AI Citation Explained: How AI Chooses Which Sources to Cite

AI citation comes down to two things: whether your page is found when the AI searches, and whether it is chosen once found. When ChatGPT, Gemini, Perplexity or Google AI Overviews answer a question that needs current information, the answer goes through five stages: the user asks, the AI rewrites the question into search queries, it retrieves pages, it picks passages, and it writes an answer with cited sources. GEO can influence the two middle stages: retrieval and selection.

This guide explains where AI answers come from, what each of the five stages does, and where your page most often gets dropped.

What are the two sources of AI answers?

AI answers come from training data and retrieved results. Training data is text the model read when it was built; retrieved results are pages the AI finds on the web at the moment it answers.

Retrieval is where GEO happens, because it updates quickly and comes with cited sources. The major AI tools document how they retrieve:

• ChatGPT search indexes pages with OpenAI's OAI-SearchBot crawler, which is separate from GPTBot, the crawler that collects training data (OpenAI crawler documentation).
• Gemini can ground its answers in Google Search results and returns the pages it used (Gemini grounding documentation).
• Google AI Overviews cites pages Google has indexed and can show as snippets, with no additional technical requirements (Google Search Central).
• Claude's web search tool always includes citations when it uses search results (Claude web search tool documentation).

For a detailed comparison of each platform's retrieval, see ChatGPT vs Gemini vs Perplexity.

The five stages of a search-grounded answer

A search-grounded AI answer follows a fixed pipeline. To appear among the cited sources, your page has to make it through every stage.

1. The user asks: for example, "Which bookkeeping services in Taipei do you recommend?" Questions are usually longer and more conversational than search keywords.
2. Rewrite into queries: the AI breaks the question into several search queries, such as "Taipei bookkeeping service" and "bookkeeping firm fees".
3. Search and retrieve: the AI runs those queries and retrieves a set of relevant pages. Pages that are not indexed, or that block AI crawlers, do not exist at this stage.
4. Pick passages: from the retrieved pages, the AI picks the passages that answer the question most directly and with evidence.
5. Write the answer: the AI combines the chosen passages into one response and marks the cited sources next to sentences or below the answer.

Stages 1, 2 and 5 are up to the AI. Stage 3 decides whether you are found, and stage 4 decides whether you are chosen. Those two stages are what GEO works on.

Where does your page get dropped?

The first three rows are about being found; the last three are about being chosen. Fix being found first: if your page is never retrieved, great content has no chance.

What kind of passages get chosen?

AI favors passages that are direct, specific and supported by evidence. There is experimental evidence for this: the GEO paper (Aggarwal et al., KDD 2024) rewrote the same pages using different methods, had generative engines answer queries, and measured how much of each answer came from each page.

Overall scores are from Table 1 of the paper; the relative changes are calculated from those scores, and the paper's own conclusion is that these three methods improve visibility by roughly 30–40%. Keyword stuffing, common in traditional SEO, performed worse than doing nothing.

The reason is that AI needs a sentence it can quote directly. "Outsourced bookkeeping for a small company typically costs a few hundred dollars per month, depending on transaction volume" is easier to cite than "fees are reasonable", and a claim backed by an official source is more credible than one without.

Why doesn't ranking first guarantee a citation?

AI does not cite in ranking order, because it picks passages, not pages. Table 2 of the GEO paper found that with sources added, a site ranked 5th in search results gained 115.1% in answer visibility, while the site ranked 1st dropped 30.3%.

This means a lower-ranked site can overtake higher-ranked ones in AI answers by being more direct and better supported. For small brands, this is an opportunity traditional SEO rarely offers.

Does training data still matter?

Training data still matters, but it cannot be changed quickly. Even when AI searches the web, whether it recognizes your brand, links you to a topic, and how it describes you are shaped by what the web has said about you in the past.

The way to influence training data is to be mentioned steadily, over time, in credible places: press coverage, industry articles, reviews and official directories. That content gradually enters future training data and is also retrieved when AI searches. In practice, focus on retrieval first while building up outside mentions.

Frequently Asked Questions

Q: How does AI decide which page to cite?

AI rewrites the question into search queries, retrieves a set of pages, then cites the passages that answer the question most directly and with evidence.

To be cited, a page must be retrieved at the search stage and then beat other pages at the selection stage. GEO works on these two stages.

Q: Do AI answers come from training data or the web?

Both. AI answers draw on training data and on results retrieved at the moment of answering.

For questions that need current information, ChatGPT, Gemini, Perplexity and Google AI Overviews search the web and cite sources. Training data shapes the AI's basic impression of your brand and only updates with the next model version.

Q: Will ranking first in search get me cited by AI?

Not necessarily. AI picks the most citable passages, not the highest-ranked pages.

The GEO paper found that with sources added, the site ranked 5th gained 115.1% while the site ranked 1st dropped 30.3%. Ranking decides whether you are retrieved; content decides whether you are chosen.

Q: What kind of content is most likely to be chosen by AI?

Content that answers directly at the top, uses specific numbers and cites verifiable sources is most likely to be chosen.

In the GEO paper's experiments, quotation addition scored 27.8, statistics addition 25.9 and citing sources 24.9, all above the unoptimized 19.5.

Q: Do more keywords make AI more likely to cite me?

No. Keyword stuffing reduces how much of the answer comes from your page.

In the GEO paper, keyword stuffing scored 17.8, below the unoptimized 19.5. Let key terms appear naturally in the title, opening and subheadings.

Q: Why does AI mention my brand but cite someone else's page?

Because AI recognizes you, but your page was not retrieved or not chosen at the search and selection stages.

First check that robots.txt allows AI crawlers and that the page is indexed. Once it can be found, rewrite the opening to answer directly and add numbers and sources.

Q: Does blocking GPTBot keep my pages out of ChatGPT search?

No. GPTBot collects training data; ChatGPT search uses a different crawler, OAI-SearchBot.

OpenAI documents that the two are controlled separately. To stay out of training but appear in ChatGPT search, block GPTBot and allow OAI-SearchBot in robots.txt.

Q: How long does it take to influence training data?

Usually months to years, and only through long-term presence.

Training data updates only with the next model version. Being mentioned steadily in press, industry articles and reviews gradually teaches AI to recognize your brand; in the short term, focus on retrieval.