How AI Search Engines Cite Sources

By Bill Hartzer · Updated 19 August 2026

What a citation in an AI answer actually is

A citation in a generated answer is a record that a document was retrieved and was available to the model while it wrote. It is a claim about provenance. It is not a quotation, not a page reference in the academic sense, and not a guarantee that the sentence it is attached to reflects what that document says.

This distinction is the single most important thing to understand about the format, because the visual grammar of a small numbered marker after a sentence is borrowed directly from academic writing, where it means something much stronger. In a paper, a citation means an author read a source and is asserting that it supports the claim. In a generated answer it usually means a retrieval system returned that URL, a model wrote a paragraph while conditioned on some retrieved text, and an attribution step associated the two afterwards.

Perplexity introduced the format when it launched on 7 December 2022, and it was a genuine contribution — an answer with sources is better than an answer without. Everyone copied it, correctly. But the copying spread the visual convention faster than it spread the caveat. A whole vocabulary has grown up around trying to influence those citations, usually under the banner of answer engine optimization, and it is worth separating what that term describes from what it actually delivers.

Where in the pipeline citations are produced

Four stages sit between a question and a cited answer, and citations are attached at the last one.

Query rewriting. The system composes its own search query from the prompt. Microsoft documents this explicitly for Copilot; OpenAI says queries are rewritten before being sent to search providers.
Retrieval. The rewritten query hits a web index and returns a candidate set — typically far more documents than the model will use.
Selection and reading. Some candidates are actually fetched and read; most are not. This split has measurable consequences. Independent measurement of ChatGPT's retrieval stack published on 17 August 2026 found that pages the system actually opened were cited about 74% of the time, against about 7% for pages that were only retrieved into the candidate set.
Generation and attribution. The model writes prose conditioned on the selected passages, and citations are attached to sentences.

The same study found another detail worth knowing: the snippets ChatGPT stores are query-independent, cut at index time at roughly 200 characters following the page's main heading. That is a materially different object from a classic search snippet, which is assembled per query from the passage that matched. It means the text a system holds about a page may have nothing to do with the question being asked.

How the major products do it

The implementations differ more than the uniform look suggests.

Perplexity — numbered inline markers after individual sentences, plus a source list. The densest citation style in the category and the one that set the pattern.
ChatGPT Search — inline links and a source panel, generated over a mixture of OpenAI's own index and purchased web data. Which sources are even available depends on the account tier.
Microsoft Copilot — inline citations over Bing results. Since April 2025, Copilot Search in Bing shows the generated answer and conventional ranked results on the same page, which Microsoft framed at launch as supporting a healthy web ecosystem. Keeping the links visible next to the answer is the most publisher-friendly design any major product ships.
Google AI Overviews — links to supporting pages within and beside the summary, with the ranked results still below. Since 27 January 2026 the "Show more" control hands the reader into AI Mode's chat interface on mobile, at which point the ranked list is left behind.
Yahoo Scout — cited sources with rich media, generated by Anthropic's Claude over Microsoft Bing's grounding API.
Brave Search's Summarizer — an AI answer above the results with citations back to source pages, launched March 2023 and switchable off in settings.
Kagi — publishes a policy that its AI answers must cite sources, and its Quick Answer is invoked by the reader rather than fired on every query.

The citable universe is fixed by the index

This is the part that almost no coverage states, and it explains more about citation patterns than any property of the model does.

An answer engine can only cite what its grounding index contains. A page that a given crawler never fetched, or that the index dropped, is not a candidate for citation no matter how relevant it is. So the question "why does this engine keep citing the same handful of sites?" is often a question about an index, not about a language model.

Microsoft Copilot has no crawler of its own and grounds on Bing, so its citable set is Bing's index — Microsoft's own crawl, not a rival's, but a single one.
Yahoo Scout, launched 27 January 2026, grounds on Bing's API. Yahoo's own framing was that the underlying index is Bing's while the responses and ranking are Yahoo's. Yahoo has not crawled the web since around 2010, so Scout's citable set is also Bing's.
ChatGPT draws on OpenAI's own index built by OAI-SearchBot, on purchased web data, and on Bing for Enterprise and Edu workspaces. The August 2026 measurement found free and paying users served from materially different backends — which means the same question can produce different citable sources depending on the subscription.
Perplexity cites from its own index, built by PerplexityBot, described in September 2025 as tracking over 200 billion unique URLs.
Google cites from its own Search index, the same one Googlebot builds for the ranked links beneath.

