Yeşilyurt rebuilds his site so only AI agents can read it
The practitioner replaced his WordPress homepage with a Markdown, API-first version served to about 30 agent families.
Metehan Yeşilyurt announced on LinkedIn on 9 October 2026 that he has rewritten metehan.ai from scratch for AI agents. Humans still see a normal site at the same URL. Claude, ChatGPT, Codex, Cursor, Perplexity, Gemini, Grok, Hermes, OpenClaw and about 30 other agent families get clean Markdown, an API-first homepage, no menus, no scripts and no images.
The agent homepage names the topics he covers and lists each endpoint with a working example. One endpoint, /api/ask?q=, takes a question and returns an answer with sources. Another, /api/agent/hello, lets a visiting agent identify itself, its model, its operator and its reason for the visit. IPs are not stored and only aggregate counts are published at /api/agent/stats.json.
Yeşilyurt also pulled 313 items from his own writing. He then scored each one with a decision model on OpenRouter for whether it is supported, independent, original and useful. He says the scoring is still in progress.
He frames the build as an experiment with four questions. Do agents read the main page first or go straight to the API. Which agents introduce themselves when asked. Does a smaller verified set get cited more than a large archive. Two further tests he is keeping private for 30 days.
The move lands against a wider argument that the human-facing site is losing its role as the primary surface. James Ward, Head of Digital Experience at Personio, asked on the Masters of Search podcast whether websites as we know them are heading toward becoming databases for LLM front ends.
What we think this is, and is not
We think this is a useful public test, not a template. One practitioner has rebuilt one site and has not yet reported citation or traffic outcomes. The questions he is asking are the right ones for the category, and the answers will be worth reading in 30 days.
Before stripping your own site for agents, measure how current crawlers already fetch it. Check which pages get cited and for which queries. Run the rebuild as a hypothesis on a small subset of URLs, with the full version still live for comparison. A site that serves humans badly to serve agents well is a bet on a retrieval behaviour none of us has measured yet.