How to use the seo-agentic skill for agent readiness audits

What the seo-agentic skill makes the agent do

An agent skill is a folder of instructions and supporting files that Claude loads when a request matches the skill's description. This one covers agent readiness: whether AI agents can read a page, find its controls and complete tasks on it.

When asked, the agent runs scripts that pull Google's Lighthouse Agentic Browsing score and check robots.txt, llms.txt, Markdown delivery, discovery files and the quality of the accessibility tree. It then reports findings in priority order and can draft fixes. It also explains what the Lighthouse score means.

• Full audit (default mode) of a URL
• Explanation of a Lighthouse Agentic Browsing result, from a live URL or a saved JSON report
• Draft fixes for robots.txt Content-Signal, llms.txt, ai-catalog.json and WebMCP
• Re-check of the dated facts the skill relies on

Install the seo-agentic skill in Claude

Download the ZIP from the skill's page in this directory. Unzip it into your personal skills folder, ~/.claude/skills/, or into a project's .claude/skills/ folder if you want the skill available only in that project.

If you use the Claude apps, you can instead upload the skill under Settings → Capabilities → Skills.

The download includes SKILL.md, a LICENSE.txt and reference files covering access policy, agent-friendly pages, discovery and Markdown, the Lighthouse category, the vendor matrix and WebMCP.

• Personal install: ~/.claude/skills/
• Project install: .claude/skills/ inside the project
• Claude apps: Settings → Capabilities → Skills, then upload

How to trigger the audit with natural requests

You do not need to type a command. Ask in plain language about whether AI agents can use or browse the site, or about llms.txt or WebMCP. The skill's description lists these topics, so matching requests should load it.

The skill also defines slash-style commands. The audit mode takes a URL, and the other modes take a URL or a saved report, as shown in the command table in SKILL.md.

• Can AI agents use and browse example.com?
• Audit example.com for agent readiness.
• What does our Lighthouse Agentic Browsing score of 3/5 mean?
• Draft an llms.txt for example.com.
• Do we need WebMCP tools on our checkout form?

What the audit checks and how to read the results

A full audit runs three checks in order: the Lighthouse fraction, HTTP and markup checks, and an accessibility-tree heuristic. The Lighthouse result is reported as X/N exactly as computed, not as a percentage. The accessibility score is a local 0-100 heuristic and is kept separate from the Lighthouse result.

Findings are grouped as P0 (accessible names and roles, CLS at or under 0.1, primary content present without JavaScript, robots.txt reachable, and WAF handling of verified bots), then P1, P2 and P3. Fix P0 items before lower priorities. Standards items such as WebMCP and Content-Signal are labelled as drafts or proposals.

• Lighthouse Agentic Browsing: X/N, with each audit marked pass, fail, informative or N/A
• Accessibility-tree heuristic: 0-100, reported on its own
• P0 failures: the count that limits agents most right now
• Access policy: training, search and user-triggered agents reported on separate lines

Run the audit safely: authorization, scripts and network use

The skill runs Python scripts (lighthouse_agentic.py, agentic_check.py and agent_ux_check.py) and makes network requests to the target site and to Google's PageSpeed Insights API. Only audit sites you own or have permission to test.

The WAF behaviour test is the sensitive part. It is added only when you control the site or confirm you are authorized, because it sends requests carrying AI agents' user-agent tokens. Its result describes how the site treats unverified traffic; it does not prove a real agent is blocked.

A Google API key is recommended so you are not sharing anonymous quota. The optional isitagentready.com scan is a third-party check; treat it as one operator's method, not a conformance test.

The SKILL.md commands call a wrapper at ${CLAUDE_PLUGIN_ROOT}/scripts/claude-seo. The source does not explain how that path is set up for a standalone download, so confirm the scripts are present and that Python runs before relying on a result.

• Requires a Python runtime
• Requires network access to the target and the PageSpeed Insights API
• Runs scripts: yes. Uses network: yes
• Optional second opinion: isitagentready.com

Example requests and what you get back

Ask for a full audit of your own site and you get a summary: the Lighthouse X/N on mobile and desktop, the accessibility heuristic, the count of P0 failures and one sentence on what most limits agents.

Ask for a fix and you get drafts printed for review. Nothing is deployed and no file is overwritten. The robots fix adds Content-Signal to each group and never changes Allow or Disallow. The skill asks for your training policy before choosing ai-train=yes or no. WebMCP drafts mark sends, purchases and deletes as consequential and keep a human confirmation step.

• Audit: summary, Lighthouse breakdown, prioritised findings, access policy, standards status and recommendations
• Fix: drafts for robots.txt, llms.txt, ai-catalog.json or WebMCP tools, with no deployment
• Explain: a plain-language reading of an existing Lighthouse report

Tips for getting useful results

Give the skill a saved Lighthouse JSON report if you already have one, so the explanation matches the exact run you are discussing.

Treat robots.txt as a preference signal, not access control. Protect private paths with authentication, because some user-triggered agents may ignore robots.txt.

Do not expect the skill to measure AI citations or brand visibility. The skill hands that work to a separate seo-geo skill, and commerce protocol depth goes to seo-ecommerce.

• Keep the report's evidence (status codes, headers, rule ids, selectors) when you pass recommendations to a developer
• Rerun the named check after each fix to confirm it worked
• Ask for the standards status if you are unsure whether a feature is ready to rely on