AI citation: definition and how it works

An AI citation is a link, footnote or reference that an AI search engine displays with its answer to show where information came from. ChatGPT search, Perplexity, Claude with web search, Google AI Overviews and AI Mode all show citations in some form.

A citation is a display decision. It records which sources the engine chose to credit. It does not record every page the engine read, and it does not measure how much any one page shaped the wording. For the full distinction, see our guide to cited vs retrieved sources.

Citations sit at the end of a pipeline with distinct stages.

1. Access. A crawler must be allowed to fetch and store your page. Blocked or unrendered content never enters the pool.
2. Retrieval. For a given query, the engine decides whether to search at all, then pulls a set of candidate pages. This is the set of retrieved sources.
3. Generation and citation. The model writes the answer and attaches citations to some of the retrieved pages. The cited set is a subset.

If the engine answers from its training data with no live search, there is usually no citation at all. Retrieval is discretionary per query on every major platform.

The citation unit also varies, as of October 2026:

• Claude (API web search): cites text spans, with a cited_text excerpt of up to 150 characters. See how Claude picks and cites sources.
• Perplexity: cites whole sources by an integer index.
• Google AI Overviews and AI Mode: show supporting links. Google says a page must be indexed and eligible to show with a snippet to appear, and that the two surfaces may use different models and return different links.

Because a span and a whole-page link are different things, citation counts from different platforms do not add up cleanly.

Citations are the part of an AI answer you can actually observe. They show which domains an engine credits for a topic, and they are where referral clicks come from.

They also concentrate. A 2025 study of AI Search Arena data (arXiv 2507.05301) found the top 20 news sources accounted for 67.3% of all citations for OpenAI models. When engines keep citing other sites instead of yours, being named on those sites is associated with appearing in the answer. Our playbook on finding the third-party pages AI engines cite covers how to identify them.

What is in your control is the access stage: crawler permission per bot, firewall reachability, content in the HTML before JavaScript runs, and index and snippet eligibility. What is not in your control is which retrieved page the model chooses to credit. The free checker on this site shows whether engines name you for up to three buyer questions, as a one-time snapshot.

"Cited means used"

Not reliably. A mechanistic study (arXiv 2606.28358) repaired over 90% of missed citations by amplifying specific attention heads, without changing the sources. Citation emission is partly separate from source influence. Our summary of the attribution mechanism study explains the method.

"Not cited means not read"

Engines retrieve more pages than they cite. A page can inform an answer and receive no credit.

"A citation proves the claim"

Citations can resolve and look relevant while not supporting the sentence they sit next to.

"Schema markup earns citations"

The evidence does not support this. Ahrefs' matched-control test of 1,885 pages found adding schema produced no major uplift on any platform, and AI Overview citations fell 4.6%. For what studies do and do not show, see LLM citation optimization.

"Citations have a rank"

Google's own AI report exposes impressions only. No position data exists to check an "AI rank" against.

• Retrieved sources: every page the engine pulled for a query, of which citations are a subset.
• Prompt tracking: repeatedly asking engines set questions and logging which sources they cite.
• Grounding: answering from live retrieved content rather than training data alone.
• Knowledge cutoff: the date after which a model's training data stops, which governs answers given without retrieval.

Why do different AI platforms show citations differently? Each vendor built its own attribution format. As of October 2026, Anthropic's API cites short text spans with an excerpt of up to 150 characters, Perplexity cites whole sources by number, and Google shows supporting links in AI Overviews and AI Mode. Because these units differ, a citation count on one platform is not directly comparable with a count on another. Compare platforms separately rather than summing them.

Can I see which sources an AI engine looked at but did not cite? On some developer APIs, yes. OpenAI, Anthropic and Perplexity expose the retrieved set separately from the cited set, and the retrieved set is usually larger. Consumer apps generally show only citations. If you want to compare the two, log both fields through the API with search enabled and record which pages were retrieved but dropped.

Does an AI answer always include citations? No. Engines decide per query whether to search the web. When they answer from training data alone, there is usually nothing to cite. ChatGPT searches at the model's discretion, and web search on Claude.ai is a user toggle. A citation therefore tells you retrieval happened for that answer. An answer without citations may come from a training snapshot with an older knowledge cutoff.

What can I do to make my pages citable? Remove blocks. Allow the retrieval crawlers you want in robots.txt (for example OAI-SearchBot for ChatGPT search), confirm your firewall or CDN does not return errors to them, put key content in the HTML before JavaScript runs, and avoid accidental noindex or nosnippet directives. Stating specific facts, such as a price, on a fetchable page gives an engine a sentence it can quote. None of this guarantees a citation.