How Do LLMs Choose Which Sources to Cite?
LLMs choose which sources to cite through a series of steps. When an assistant answers with web search, it first decides whether to search at all. It then turns the question into one or more search queries and retrieves results from its search source. It reads a subset of those pages, picks the passages that best support what it wants to say, and attaches citations to parts of the answer. Pages are cited when they survive every step. That means ranking well enough for the queries the assistant runs, being readable to its crawler, and containing a clear, specific, current passage that answers part of the question. Each engine weighs these differently. And the cited sources are not always the ones that most shaped the answer.
What are the steps?
| Step | What happens | What it means for your pages |
|---|---|---|
| 1. Decide to search | The model judges whether it needs current or specific information | Questions about prices, recent events or local options usually trigger a search; general ones may not |
| 2. Write queries | The question is rewritten, often into several searches | Your page must rank for the queries the assistant runs, not only the user's wording |
| 3. Retrieve results | Results come from the engine's search source | Visibility in that index matters, whether Google, Bing, Brave or another |
| 4. Read pages | A subset of results is fetched and read | The crawler must be allowed, and key text must be in the HTML |
| 5. Select passages | Passages that best support the answer are chosen | Direct, specific, self-contained passages are easier to use |
| 6. Attach citations | Sources are linked to claims in the answer | Some engines show many citations; others show few |
Google has described its AI Mode as using "query fan-out", running many related searches at once. Analysis of ChatGPT's shopping answers has found it running multiple searches in the same way.
Do search rankings decide citations?
They matter, but less directly than many assume, and it depends on the engine:
- Google's AI Overviews: Ahrefs' 2026 analysis found about 38% of citations came from the top 10 organic results, and about 31% from pages beyond the top 100.
- ChatGPT search: ChatGPT draws on third-party search providers and its own indexes, and the evidence on overlap with Bing is mixed. Seer Interactive found in February 2025 that 87% of SearchGPT citations matched Bing's top 20 for the same question, but Ahrefs found in July 2025 that only about 8% of ChatGPT's citations were in Bing's top 10, and about the same share in Google's.
- Claude: its web search uses Brave. One analysis of about 35,000 Claude citations found 79.2% came from Brave's top 10.
Because assistants run their own reworded queries, a page can be cited without ranking for the user's exact question. It only needs to rank for one of the searches the assistant ran.
What makes a page more likely to be cited once it is retrieved?
- A direct answer: a passage that answers part of the question in a sentence or two, near the top of a section.
- Specific facts: numbers, prices, dates, names and steps are easier to cite than general claims.
- Freshness: Seer Interactive found in 2026 that 75% of pages cited by ChatGPT, Gemini and Perplexity had been updated within the past year, and 88% within two years.
- Readable structure: clear headings, lists and tables in HTML.
- Independence and authority: Muck Rack's analysis of more than 25 million cited links found about 84% came from earned sources, such as journalism, reference sites and third-party content.
Length on its own does not help. Ahrefs found almost no correlation between word count and being cited in AI Overviews.
Why do different engines cite different sources?
Each uses a different search source, a different model and different rules for how many citations to show. A Semrush and Growth Memo study found ChatGPT cited brands' domains far more often than it named them, while Gemini named brands far more often than it cited them. Citation habits also change over time. seoClarity recorded sharp platform-level drops in ChatGPT citations in March and April 2026, followed by a rebound. Why sources differ across engines is covered in why ChatGPT, Perplexity and Google cite different sources for the same query.
Are cited sources the ones the answer actually used?
Not always. Research published in 2026 examined how a language model decides whether to attach a citation. It found the decision is spread across many internal components. The authors concluded that inline citations can create a false sense of security. Answers can miss citations for sources they relied on, and add citations that did not really support the claim. Journalism researchers have documented the practical side. In a 2025 Columbia Journalism Review study, eight AI search tools gave incorrect answers to more than 60% of queries asking them to identify news sources.
For brands, this means two things. A citation is evidence that your page was retrieved, not proof that it shaped the answer. And an answer can reflect your content without citing you, which is one reason unlinked brand mentions matter.
What might an illustrative trace look like?
Consider an illustrative question to an assistant: "What's the best CRM for a 20-person real estate team?" The assistant decides to search, because the answer depends on current products. It runs several queries, such as "best CRM for real estate teams", "real estate CRM pricing" and "CRM for small real estate brokerages". The results include review platform category pages, two "best real estate CRM" articles, three vendors' pricing pages and a trade publication's roundup. It reads six of them. One vendor's page states its per-user price and real estate features in a short table, so that passage is used for pricing. The trade roundup supplies the list of options. A vendor whose site loads its pricing through a script is never read, so its price is taken from an old review instead. The final answer cites four of the six pages, and names two brands that appear only in text the assistant read but did not cite.
How can you use this?
- Check crawler access and rendering for the assistants you care about.
- Rank for the searches assistants run, not just your head term. Bing's AI Performance report shows the grounding queries used to find your content in Copilot.
- Write passages that answer sub-questions directly, with specific facts.
- Keep important pages current, and show when they were updated.
- Earn third-party coverage, since much of what gets cited is independent.
- Track citations by engine, since each behaves differently.
Why some page types are cited more than others is covered in which types of pages LLMs cite.
What do people ask about how LLMs cite sources?
Do LLMs cite the most authoritative source?
Not necessarily. They cite retrieved pages that best support the answer, which favours relevant, clear and current passages as well as authority.
Can a small site be cited?
Yes. Many citations go to pages outside the top search results, especially for specific questions.
Does structured data make citation more likely?
It helps search engines understand pages, but assistants mostly quote visible text. Structured data must match what is on the page.
Why does the same question cite different sources each time?
Assistants vary their queries and choices between runs, and results change as indexes update. Measure over repeated runs.
Published by The Kovmere Team · updated Sep 29, 2026 · More resources