AI's own search decides who gets cited, not page tweaks, study finds

What did the study find?

The study found that which pages an AI engine cites depends far more on what its own search returns, and on what the model already believes about brands, than on how individual pages are written. The paper, What Drives AI Citations?, by Mahe Chen and Nick Koudas of the University of Toronto, was posted on arXiv on 10 October 2026.

Four findings matter most for anyone trying to be recommended by AI.

• The engine's own search is the gate. Every page ChatGPT cited had first been returned by its own search, and 97% of the pages it retrieved did not appear in Google's top 50. Ranking well on Google is no guarantee of being in the pool AI chooses from.
• Retrieval outweighs page tweaks. Comparing pages on the same website, what ChatGPT's search returned explained about 32% of the difference in how often pages were cited. All the page features the researchers measured explained 2.8%.
• Popular advice looks better than it is. Practices often recommended for AI visibility, such as question-and-answer sections and summary boxes, were linked to more citations when different websites were compared, but showed little or no effect between pages of the same website. The authors conclude that these practices tend to sit on websites already cited more for other reasons.
• What the model already knows counts. 47% of ChatGPT's citations came from websites its model names without searching at all. And 46% of its mentions of brands the question did not name were not supported by any page it cited.

The paper does find one page-level effect that holds: pages whose title and best paragraph closely matched the question asked were cited more often. The effect was larger on third-party pages than on a company's own.

How was the study done?

The researchers asked ChatGPT, Gemini and Perplexity the same questions once a day for 90 days, through each company's public developer interface with web search switched on, then analysed which pages each answer cited.

The main collection ran from 24 June to 21 September 2026. The researchers then replayed questions through ChatGPT and Gemini to see every page each engine retrieved, not just the ones it cited, and ran controlled experiments in which they changed only the names attached to otherwise identical sources.

The limits are stated plainly. The paper is a preprint and has not yet been peer reviewed. The results are associations, not proof of cause. Both markets were English-language questions about the US. The replay used developer interfaces that may not match the consumer apps, Perplexity could not be replayed, and the experiments used the smaller models behind the engines. The B2B dataset is small: in it, retrieval still explained more than page features, but the difference was not statistically significant.

What does it mean for B2B shortlists?

It means a company's place on an AI shortlist is decided mostly upstream of its own pages: by whether the engine's search finds it, and by what the model already associates with its name. Rewriting pages helps at the margin; it does not change which sources the engine looks at.

The naming experiment makes the point sharply. With everything else held equal, one company profile reached ChatGPT's top three in 80% of answers under one name and 39% under another. What the model already believes about a name can move a supplier on or off the shortlist before the evidence is weighed.

Platforms also diverge. In the B2B market studied, the model behind ChatGPT named the tracked vendors for 2% to 16% of questions, while the model behind Gemini named them for 60% to 96%. A single headline figure across platforms would hide that difference.

And the pattern moves. Comparing pages within the same websites, only 14% of the citation pattern for ChatGPT, 29% for Gemini and none for Perplexity was still shared ten weeks later. A one-off check of who AI cites is out of date within a quarter.

What does it mean for companies in Ireland and the UK?

It means Irish and UK suppliers should not assume that strong Google rankings or a well-written website put them in front of AI. The study measured US questions only, so its figures are not Irish or UK measurements, but the mechanisms it describes belong to the assistants themselves.

Two risks follow for smaller suppliers. First, if the model knows a larger UK or international competitor better, that prior shapes both what it searches for and what it cites. Second, the sources AI's search returns in a local market may be different from the ones a company tracks on Google. Both are worth checking directly, in the market you sell into.

What should leadership teams do this quarter?

Start by finding out which sources AI actually uses when your buyers ask about your category. Four steps follow.

1. Look at the sources, not just the answer. Ask the questions a buyer would ask, in a fresh chat on more than one platform, and record which websites are cited. Compare them with your Google results; the study suggests they will differ.
2. Test what AI knows about you before it searches. Ask what it knows about your company and your category leaders. Gaps or errors there point to the independent sources that describe your market.
3. Answer the buyer's question in plain words where it is asked. On your own pages and, above all, on the third-party pages AI cites, the title and opening paragraph should match the question a buyer types.
4. Measure each platform separately, and repeat. ChatGPT and Gemini named vendors at very different rates in this study, and the pattern shifted within weeks.

A Gap Check report shows how far you are behind the most-named of your top three competitors, in your market's own terms. Send your website, the market you want measured and your top three competitors for that market, and Genivista sends the report, at no charge, walked through with Neil.