), and large-scale empirical studies (notably the Ahrefs DiD study on Schema markup with n=1,885 + 4,000 matched controls, which shows no measurable Schema effect on AI citation rate).

Installation

Option 1 — Install via the Claude desktop app (recommended)

1. Download claude-agent-readiness-skill-v1.0.0.skill from the latest release.
2. In the Claude desktop app, open Customize → Skills.
3. Click + → Skill erstellen / Create skill → Skill hochladen / Upload skill and select the downloaded .skill file.
4. The skill activates immediately on any of the trigger phrases listed in SKILL.md. No restart required.

That's it — the app handles unpacking, validation, and registration. No manual file manipulation.

Option 2 — Clone the repository (for contributors and forks)

If you want to modify the skill, contribute changes back, or maintain a fork:

git clone https://github.com/byte5ai/claude-agent-readiness-skill.git
cd claude-agent-readiness-skill

To rebuild the .skill package after edits, use the skill-creator skill's packager (Python script bundled with Claude) or zip the contents with a .skill extension. Validate first to ensure the SKILL.md frontmatter stays schema-compliant (name, description, optional license, compatibility, metadata, allowed-tools).

Option 3 — Paste into a system prompt or project context

If you don't use the Claude desktop app's skill loader, the skill is plain Markdown. Paste the contents of SKILL.md and the relevant assets/ files directly into a Claude project's system prompt or context. The tier and confidence rules still apply.

How to use

Trigger the skill with any natural-language request that signals "audit a site for AI visibility", for example:

• audit this site for AI search visibility: https://example.com
• GEO audit for our docs subdomain
• is my site AI-ready?
• LLM citation optimization check on https://example.com
• agent readability check on our open-source repo

The skill clarifies mode (audit / discussion / both), scope (site / repo / both) and depth (full scan / quick) before starting if these are not specified in the request.

Repository layout

claude-agent-readiness-skill/
├── SKILL.md                              # Skill manifest, trigger phrases, workflow
├── knowledge-base.md                     # Evidence base with all tier/confidence tags
├── assets/
│   ├── audit-checks.md                   # Deterministic checks (~50 codes in clusters A–K)
│   ├── qualitative-rubric.md             # 1–5 rubric for content evaluation
│   ├── recommendations-library.md        # Actions sorted by evidence × effort
│   ├── report-template.md                # Output format for the audit report
│   └── discussion-prompts.md             # Sparring prompts on four trade-off axes
├── LICENSE                               # MIT
├── CHANGELOG.md
└── README.md                             # this file

Tier and confidence system

Every claim and recommendation inside the skill is tagged. The system is documented in detail in SKILL.md and knowledge-base.md. Short version:

Tag	Meaning
Tier S	Platform primary source (Google, OpenAI, Anthropic, Bing, Cloudflare, Vercel, published spec sites).
Tier A	Peer-reviewed or methodologically sound academic research.
Tier B	Practitioner empirical studies with their own dataset and transparent method.
🟢🟢	Multi-tested, method-sound (e.g., DiD with matched controls), finding contradicts vendor interest.
🟢	Solid finding, transparent methodology.
🟡	Plausible but methodologically limited or n=1.
🔴	Marketing claim, secondary synthesis without clear primary source, or anecdote.

Reports must always carry these tags inline with every concrete recommendation; output without tags is not allowed by the skill.

Limitations

• Knowledge base is dated 2026-05-16. Platform algorithms drift quarterly — recommendations have a finite half-life. See knowledge-base.md section 11 for open questions.
• Most underlying studies are English-language and US/global in scope. Language- and region-specific AI visibility patterns may differ.
• No active crawler telemetry integration (Bing Webmaster Tools, Cloudflare AI Crawl Control). The skill recommends checking those manually.
• The skill does not perform code quality analysis on repositories — only AGENTS.md coverage and structural readability for coding agents.

Maintenance

Findings in this field age fast. We recommend:

• Reviewing the knowledge-base.md quarterly against new platform documentation and major empirical studies.
• Re-checking the headline numbers in tier B (especially Ahrefs follow-ups, which have already corrected the 76% → 38% top-10 overlap once).
• Updating assets/recommendations-library.md when new evidence shifts the effort × impact balance.

Contributing

Contributions are welcome under the MIT license. Useful contributions include:

• New empirical studies with their own dataset that change one of the recommendations.
• Verification of existing tier B numbers against the primary sources.
• Translations (the skill is currently English-only).
• Bug reports for false claims or incorrect tier/confidence assignments.

Please open an issue before submitting a large PR so the change can be discussed first.

License

MIT — see LICENSE.

Author

Maintained by byte5 GmbH — Christian Wendler, cwendler@byte5.de.