How AI models choose sources: what research says about citations

When an AI model answers a question about, say, the best heat pump or a reliable accountant, it often pulls information from the web and cites a few sources. Which sources those are largely determines which businesses end up in the answer. In this article, we summarize what academic research and large citation studies say about this. We flag what was measured and what's our interpretation.

The GEO study: what helps you get cited

The term generative engine optimization (GEO) comes from a paper by Pranjal Aggarwal and colleagues, including researchers at Princeton University, presented at the KDD conference in 2024¹. They built a benchmark of 10,000 queries across domains, GEO-bench, and tested how changes to a web page's text affect visibility in AI answers.

The headline finding: relatively small changes boosted visibility by up to 40%¹. The best-performing methods were:

• Cite sources: add references to credible sources.
• Add quotations: include relevant quotes from credible parties.
• Add statistics: replace qualitative claims with concrete numbers.

These three delivered a relative improvement of 30 to 40% on the main visibility metric, and 15 to 30% on a subjective measure of how prominent the source is in the answer¹. Keyword stuffing, the classic SEO move of repeating query terms, performed poorly¹.

For small businesses, a second finding is at least as interesting. Websites that ranked lower in search results gained the most. Citing sources gave a site ranked fifth a 115% increase in visibility, while the top-ranked site's visibility dropped by 30.3% on average¹. The researchers point out that many lower-ranked sites belong to small creators and independent businesses. They also found that what works differs by domain¹.

One caveat: this is a controlled experiment, not a measurement of ChatGPT or Gemini in the wild, and the AI models have changed since. But the direction, that factual, well-supported and well-structured text gets used more often, is consistent with what other studies see.

Why would these changes work? A plausible explanation, and this is our interpretation, is that an AI model builds an answer from snippets of text that answer a question directly and verifiably. A line like "we deliver fast, expert service" gives an AI model little to work with. A line like "installation within three weeks, 412 installs in 2025, average rating 4.7 from 380 reviews" does. Concrete, specific information is easier to cite and easier to check against other sources. The numbers in this example are made up, of course.

How closely do AI citations match Google's rankings?

This varies a lot by system. Ahrefs analyzed 1.9 million citations from 1 million Google AI Overviews. Of those, 76.1% came from pages in Google's top 10, and 86% from pages somewhere in the top 100². For AI Overviews, a strong Google ranking remains a solid foundation. Still, about 14% of cited pages weren't in the top 100 at all: you can show up as a source without a high ranking.

Standalone AI assistants are a different story. In a study of 15,000 prompts, Ahrefs found that on average only 12% of links cited by ChatGPT, Gemini and Copilot ranked in Google's top 10 for the same prompt, and 80% of those citations didn't rank in Google at all for that query³. Perplexity is the exception: 28.6% of its cited URLs were in the top 10³. One explanation Ahrefs offers: AI assistants turn the user's question into their own, often different, searches and use different indexes.

Even within Google, things differ. In an Ahrefs analysis of 730,000 response pairs, AI Mode and AI Overviews gave substantively the same answer 86% of the time, but cited the same URLs only 13.7% of the time⁴.

Why AI assistants find different pages

The small overlap with Google comes down to how AI assistants search. They don't paste the user's question into a search engine word for word. They first decide whether to search, and if so, with which queries. Ahrefs describes how ChatGPT appears to use a classifier that only triggers a web search when the answer is hard to predict from training data³. It then often runs several narrower searches, known as fan-out. The pages found that way often don't rank for the original question.

The AI model's training knowledge plays a role too. A business that's mentioned often and consistently across the web can end up in an answer without any source attached. So citations show part of the story: which sources the AI model used at that moment. Why it already knew a certain name stays partly hidden. That's a limitation of any AI visibility measurement, including ours.

Which kinds of sources get cited?

Pew Research Center found that Wikipedia, YouTube and Reddit are the most frequently cited sources, both in Google's AI summaries and in standard results. Government (.gov) websites appeared more often in AI summaries (6%) than in standard results (2%), and 88% of summaries cited three or more sources⁵.

Semrush tracked more than 230,000 prompts in ChatGPT search, Google AI Mode and Perplexity for 13 weeks, covering more than 100 million citations. Reddit and LinkedIn ranked among the top five most-cited domains on all three⁶. The same study showed how fast this shifts: ChatGPT cited Reddit in close to 60% of responses in early August 2025, and in around 10% by mid-September⁶.

For local businesses, your own website carries a lot of weight. BrightLocal found that for local queries, ChatGPT search used business websites for 58% of its sources, followed by business mentions on other sites (27%) and online directories (15%)⁷. But here too, things go wrong when the basics are off: Adobe estimated that 39% of the homepages of a group of US online retailers weren't machine-readable for AI models⁸. The gap by category was large: apparel scored 76% readable, grocery 59%. Adobe ties better scores to consistent, structured product content⁸.

What stands out across these studies: AI models rarely cite just one source. They combine a reference work, a forum thread, a comparison site and a business website into one answer. A business that's only well described on its own website lacks outside confirmation. A business that's often mentioned elsewhere but has an unclear website lacks the facts. You're in the strongest position when both are right and back each other up.

What this means for your business

Taken together, the studies suggest a few practical conclusions. They're interpretations, not guarantees.

• Write in facts. Concrete numbers (prices, lead times, volumes, certifications), clear definitions and references to credible sources match what worked best in the GEO study.
• Make your website machine-readable. Key information belongs in plain text on the page, not just in images, PDFs or scripts. Structured data and consistent contact details help.
• Show up beyond your own website. Because AI assistants often cite sources outside Google's top 10, mentions on comparison sites, directories, trade media, forums and review platforms count.
• Look at each AI model separately. Perplexity leans more on Google's rankings than ChatGPT does; AI Mode cites different pages than AI Overviews. What works for one AI model may not work for another.
• Expect shifts. ChatGPT's drop in Reddit citations shows source preferences can change within weeks. Optimizing once isn't enough.

How GeoReady measures this

For every prompt and every AI model, GeoReady shows which sources ChatGPT, Perplexity, Gemini and Claude cite, down to the URL. You see whether your own pages get used, which third-party sites name your competitors and where you're missing. Tracking runs through the official APIs, 3× per week, with repeated runs. Alongside citations, you see your win rate, share of voice and sentiment with the source sentence. GeoReady turns that into actions, like adding facts to a page or earning a mention on a frequently cited source, and shows the measured effect in your GeoReady score after 30 days.