seo-agentic audits how ready a site is for AI agents that browse, fill forms and act for people. It reads Google’s Lighthouse Agentic Browsing result as an X/N fraction, checks server-rendered content, robots.txt groups and Content-Signal, llms.txt, Markdown delivery, well-known discovery files and WebMCP markup, and scores the accessibility tree for agents with a local heuristic.
Findings are ranked P0 to P3 and each recommendation names the check to rerun. A fix mode prints drafts to the screen for review without deploying or overwriting anything.
What You Get
- An X/N Lighthouse Agentic Browsing result for mobile and desktop
- A prioritized findings list from P0 to P3 with evidence and the check to rerun
- A separate agent-UX score for the accessibility tree
- One-line access policy for training, search and user-triggered agents
- Draft robots.txt Content-Signal lines, llms.txt, ai-catalog.json and WebMCP tool sketches printed for review
Common Use Cases
- Check whether ChatGPT, Gemini or Claude browsing agents can use a checkout or signup form
- Decide a robots.txt policy that separates AI training from AI search
- Walk a client through a Lighthouse Agentic Browsing score
- Draft WebMCP tools for a site with forms or transactions
What Makes It Unique
It treats agent readiness as accessibility, performance and access policy with a Markdown and discovery layer on top, and keeps WebMCP optional. Every vendor fact carries a dated source grade, the skill refuses to promise ranking or citation gains, and it keeps the Lighthouse fraction apart from its own heuristic score.
Under The Hood
The skill loads SKILL.md and reads six reference files on demand. Its steps run shell commands that start Python scripts through the claude-seo launcher: one calls the PageSpeed Insights API for the Lighthouse category, one fetches the target site, robots.txt, llms.txt and well-known paths over the network, and one drives a headless Chromium for the accessibility tree. When the API quota runs out it suggests a local npx lighthouse run. An optional user-agent matrix sends requests carrying AI agent tokens and only runs when you confirm you are allowed to test the site. Fix mode prints drafts to stdout. Refresh mode edits the dated reference files and a constant in a script. The frontmatter declares none of this. Fetched content is treated as untrusted data, and requests pass an SSRF guard.