How do AI assistants decide which sources to cite?

Practice owners usually ask this question after they have been left out. They have a better website than the practice down the street, and the assistant cited the other one anyway. The answer is rarely about the website alone, and the mechanics explain why.

How does an AI assistant find sources in the first place?

It searches. Google says AI Overviews and AI Mode use retrieval-augmented generation: its ranking systems pull relevant pages from the Search index, and the model writes from them. OpenAI says ChatGPT rewrites a question into one or more targeted queries and sends them to search providers. The answer is written from what those searches return.

Google calls the rewriting step query fan-out: one question becomes a set of related searches running at once. OpenAI gives a similar example, where a broad question turns into a specific search, then a second, narrower one after the first results come back. The practical meaning for a practice is that the assistant may never search the words your patient typed. It searches for the sub-questions it thinks it needs answered.

What has to be true before a page can be cited?

Four gates, in order. The page has to be reachable by the assistant’s crawler, present in the index it searches, retrieved for one of the rewritten queries, and then chosen as the page that supports a sentence. Most practices that never get cited are failing the first or second gate, which no amount of writing fixes.

The vendors document the first two gates plainly. Google says a page must be indexed and eligible to show with a snippet to appear in its generative AI features. OpenAI says a site that blocks OAI-SearchBot will not be shown in ChatGPT search answers, though it can still appear as a plain link. Perplexity says PerplexityBot exists to surface and link websites in its results.

What kind of content gets cited most?

Content that gives the model something specific to repeat. The peer-reviewed GEO study at KDD 2024 tested content changes on a 10,000-query benchmark. Citing sources, adding quotations and adding statistics raised visibility in generative answers by roughly 30 to 40%. Keyword stuffing produced little or no improvement.

One result from that paper should interest smaller practices. The gains were largest for pages that ranked lower. With the cite-sources change, a site ranked fifth in the search results gained 115% in visibility while the top-ranked site lost 30%. Being the biggest name in town doesn’t lock the answer. Being the clearest, best-sourced page on a narrow question can move you into it.

Microsoft gives similar advice to publishers in its Bing Webmaster Tools guidance: clear headings, tables and FAQ sections, claims backed by evidence and data, and pages kept up to date. That is a search company telling you what its assistant finds easy to cite, and it lines up with the research.

Do assistants prefer other websites over the practice’s own site?

Often, yes. A 2025 study by Chen and colleagues compared AI search citations with Google results and found a systematic, strong bias toward earned media, meaning independent third-party sources, over brand-owned pages and social content. Google’s own results showed a more balanced mix. That paper is a preprint, not yet peer reviewed.

For a practice, earned media means the directories, local news, hospital and association listings, and review platforms that talk about you without you writing the words. It explains the complaint I hear most. Your site can be excellent and still be the second source an assistant reaches for, behind a directory that describes you in one line. That line had better be right.

Does ranking on Google mean an assistant will cite you?

Less than you would expect. Ahrefs compared citations from ChatGPT, Gemini, Copilot and Perplexity across 15,000 long-tail queries and found only 12% of links cited by ChatGPT, Gemini and Copilot were in Google’s top ten for the same prompt. Perplexity overlapped more, at about 28.6%. That is a 2025 vendor study.

Some of the gap comes from fan-out. The assistant is searching sub-questions you never ranked for, and citing whichever page answers each one. The rest comes from selection. A page that ranks well on a broad term but never states a specific fact in a sentence can be retrieved and still not be cited.

What can a practice actually influence?

Gates one and two are yours to fix outright, and they are the cheapest. Gate three improves when each important page answers one question a patient would type, in the words a patient uses. Gate four improves when your pages carry specific, sourced statements and when independent sites say the same things about you that you say about yourself. None of it is guaranteed. OpenAI says so directly: placement is not guaranteed. What you can do is stop failing the gates you control.

You can watch part of gate four from the outside. Bing Webmaster Tools now reports how often Microsoft Copilot and Bing’s AI summaries cite your pages, which URLs they cite, and the grounding queries the AI used to find them. It covers one family of assistants, and it is the clearest view of citation any vendor offers today.