Permission comes first
Before an agent reads anything, it checks whether it's allowed in. OpenAI, Anthropic and Perplexity each publish named user agents for this — GPTBot and OAI-SearchBot for OpenAI, ClaudeBot and Claude-SearchBot for Anthropic, PerplexityBot for Perplexity — and each respects a site's robots.txt file for that agent specifically (OpenAI; Anthropic; Perplexity). A site can block one of these and allow another without realizing it, often because a security tool or CDN default quietly disallowed a bot nobody remembers configuring. The first thing an agent "sees" is whether it was let in at all — and a site that blocks it never gets to the next step.
Then it looks for structure it doesn't have to guess at
Once in, an agent parses the page. A person tolerates a cluttered layout because they can visually skim past the noise. An agent works from the underlying HTML, headings and metadata, and structured data where it exists. A page that relies on a person's visual judgment to understand — a logo that implies the business name, a photo that implies a service, a phone number rendered as an image — hands the agent nothing it can parse with confidence. It has to infer, and an agent that has to infer too much moves on to a site that made it easier.
Facts it can verify, not facts it has to trust
A person reading "family owned since 1998" on a homepage takes it at face value. An agent weighing whether to recommend a business looks for that same claim stated as a fact it can check against other signals — a consistent name, address and phone number across the places it looked, a clear statement of what the business actually does, contact details that resolve to something real. The gap between a claim a person accepts on sight and a fact an agent can verify is exactly where a lot of sites lose ground without realizing it.
A way to finish the job
Finally, an agent looks for what happens next. If it has decided a business is worth recommending, can it point the person it's helping to an actual next step — a booking link, a working phone number, a clear path to buy? A site that makes a person hunt for a phone number, or renders it in a way a machine can't read, gives the agent nothing to hand off. The recommendation dies at the last step, and the agent recommends whoever made that step obvious instead.
Why this adds up to more than the sum of its parts
None of these four things — permission, structure, verifiable facts, a clear next step — is exotic on its own. Most sites get at least one of them right by accident. The problem is that an agent needs all four in the same visit to recommend a business with confidence, and most sites were built to look good to a person, not to be legible to a system checking all four at once.