Which AI actually reads your site? Two months of LLM traffic, measured—Martian Chronicles, Evil Martians’ team blog

Evil Martians instrumented their own site with server-side middleware to track LLM agent traffic over two months, capturing ~268,000 agent requests vs ~107,000 human pageviews. Key findings: ChatGPT-User accounts for 73% of all agent traffic and reads rendered HTML almost exclusively (0.1% Markdown); Claude Code requests Markdown 76% of the time via HTTP content negotiation. The llms.txt file received ~770 direct fetches but only 37 came from named AI assistants — the rest were SEO bots and crawlers. A hidden AI hint div earned zero attributable fetches. The practical takeaway: keep HTML clean for ChatGPT, implement Accept: text/markdown content negotiation for coding agents, and measure your own logs rather than following generic GEO advice.

Jul 21•15m read time•From evilmartians.com

Table of contents

Making your site visible to LLMs: 6 techniques that work, 8 that don't How do you measure LLM traffic when the clients don’t run JavaScript?How much of your traffic is even human?About 15% of agent reads are Markdown Why “optimize for LLMs” is the wrong instruction Does anything actually read llms.txt?What surprised us?What we’d change, knowing all this FAQ

407 Impressions

Comment

Bookmark

Copy