Agentic AI coding assistants and your website: how Claude Code and OpenCode read your docs in 2026

Key takeawaysn- Coding agents don't browse like ChatGPT Search. Claude Code, OpenCode and Codex fetch specific documentation pages, READMEs and API references to finish a task, which makes your docs a runtime dependency for other people's code.n- Claude Code never reads your raw page. It fetches HTML, converts it to markdown, cuts it at 100KB of text, and gets back a summary written by a small, fast model answering one question at a time.n- Four bots decide whether agents can read you live: Claude-User, Claude-SearchBot, OAI-SearchBot and ChatGPT-User. Blocking the famous training bots (ClaudeBot, GPTBot) doesn't stop live reads, and blocking the user bots makes you invisible to agents.n- No major AI crawler executes JavaScript. If your docs need JS to display their content, agents see an empty shell.n- Markdown and llms.txt serve agents, not AI search rankings. Markdown files are just 0.05% of AI search citations, but agent crawlers hammer /docs/*.md paths. Serve both formats, and expect llms.txt itself to be read rarely.n## Why coding agents suddenly read your docsnnA few years ago your documentation had two audiences: humans and Googlebot. Now there's a third, and it behaves like neither. When a developer asks Claude Code to wire up your API, the agent reads your docs itself, interprets them, and writes code based on that reading. If the docs are unreadable to machines, the integration quietly fails and the developer concludes your product is hard to use. Your docs have become a runtime dependency for code you don't control.nnData from Promptwatch makes the shift explicit. Its markdown in AI search report puts it plainly: agents powering Claude Code and OpenAI's Codex