How do AI citations work in ChatGPT, Claude and Grok?

Published 31 October 2026. Updated 3 November 2026.

An AI citation is a link that ChatGPT, Claude or Grok attaches to its answer to show where a claim came from. Each assistant cites pages its own web search found during that turn, and it cites only some of the pages it read.

This page covers how each assistant shows its citations, how many of the pages it reads end up cited, and what published studies say about which pages get picked. The search that comes first is covered in how ChatGPT, Claude and Grok search the web, and our guide to query fan-out explains how one prompt turns into several searches.

What is an AI citation?

An AI citation links a sentence in an assistant's answer to a web page the assistant read while answering. It comes at the end of a longer process. The assistant writes search queries, gets results back, opens some pages and then credits a few of them in the text.

That makes a citation different from a mention. An assistant can name a brand or a product without linking any page about it, and it can read a page closely without crediting it. People searching for "AI citations" sometimes mean how to cite ChatGPT in an essay. That is a question for style guides, and this page does not cover it.

How do ChatGPT, Claude and Grok show their citations?

All three cite inline, next to the text that relies on a source. Each also keeps a longer list of sources behind the answer, and the companies' own docs describe that list as fuller than what gets cited.

Assistant	How citations appear	What the docs say about the longer list
ChatGPT	Inline links in the answer, plus a Sources button	The sources field "returns the complete list of URLs the model consulted", and "The number of sources is often greater than the number of citations."
Claude	Citations attached to the sentences that use them, each with up to 150 characters of quoted text in the API	"Citations are always enabled for web search."
Grok	Numbered inline links, starting at 1 and reused when a source comes up again	"Not every URL in this list will necessarily be directly referenced in the final answer."

The quotes come from OpenAI's web search guide for developers, Anthropic's web search tool documentation and xAI's citations documentation. OpenAI's guide also says inline citations "show only the most relevant references", and that apps built on the API must make them "clearly visible and clickable". In the ChatGPT app, OpenAI's help article on searching the web says the Sources button shows "cited sources and other relevant links".

In conversations recorded with Only Fan Out, ChatGPT's product cards for shopping questions sit apart from its cited sources, and Claude plots places on a map without citing them. For Grok, the extension resolves each numbered citation back to its source.

How many of the pages an AI reads does it cite?

For ChatGPT, about half. Ahrefs looked at 1.4 million ChatGPT prompts on desktop and recorded every URL ChatGPT retrieved. In Ahrefs' study of why ChatGPT cites one page over another (published 15 April 2026, updated 27 August 2026), 49.98% of retrieved URLs were cited.

Ahrefs adds a caveat: some URLs that were not cited may never have been opened, so the figure covers the whole path from search result to citation. The split also depends on where the URL came from. Pages from ChatGPT's search results were cited 88.46% of the time, while Reddit URLs were cited 1.93% of the time and made up 67.8% of the URLs that went uncited.

We have not found a comparable published figure for Claude or Grok. The query fan-out guide explains the difference between a fetched source and a cited one, and why the fetched list is nearly always the longer of the two. Our page on fetched vs cited sources lines up the published gap figures by what each study counted as fetched.

Does a page have to come up in the assistant's own search to be cited?

For ChatGPT, yes. Mahe Chen and Nick Koudas studied 158,099 answers and 468,037 citations from ChatGPT, Gemini and Perplexity, collected over two 90-day periods of commercial questions. In their paper, "What Drives AI Citations?" (submitted 10 October 2026), they replayed ChatGPT's searches through its API and compared the results with Google and Bing.

Every page ChatGPT cited was one its own search had retrieved. Which pages that search returned explained far more of the difference between cited and uncited pages than the pages' own properties did: 32% against 2.8%. The sub-queries ChatGPT wrote rarely matched what Google or Bing returned for the same question, so a high Google ranking for the prompt does not mean a page will turn up in ChatGPT's search.

The authors say the findings held in a second market of small B2B vendors, and they released their data and code.

What makes one retrieved page more likely to be cited than another?

Two studies found that pages whose titles and text match the questions being asked do better. Both describe correlations in closed systems, so treat them as signals worth testing on your own pages.

Study	Setup	Finding
Ahrefs, 2026	1.4 million ChatGPT prompts; title similarity measured with open-source embedding models	Cited pages had titles closer to the prompt (0.602 against 0.484 for uncited pages) and closer still to the fan-out query (0.656)
Ahrefs, 2026	Same data, search results only	Pages with natural-language URL slugs were cited 89.78% of the time, against 81.11% without
Chen and Koudas, 2026	158,099 answers across ChatGPT, Gemini and Perplexity	Within one website, pages whose titles and paragraphs match the questions asked are cited more often
Chen and Koudas, 2026	Same data	FAQ blocks, structured data, summary boxes and page speed "appear effective only across websites", not between pages of the same site

Ahrefs notes that ChatGPT is closed, so its embedding scores only approximate how ChatGPT itself compares text. Neither study shows how OpenAI, Anthropic or xAI choose sources internally.

Can an AI name a brand without citing a page?

Yes. Chen and Koudas found that ChatGPT writes brands it already knows into its search queries, prefers websites it knows, and names brands with no page behind the mention. In controlled tests, brand names alone changed how ChatGPT ranked products that had identical facts. Their summary: "Pages are thus cited when searched; brands are also named when remembered."

For anyone measuring AI visibility, that means citations and mentions are two separate counts. A citation shows that a page was found and used. A mention without a citation shows what the model already believed about a brand.

How can you see which pages an AI cited?

In the chat apps you see the cited links and, in ChatGPT, the Sources button. The full list of pages the assistant read stays out of view, and so do the queries it ran. Our walkthrough on how to see what ChatGPT searched covers the ways to recover ChatGPT's queries since it stopped saving them in September 2026.

Only Fan Out is a free Chrome extension that shows, for each turn in ChatGPT, Claude and Grok, the queries the assistant ran, every page it fetched and which of those it cited, side by side in a panel next to the chat. Add a brand or domain to its watchlist and it highlights every mention across queries, fetched sources, cited sources and answers. You can export a turn or a whole conversation as CSV, Markdown or JSON. Nothing leaves your browser.

What can a site owner do about AI citations?

No one can promise citations, and the extension does not change them either. The studies above suggest where to look first:

• Find the queries the assistant writes for your topic and check whether your pages rank for them. A page that its search never returns cannot be cited.
• Write titles and opening paragraphs that answer the question in plain words, so the page matches what is being searched.
• Let OAI-SearchBot and Claude-SearchBot crawl the site. The crawler table in how ChatGPT, Claude and Grok search the web lists which bot does what.
• Track mentions and citations separately, since an assistant can name your brand without linking you.

Our GEO workflow with fan-out data sets these checks out as six steps you can repeat each month.

Related questions