How do LLMs choose sources to cite? In short, AI search tools search an index for pages that match the question, read the most relevant passages, and cite the ones that state a clear answer with specific facts. There’s no secret switch. If a search system can’t crawl and index your page, it can’t be a candidate, and if your passage doesn’t answer the question plainly, another page’s will.
Key Takeaways
- LLMs use web content in two ways: what they learned in training, and what they retrieve live. Citations come almost entirely from live retrieval, which is the part you can influence.
- Retrieval works on passages, not whole pages. The system looks for a chunk of text that answers the question on its own.
- Google says no special markup or files are needed for AI Overviews or AI Mode. A page must be indexed and eligible to show with a snippet.
- Research on generative engines found that adding citations, quotations and statistics made content more visible in AI answers, while keyword stuffing did not help.
- The exact ranking weights are not public. Be wary of anyone who claims to know them or guarantees a citation.
This guide is for owners and marketing leads who want to understand the mechanism before changing their site. When you’re ready to act, our generative engine optimization checklist turns it into 12 on-site changes.
Two Ways LLMs Use Web Content
An LLM uses web content in two different ways: what it absorbed during training, and what it fetches when you ask a question. Citations almost always come from the second route.
A large language model learns to predict likely language from huge amounts of text. That training is a snapshot. It can shape what the model “knows” about your brand, but it doesn’t come with neat footnotes, and you can’t edit it after the fact. Some AI companies let you opt out of training crawls separately from search crawls, so you can make that choice on its own terms.
How Retrieval Picks Candidate Pages
The citations you see in ChatGPT search, Perplexity, Claude and Google’s AI Overviews come from retrieval, sometimes called retrieval-augmented generation (RAG). The steps look roughly like this:
- The question becomes searches. The system rewrites your prompt into one or more search queries, often several for a complex question.
- An index returns candidates. Pages come from a search index the tool can access: its own crawl, a search partner or both.
- Passages are scored. The system reads the candidates in chunks and keeps the passages most relevant to the question.
- The model writes the answer. It drafts a response from those passages and links the sources it relied on.
Two consequences follow. First, if your page isn’t in the index, nothing else matters. Second, you are competing at the passage level. A long page with one buried answer can lose to a short page that states it in the first sentence.
What Each Engine Has Said Publicly
- Google: its documentation on AI features and your website says a page must be indexed and eligible to show in Search with a snippet to appear as a supporting link in AI Overviews or AI Mode, with no additional requirements. Its guide to generative AI features adds that structured data isn’t required, that Search ignores llms.txt files, and that optimizing for these features is still SEO.
- OpenAI: ChatGPT search relies on its OAI-SearchBot crawler. Sites that block it won’t be shown in ChatGPT search answers, while GPTBot is the separate training crawler (OpenAI).
- Perplexity: PerplexityBot surfaces sites in Perplexity’s search results and, according to Perplexity, is not used for training (Perplexity).
- Anthropic: Claude-SearchBot indexes content for Claude’s search, Claude-User fetches pages when a user asks, and ClaudeBot collects training data (Anthropic).
None of them publishes how candidate passages are ranked. Anyone who tells you the precise weighting is guessing, or selling something.
What the Research Says
The best-known study is “GEO: Generative Engine Optimization” by Aggarwal and colleagues, presented at KDD 2024. The researchers tested content changes against generative engines and found that adding citations to sources, quotations from relevant experts and statistics improved a source’s visibility in AI answers by roughly 30 to 40% in their tests. Keyword stuffing gave little or no improvement (arXiv).
Treat this as direction, not law. It was a controlled experiment, and the commercial engines have changed since. But it matches what you would expect from a system looking for passages it can trust: evidence beats repetition.
What Makes a Page Easy for AI to Quote?
A quotable page answers the question in its first sentence, backs it with specific facts, and uses headings that match how people phrase the question. The model is looking for a passage it can reuse with confidence.
