Misconception: every AI crawler is the same kind of traffic. This leads teams to either block everything or allow everything. Both choices can damage GEO because training, search discovery, user-triggered fetching, and preview generation are different jobs.
Verified signal: Hermes recorded official documentation from OpenAI, Anthropic, Perplexity, Apple, Meta, and Amazon. These sources repeatedly separate training bots, search bots, user-triggered fetchers, and preview or ads bots. Robots rules, IP JSON, WAF rules, and update windows are not interchangeable.
Action checklist: build a control matrix before changing access. Track provider, bot type, user agent, IP JSON URL, robots rule, meta rule, last synced time, allowed domains, WAF action, and notes access result. Do not use a training opt-out as proof that search visibility is blocked or preserved.
19LAB test result: 19LAB and yijiu.me should protect search and user-triggered reading while still making deliberate choices about training access. The current goal is not to block AI. The goal is to let the right agents read the right pages.
Reusable principle: GEO access control starts with classification. If bot roles are mixed, every downstream visibility result becomes hard to interpret.