How LLMs Choose Which Sources to Cite

It's Not Traditional Search

When Google shows classic results, it gives you a list of pages to choose from. The final decision is yours.

When an LLM answers a question, it pulls information from several sources and hands you one response. There's no list to pick from. The model has already decided which information to use, and often which sources to show as citations.

That changes the rules for anyone who publishes.

The Selection Process

LLMs with web access run something close to a search, then treat the results differently. The details vary by product, and the vendors document more of them than they used to.

First, the system decides whether to search at all. Anthropic's docs say Claude searches when a request depends on current or changing information, and answers stable-knowledge questions directly. When a model answers from its training data alone, you get no citations.

Then it turns your question into queries. ChatGPT typically rewrites a prompt into one or more targeted queries and sends them to its search partners, sometimes followed by more specific ones. Google says AI Overviews and AI Mode may use "query fan-out", issuing several related searches across subtopics. So your page competes for sub-questions you never typed into a keyword tool.

Next, it reads the candidates and writes the answer. It doesn't paste whole paragraphs. It extracts and rephrases.

Finally, it shows sources, and how it does that depends on the product. Perplexity's help center says every answer carries numbered citations linking to the original sources. In Anthropic's API, citations are always on when Claude uses web search. ChatGPT responses that use search "may include citations," in OpenAI's words.

There is also a filter before any of this: access. OpenAI says sites that block OAI-SearchBot won't be shown in ChatGPT search answers. Perplexity asks you to allow PerplexityBot, and Anthropic runs a separate Claude-SearchBot for search. For Google, a page has to be indexed and eligible for a snippet; Google says no special files, markup or schema are needed for its AI features.

What Makes a Source Get Chosen

Clarity of information. LLMs handle well-structured content with plain statements better. Vague or confusing text is less useful to them.

Perceived authority. Recognized sources in a field carry more weight. That is built over time, not through shortcuts.

Timeliness when it matters. For topics where information changes, up-to-date sources get preference. It's the same reason Claude reaches for search on current questions, and Microsoft's Bing team tells publishers to keep content fresh and accurate so AI answers reference the latest version.

Consistency with other sources. Information that shows up the same way across several reliable sources is more likely to be used.

What Doesn't Work

Traditional keyword optimization. LLMs work with meaning, not word matches. The research backs this up: in the GEO study presented at KDD 2024, keyword stuffing gave little or no gain in generative-engine visibility, and on Perplexity it did about 10% worse than the unedited page. Adding citations, quotations and statistics raised visibility by up to 40% on the authors' benchmark. That was a 2023-24 test setup, so treat it as direction, not a formula.

Padding content to look more complete. More words don't equal more authority.

Ranking manipulation tactics. LLMs don't use the same ranking systems as traditional search engines, and none of them promise a citation. Google says indexing and serving aren't guaranteed; OpenAI says placement in ChatGPT search isn't either.

Strategies That Do Work

Build real authority in your field. Get recognized as an expert by other trusted sources.

Create citeable content. Plain statements, specific data with its source, and precise definitions an LLM can extract and reuse.

Keep your information consistent. Your website, profiles and outside mentions should tell the same story, which is most of what building your brand as an entity means.

Structure content so it's easy to process. Clear headings, focused paragraphs, lists where they help. Structured data that matches the visible text helps machines read the facts without guessing.

Check whether any of it works. A fixed set of questions, logged every month, tells you more than any single test. Here's how to measure your visibility in ChatGPT and Perplexity.

The Indirect Human Factor

LLMs learn from content people write. When people cite you, reference you or discuss you, that shapes how models describe you.

GEO doesn't replace building a real reputation. It speeds it up and makes it easier to see.

LLMs are changing how information reaches people. Knowing how they pick sources is the first step toward being picked.

There are no shortcuts. The base is still real authority, useful content and a consistent presence. GEO is about how that base is presented to the new intermediaries.

If you're not sure whether the gap is your brand, your site or your content, start with a free audit. We look at where the system breaks before recommending more content.