How ChatGPT, Claude and Perplexity decide which pages to cite

When someone asks ChatGPT, Claude or Perplexity a question that needs fresh facts, the assistant does not answer from memory alone. It runs a web search, fetches a handful of pages, reads them, and writes an answer that links to the sources it leaned on. Those links are the new search result. If your page is not among them, the visitor never sees you.

None of the three companies publishes a ranking formula. But the pipeline is the same everywhere, and every stage has a way to drop your page.

Step one: the assistant has to be allowed in

Each assistant fetches pages with its own crawler, and each crawler reads your robots.txt. OpenAI uses OAI-SearchBot for ChatGPT search, Anthropic uses Claude-SearchBot, and Perplexity uses PerplexityBot. If your robots.txt blocks them, or a firewall rule treats them as bots to challenge, your page cannot be read, so it cannot be cited.

This is the most common failure we see in audits, and it is usually an accident. A blanket "block AI" rule written to stop training crawlers also blocks the search crawlers that send visitors. We wrote a separate guide on telling the two apart.

Step two: the page has to be readable without a browser

AI crawlers fetch HTML. Most do not run JavaScript the way a browser does. If your content only appears after a client-side app boots, the crawler sees an empty shell.

Check by viewing the page source, not the inspector. If the headline and the main text are in the raw HTML, you are fine. If all you see is a root div and a script tag, the assistant sees the same.

Step three: the page has to be understood

Once a page is fetched, the assistant looks for passages that answer the question. Pages that make this easy get quoted more often. In practice that means:

• One clear H1 that says what the page is about.
• Headings that read like the questions people ask.
• Short paragraphs that state the answer first and explain after.
• Facts written as text, not hidden in images or tabs.

Structured data helps the assistant be sure what kind of thing the page describes. An Organization or Product block in JSON-LD removes guesswork about names, prices and who is behind the site. Our schema guide covers the five types most sites need.

Step four: the page has to be worth trusting

Given two pages that say the same thing, an assistant tends to prefer the one that looks maintained and specific. A visible date, a named author, concrete numbers and a clear owner all signal that. Thin pages that restate what everyone else says are easy to skip, because there is nothing in them to cite.

Where to start

Run a free audit on the page you most want cited. It checks crawler access, server-rendered content, headings and structured data, and tells you which fix moves the score most. Fix the access problems first, because nothing else matters until the assistant can read the page.