seo-ai-search — a Claude Code skill for SEO and AI-search visibility

Left to their training data, coding agents "do SEO" from folklore: deprecated schema types, meta-keyword habits and confident claims about AI crawlers that are simply wrong. Anyone using an agent to build or audit a site inherits those mistakes. The problem was to encode current, evidenced guidance in a form an agent can actually apply — and that flags which facts go stale.

Open-source · MIT · Published September 2026

Approach

• Treat the skill as a knowledge product with an engineering discipline: every load-bearing claim cites a primary source (Google and Bing documentation, crawler datasets from Vercel/MERJ and Cloudflare, the Princeton GEO study, click-through studies).
• Split the content into a stable core (SKILL.md) and lazily loaded references so the agent only reads what the task needs — technical SEO, structured data, AI search, local search, keywords and content, and an eight-layer audit checklist.
• Mark volatile facts explicitly so the skill instructs the agent to re-verify rather than repeat stale numbers.
• Ship tooling, not just prose: a zero-dependency audit script for robots, sitemaps, raw-HTML completeness and per-AI-bot access, and an IndexNow ping script for the Bing/ChatGPT indexing pipeline.

A Claude Code skill is a directory the agent loads on demand. SKILL.md holds the workflows and stable rules; the references directory is read lazily per task; the scripts run as plain Node against a live site and return CI-friendly exit codes.

• 01SKILL.md defines when the skill triggers (implementing SEO, auditing visibility, choosing schema, keyword research, local search, AI-answer visibility) and the workflows to follow.
• 02Six reference documents cover JavaScript rendering by crawler, the current rich-result gallery with deprecation dates, what feeds each AI assistant, the profile-first local playbook, SERP-composition keyword research with honesty tiers, and a strict audit order.
• 03scripts/audit.mjs checks layers one and two of the audit — robots and sitemap health, raw-HTML completeness and access for each AI crawler — and exits non-zero on critical findings so it can run in CI.
• 04scripts/indexnow-ping.mjs pushes the sitemap to IndexNow after each deploy.
• 05Installable per user or per project with a single git clone; validated with `claude plugin validate`.

The project is AI-native in the other direction: it is guidance and tooling for an AI coding agent. The engineering decisions — stable core vs lazy references, explicit volatility flags, a refusal to promise rankings — exist because agents otherwise over-generalise from training data.

• Claude Code skills
• Node.js (zero dependencies)
• Markdown
• IndexNow API

Outcome to be added. Results are only published here when they can be stated factually — no estimated metrics.

• Knowledge for agents needs the same rigour as code: sources, versioning and explicit statements about what will go stale.
• Lazy loading matters for context budgets — the agent should read the reference it needs, not the whole corpus.
• A script that returns a CI exit code is worth more than a paragraph explaining what to check.