How AI engines choose what to cite: the evidence, not the folklore.

Every week someone publishes "the 27 AI ranking factors". Few of them cite anything. Here is what the studies we trust actually found, and the limits of each.

1. Retrieval comes first

Assistants that search the web can only cite what their searches return. Ahrefs analysed 1.4 million ChatGPT prompts and found roughly half of retrieved URLs ended up cited. URLs that came from search results were cited 88.46% of the time; Reddit URLs, 1.93%. Classic indexing and ranking are the entry ticket.

2. The title has to match the question

In the same study, cited page titles were more similar to the prompt than non-cited ones (0.602 against 0.484), and even closer to ChatGPT's own background searches (0.656). Natural-language URL slugs were cited 89.78% of the time against 81.11% for opaque ones. Write titles and slugs the way people phrase questions.

3. Statistics, quotes and sources help

The GEO paper by Aggarwal and colleagues (Princeton and IIT Delhi, KDD 2024) tested content changes on a 10,000-query benchmark. Methods like adding statistics, quotations and citations boosted visibility in generative answers by up to about 40%, with results varying by domain. Keyword stuffing did not help.

4. Some formats are cited more

Otterly's study of about one million URLs found pages with /guide/ in the URL averaged 2.7 AI citations, against a 1.9 average; blog posts averaged 2.0 and pricing pages 1.5. That is correlation, not proof, but it matches what we see: complete, structured guides get lifted.

5. Freshness matters, unevenly

Ahrefs' freshness research reports AI-cited content is about 25.7% fresher than Google's organic results, with ChatGPT leaning newest and AI Overviews slightly older than organic. Yet in the 1.4M-prompt study, the median cited search result was about 500 days old. Relevance first; freshness breaks ties, and decides news.

6. Brand mentions beat links for AI Overviews

Across 75,000 brands, Ahrefs found branded web mentions correlated 0.664 with AI Overview visibility, against 0.218 for referring domains. Being talked about matters more than being linked to.

Folklore the evidence doesn't support

The short version

1. Get indexed and ranking in Google and Bing.
2. Title every page as the question it answers, and answer it in the first paragraph.
3. Add numbers you measured, named sources and quotes.
4. Update important pages with real changes, and show the date.
5. Get mentioned by others in your topic.

Want to see what AI answers already cost you in clicks? Our zero-click calculator estimates it from your Search Console data.

Questions people ask

How does ChatGPT choose which sources to cite?

It cites pages its searches retrieve, favouring results whose titles closely match the question and that answer it clearly. In Ahrefs' study, retrieved search results were cited 88% of the time.

Does llms.txt help AI citations?

There is no evidence it does. Log studies found AI bots almost never request it and saw no citation lift, and Google says no AI-specific files are needed.

What content gets cited most by AI engines?

Pages that rank, match the question in their title, answer it directly, and include statistics, quotes and sources. Guides are cited more often than blog posts or pricing pages.