How AI engines pick the sources they cite

Watch · 3:08

• 0:00Welcome
• 0:27The retrieval pipeline
• 1:11Each engine has its own index
• 1:43The four gates
• 2:22Answers vary from run to run
• 2:41How AI Rank Lab helps

In shortMost AI answers that cite sources are built in two steps. First the engine turns your question into one or more web searches (Google calls this query fan-out) and retrieves candidate pages from its index. Then the model reads those pages and writes an answer, citing the few passages it actually used. To be cited, a page has to be crawlable, indexed, retrieved for one of those searches and contain a clear answer worth quoting.

Two places an answer can come from

A large language model (LLM) can answer from what it learned in training, or from pages it fetches while answering. Training knowledge is fixed at a cutoff date and has no links attached. Fetched pages are fresh, and they are where citations come from. When ChatGPT, Gemini, Claude, Perplexity or Google's AI features show a list of sources, the engine searched the web for that answer.

Most answers mix the two. The model uses training knowledge for general framing and retrieved pages for specifics: prices, current products, recent news, comparisons. That is why a brand the model "knows" can still go uncited, and why a new page can be cited within days if search engines find it.

The retrieval pipeline, step by step

1. Interpret the question. The engine decides whether it needs fresh information at all. A question like "what is a sitemap" may be answered from memory; "best CRM for small agencies in 2026" almost always triggers a search.
2. Fan out. The question is rewritten into several narrower searches. Google says AI Overviews and AI Mode may use a "query fan-out" technique, issuing multiple related searches across subtopics and data sources (Google Search Central (opens in a new tab)). A buyer question might become searches for pricing, reviews, alternatives and a feature comparison.
3. Retrieve. Each search returns candidate pages from the engine's index. If your page does not rank for any of the fan-out searches, the model never sees it.
4. Read and select. The model reads the retrieved pages, or passages from them, and picks the ones that answer each part of the question most directly and credibly.
5. Write and cite. The model writes one answer and attaches a handful of citations. Pages it read but didn't use are not shown.

Each engine searches its own index

There is no single "AI index". Each engine retrieves from its own search system and uses its own crawler, which you can allow or block in robots.txt. The table lists what each company documents.

Why some pages get cited and others are ignored

Think of citation as four gates. A page has to pass all four.

• Eligible. The page can be crawled and indexed. It isn't blocked in robots.txt, isn't noindex, and its main content is visible without a login. For Google, the stated bar is "indexed and eligible to be shown in Google Search with a snippet".
• Retrieved. It ranks for at least one of the fan-out searches. This is where classic SEO still matters: relevance, links and a page that matches what the searcher wants.
• Selected. It contains a passage that answers a sub-question directly: a clear definition, a specific number, a comparison table, a step list. Vague marketing copy gives the model nothing to quote.
• Corroborated. For opinion questions ("best", "top", "vs"), engines often lean on independent sources such as review sites, comparison articles and forums. Being named on those pages can matter as much as your own site.

What this means for you

• Make sure the search crawlers listed above can reach your important pages.
• Rank for the narrower questions behind a buyer question, not only the head term.
• Write passages that can be quoted on their own: answer first, then detail.
• Get your brand onto the third-party pages engines already cite for your category.

How AI Rank Lab helps

Brand visibility runs the questions you choose on ChatGPT, Claude, Gemini, Perplexity, Google AI Overview and Google AI Mode, and stores which URLs each engine cited. For the four chat engines it calls each company's API with web search switched on, so you see real retrieved sources. API answers can still differ from what a signed-in person sees in the consumer app, and prompts run only when you run them. The Citations page then shows which domains and pages win your questions.