Which Sources Does AI Cite? Anatomy of LLM Citations

AI assistants cite the sources their retrieval system actually fetched for a given question, not "the best website on the internet". In practice that means pages that are crawlable, indexed by the search layer the assistant uses, closely matched to the wording of the query, and written so that a single passage can support a single claim. Which types of sources win varies by assistant, market, language and question, so the only reliable answer for your brand comes from measuring real answers.

Key takeaways

A citation is produced by a retrieval step (search, ranking, passage selection), not by the model "remembering" a website; answers built purely from training data usually carry no links at all.
Being mentioned and being cited are different outcomes: a brand can appear in the text with no link, or a page can be cited without the brand being named.
Cited sources tend to be pages that answer the question directly in an extractable passage: documentation, reference pages, comparisons, reviews, forums and news.
Citations are not always accurate. Academic audits have found many generated sentences that the cited page does not fully support.
Treat citation share as a statistic: sample many prompts, repeat them, exclude failed answers and compare competitors on the same sample.

How does an AI assistant choose which sources to cite?

Large language models have two very different ways of "knowing" something. Parametric knowledge is what the model absorbed during training; it is compressed into weights and cannot point to a URL. Retrieved knowledge is fetched at answer time from a search index or the live web and placed into the model's context. Only the second kind can be cited honestly. This architecture is usually called retrieval-augmented generation (RAG), a term popularised by Lewis et al. (2020).

A typical citation pipeline in an answer engine looks like this:

Query understanding and rewriting. The user's question is turned into one or several search queries. Google, for example, describes a "query fan-out" technique in AI Mode that issues multiple related searches across subtopics.
Retrieval. A search index returns candidate documents. Which index matters: Google's AI features draw on Google Search, while ChatGPT search relies on crawlers such as OAI-SearchBot plus search partners, and Perplexity runs its own PerplexityBot.
Ranking and passage selection. Candidates are scored and cut into passages; only a handful of passages fit into the model's context window.
Generation with attribution. The model writes the answer and attaches citations to the passages it relied on; the interface then decides how many links to show and where.

Every step is a filter: a page blocked for the relevant crawler never reaches step 2, and a page that buries the answer in paragraph twelve often loses at step 3.

Which types of sources get cited most often?

No public dataset covers all assistants, markets and languages, so be wary of universal "top cited domains" lists. What can be said is which properties make a source easy to retrieve and quote; the examples below are illustrative, not measurements.

Source type	Why retrieval favours it	Typical query (example)
Official documentation and product pages	Authoritative, specific, stable URLs	"Does tool X support SSO?"
Reference and encyclopedic pages	Clear definitions, broad coverage	"What is net revenue retention?"
Comparison and "best of" articles	Match list-style intent directly	"Best CRM for small agencies"
Forums and community Q&A	First-hand experience, long-tail phrasing	"Is brand Y worth it after a year?"
News and trade media	Freshness for time-sensitive questions	"Latest pricing change at company Z"
Research papers and statistics pages	Numbers and methods that support claims	"How accurate are AI citations?"

Two consequences follow. Much of your citation footprint lives on third-party pages you do not own, and the mix shifts with question type: definitions reward reference pages, "which brand should I buy" rewards comparisons and community discussion. We explore the recommendation side in how ChatGPT recommends brands, and platform-specific behaviour in visibility in Perplexity and our guide to Google AI Overviews and AI Mode.

Mention versus citation: three outcomes worth separating

When you analyse answers, keep three outcomes apart:

Mention without citation. The brand name appears in the text but no link points to your domain. This often reflects parametric knowledge or third-party pages that discuss you.
Citation without mention. Your page is listed as a source, yet the answer never names your brand. This is common for informational content: you supplied a definition, a competitor got the recommendation.
Mention with citation. The strongest outcome, because the user sees the name and has a path to your site.

Mixing these into one score hides the diagnosis: many mentions with few citations calls for better owned content, while many citations with few mentions means you educate the market without capturing demand. Our glossary entries on AI citation and presence rate define these metrics precisely.

How reliable are AI citations?

A citation looks like proof, but it is only a pointer. Liu, Zhang and Liang (2023) audited four generative search engines and found that, on average, only 51.5% of generated sentences were fully supported by their citations, and only 74.5% of citations actually supported the sentence they were attached to. In 2025 the Tow Center for Digital Journalism at Columbia tested eight AI search tools on identifying the source of news excerpts and reported incorrect answers to more than 60% of queries, including links to syndicated copies and URLs that did not work.

For brands this cuts both ways: you may be cited for a claim your page never made, and "we were cited" is not the same as "the answer was correct". Good monitoring stores the full answer together with its sources, so a human can check what was actually attributed.

What increases the probability of being cited?

There is no guaranteed placement. The evidence points to a few levers that improve the odds:

Be retrievable

Allow the crawlers that feed answer engines (for example OAI-SearchBot for ChatGPT search, PerplexityBot for Perplexity) if you want to appear there, and make sure important pages are indexed. Google states that there are no extra technical requirements for AI Overviews beyond being indexed and eligible to show with a snippet.

Write extractable passages

Answer the question in the first sentences of a section, keep one claim per paragraph and use descriptive headings. In the GEO study by Aggarwal et al. (KDD 2024), methods such as adding quotations, statistics and source citations raised the visibility of a source in generated answers by up to roughly 40% on their benchmark, while keyword stuffing did not help. We cover writing technique in depth in how to write content AI quotes.

Earn third-party coverage

Since many citations go to reviews, comparisons and communities, accurate presence on those pages matters as much as your own site. This is where GEO overlaps with classic digital PR.

Keep facts consistent and current

Outdated prices or conflicting specs across pages make passages less useful as evidence and can lead to misattribution.

How do you measure which sources AI cites for your topic?

Treat citations as a noisy outcome. Define a prompt set that reflects real buyer questions in your market and language, run it on several assistants, repeat it over time and store every answer with its sources. Then report the share of valid answers that cite your domain, which domains are cited instead of you, and how often you are mentioned versus linked. Failed or empty answers belong outside the denominator, and competitor comparisons only mean something when they come from the same sample. Our article on how to measure AI visibility explains sampling in detail, and our methodology page documents how GeoDog applies these rules. If you want to see the cited domains for your own questions, GeoDog's citations report lists every source attached to every stored answer.

FAQ

Does ChatGPT cite sources in every answer?

No. Links typically appear when the assistant performs a web search. Answers generated from training data alone usually contain no citations, and whether search is triggered depends on the question, the product settings and the model.

Can I pay or submit my site to be cited by AI?

There is no paid placement for organic citations in the major assistants as of this writing. What you can control is crawl access, indexation, content quality and your presence on third-party sources.

Why does the same question cite different sources each time?

Retrieval, query rewriting and generation are not fully deterministic, and indexes change, so single screenshots prove little.