Whose hand is this?
The file itself
Shown exactly as served, because it is the evidence. Lines that name an agent on the register are marked.
Three decisions, not one
Nearly every tool treats “AI crawlers” as a single switch. They are not. The operators run separate agents for separate jobs, and refusing each one costs you something different.
Training
Collects pages that may train a future model. Blocking costs you no traffic at all. This is the one most people mean when they say they want to block AI.
GPTBot · ClaudeBot · CCBot · Bytespider
Search
Builds the pool an assistant cites from. Blocking means you stop being cited when someone asks a question your page answers.
OAI-SearchBot · Claude-SearchBot · PerplexityBot
Errand
Sent because a specific person asked about your specific page. Blocking tells that reader your page cannot be read. Almost nobody chooses this on purpose.
ChatGPT-User · Claude-User · Perplexity-User
Two names you will see on block lists are not crawlers at all. Google-Extended
and Applebot-Extended are robots.txt opt-out tokens for training permission;
nothing ever sends them as a user-agent. Blocking them is meaningful, but it is not blocking
a visitor.