Why does AI choose one source over another?

Say someone asks an AI, “Who can paint a portrait of my cat?” It can answer from what its model learned in training, from what’s in the conversation, such as a photo of the cat, or from pages it looks up while it answers. Often it mixes them. A model stops learning when its training ends, so new pages reach its answers through search. That’s the part the steps below follow.

Research: Search Mode Changes the Local Businesses an AI Recommends

Before searching, the engine often rewrites a question into a few shorter searches, such as “custom cat portrait from photo”, “pet portrait artist” and “best pet portrait artists”. Google calls this query fan-out. It’s one reason a page can match the words someone typed and still not appear, and how a “best of” list can end up in the answer.

Research: Reworded Buying Questions Produced Different AI Brand Lists

Each search returns a shortlist of pages from an index. It doesn’t always match Google’s first page: Google’s own AI Overviews cite sites from further down too, so a ranking report shows only part of the picture.

Research: Track AI Overview Citations Alongside Google Rankings

A page the engine’s search crawler can’t reach never makes the list, however good it is. AI companies run several crawlers with different jobs. OpenAI’s OAI-SearchBot finds pages for ChatGPT search. GPTBot collects pages to train OpenAI’s models. ChatGPT-User visits a page when a person asks it to, and robots.txt rules may not apply to it. Google-Extended decides whether Google may use your pages for Gemini, and doesn’t affect Google Search. So a studio can block those training crawlers and still be found by ChatGPT search. Know which crawler does what before you change anything.

Research: Before Paying for AI Page Formatting, Check How Sources Get Found · Glossary: robots.txt and crawler rules

From each page on the shortlist, the engine keeps the passages that answer part of the question. A page can appear in the source list while its name, address or ordering details never reach the answer.

Research: When AI Cites a Page, Which Details Reach the Answer?

The model writes the answer from those passages and links its sources. A link makes a claim easier to check; it doesn’t prove the claim, so open the source.

Research: What an AI Citation Actually Tells You About Your Business

An answer can use a page and link it without ever naming the business behind it. Available, retrieved, mentioned, cited and recommended are separate outcomes, and none of them guarantees the next.

• In the example, a small studio that paints pets as Renaissance portraits had its page available, retrieved and cited.
• Its idea reached the answer: “some artists paint pets as Renaissance portraits”.
• The answer still named a marketplace and a list, and never mentioned the studio.

Research: Which Step Is Missing Between Your Page and an AI Mention?

See how we count this in a real audit

A large study of Chinese-language AI search found that a business mentioned in many of an answer’s sources is more likely to be named in that answer. In the example, the studio’s name appeared on one source, its own page; the marketplace’s appeared on three. The study measured an association, and its authors call for controlled tests of cause.

Research: When AI Cites a Page, Which Details Reach the Answer?

Ask the same question tomorrow, from another city or in another language, and the answer can change. One screenshot shows one answer, so it can’t show that an edit worked. Keep a small set of real customer questions, note how you asked them, and repeat them over time.

Research: What Did That GEO Experiment Actually Improve? · Separate Language from Location in a Local AI Search Audit

Before you block or allow an AI crawler, find out what it does. Then ask AI your customers’ questions from time to time, and check whether the pages it links to mention you.

The studio, the marketplace, the list, the painting and the answers in this example are made up: it’s an illustrative run.