Think about how a retrieval system reads your page: in chunks. A paragraph that opens “It depends on a number of factors” gives it nothing. A paragraph that opens “A basic business website typically takes 4 to 6 weeks to build” gives it a sentence it can cite. The second one is also better for humans, which is the point.
- Direct answers up top. Put the answer in the first one or two sentences under each heading, then explain.
- Specific facts. Numbers, names, prices, dates, steps. Vague claims get skipped.
- Evidence. Link to primary sources and name them, so the passage carries its own proof.
- Self-contained paragraphs. Each one should make sense if someone quotes it alone. Avoid “as mentioned above.”
- Corroboration elsewhere. If other trusted sites describe you the same way, a model has more reason to repeat it.
A Before and After Rewrite
This is an illustrative example we wrote to show the pattern, not a client result. Say a hotel’s page answers “Do you allow early check-in?”
Before: “We pride ourselves on flexibility and want every guest’s stay to begin on the right foot. Our team works hard to accommodate a variety of needs, so please get in touch and we’ll see what we can do.”
After: “Early check-in is available from 11 a.m. for a flat fee of $25, subject to room availability. Standard check-in is at 3 p.m. To guarantee an early room, book it at least 48 hours before arrival by emailing the front desk.”
The first version has no fact an AI could use. The second has a time, a price, a condition and an action. (Swap in your real policy, of course. Never publish details you can’t honor.)
Does Ranking on Google Mean You’ll Be Cited?
Not necessarily, but it helps. Retrieval systems draw on search indexes, so pages that are crawlable, indexed and trusted start with an advantage. Being on page one doesn’t guarantee your passage is the one quoted, though, and a lower-ranked page with a sharper answer can be cited ahead of you.
Technical health is the entry ticket: no accidental noindex tags, no blocked search bots, main text in the HTML rather than in images or scripts, and fast pages. In Apex®, the AI-native digital experience platform (DXP), we built these basics into publishing because the same problems repeat on almost every site we audit. You can see how it plays out in our case studies.
What’s Still Unproven?
Several popular claims about AI search have no solid public evidence behind them. Here’s our plain list of what we can’t verify.
- That any specific schema type raises your chance of being cited in AI Overviews. Google says none is required.
- The exact ranking signals ChatGPT, Gemini, Claude or Perplexity use to pick one source over another.
- That a particular word count, keyword density or heading count triggers citations.
- How much AI citations will affect revenue for a small business. It varies by niche, and attribution is messy.
What Should You Skip?
Skip anything that adds work without a clear mechanism. Most of these tactics are repackaged guesses.
- An llms.txt file as a magic fix. Google Search ignores these files. Other tools may read them, but they won’t replace clear pages.
- Stuffing keywords or repeating the question 10 times. The GEO research found it didn’t help, and it reads badly.
- Fake FAQ blocks. Only answer questions people actually ask.
- Paying for “AI citation guarantees.” Nobody controls these systems.
- Abandoning classic SEO. It’s still how pages get into the indexes that AI tools search.
What This Means for Your Pages
If citations come from passages that retrieval can find and trust, your job is simple to state: be indexable, answer clearly, prove what you say and be consistent across the web. To test whether it’s working, pick five questions your customers ask, run them in ChatGPT, Gemini, Claude, Perplexity and Google, record which sources get cited, make your changes and check again a few weeks later. Answers vary between runs, so treat one result as an anecdote and a pattern as a signal.
For the step-by-step changes, rated by effort and impact, use our GEO checklist. If you want help, our GEO and page speed services team can audit your site, and Apex® GEO intelligence tracks which sources get cited for your prompts over time.
Sources
- GEO: Generative Engine Optimization (Aggarwal et al., arXiv)
- What Are Large Language Models (LLMs)? (IBM)
- AI Features and Your Website (Google Search Central)
- Optimizing for Generative AI Features on Google Search (Google Search Central)
- OpenAI Bots Documentation
- Perplexity Crawlers
- Does Anthropic Crawl Data From the Web? (Anthropic)