Two practical consequences. Asking Copilot and Yahoo Scout the same question is not a cross-check; it is one index asked twice. And when an engine cannot cite something, adding detail to the question will not help, because the limit is upstream of the conversation.

Where citation goes wrong

Attribution after the fact. If a model writes a paragraph and an attribution step then matches sentences to retrieved documents, the citation can land on a plausible source rather than the actual one. The reader has no way to tell those two cases apart.

Documented failures. In June 2024 Forbes accused Perplexity of plagiarising its reporting, and Wired published parallel findings on scraping and inaccurate summarisation. The News Corp action filed in 2024 on behalf of Dow Jones and the New York Post included allegations that Perplexity produced quotes that did not exist and attributed them to those publications. Whatever the litigation eventually decides, the allegations describe the specific failure mode the format invites: fabricated content wearing a real source's name.

The appearance of verifiability. A cited answer feels checked. Almost nobody opens the links. The measurement showing that merely-retrieved pages are cited around 7% of the time is reassuring about the systems; it says nothing about readers, who treat a marker as a substitute for the source rather than a route to it.

Staleness inside the citation. The cited page can be current while the sentence generated from it is not, or vice versa. A generated answer carries no date of its own, and the index behind it has its own refresh cycle.

Citation is becoming a commercial relationship

The convention started as an attribution courtesy. It is turning into a contract.

OpenAI launched ChatGPT search on 31 October 2024 alongside content licensing agreements naming the Associated Press, Axel Springer, Condé Nast, the Financial Times, Le Monde, News Corp, Reuters and The Atlantic among others — money flowing out to publishers in exchange for licensed content appearing in answers. Yahoo joined Microsoft's Publisher Content Marketplace pilot when Scout launched in January 2026. Perplexity, facing suits from News Corp, Reddit and three Japanese newspaper publishers, runs a revenue-sharing programme with cited publishers as a partial answer to the complaint.

The significance for a reader is not the money but what it implies about the pool. When some sources are in an answer engine's citable set through a negotiated agreement and others are in it through a crawl, the citation list is partly an editorial artefact and partly a commercial one, and nothing in the interface distinguishes them.

Reading a citation properly

None of this makes cited answers useless. It makes them a starting point with a built-in audit trail, which is more than an uncited answer offers. Four habits do most of the work.

Open the link before repeating the claim. The citation exists precisely so the sentence can be checked, and the checking is the reader's job, not the system's.
Check that the cited page says the specific thing. Not the general topic — the specific number, date or quotation.
Notice what is missing. A generated answer does not show what it declined to cite, or that every source was one publisher's, or that the only primary document was never opened.
Prefer the source to the summary when citing onward. A generated paragraph is not reproducible; the page it points at is.

Frequently asked questions

How do AI search engines decide what to cite?

They retrieve a candidate set from a web index, open and read some of it, generate an answer conditioned on the selected passages, and attach citations at the end. Pages actually opened are cited far more often than pages merely retrieved — around 74% against 7% in an August 2026 measurement of ChatGPT. No engine publishes a model of why one eligible source is chosen over another.

Do citations in AI answers mean the answer is accurate?

No. A citation records that a document was retrieved and was available to the model, not that the sentence attached to it reflects that document. Reporting by Forbes and Wired in June 2024, and allegations in the News Corp litigation about quotations that did not exist, describe exactly this failure. The visual grammar borrows from academic citation, where it means something much stronger.

Why do AI answers keep citing the same websites?

Largely because of the index rather than the model. An answer engine can only cite what its grounding index contains, and several products share one index — Copilot and Yahoo Scout both ground on Bing. Content licensing agreements also place some publishers in the citable pool by contract rather than by crawl, and nothing in the interface distinguishes those cases.

Does Perplexity cite differently from ChatGPT?

In style and in substrate. Perplexity uses dense numbered markers after individual sentences and cites from its own index, built by PerplexityBot and described in September 2025 as tracking over 200 billion URLs. ChatGPT uses inline links and cites from a mixture of OpenAI's own index, purchased web data and, for Enterprise and Edu workspaces, Bing.

Can two people get different citations for the same question?

Yes, routinely. Generated answers are non-deterministic, so wording and sources vary between runs. More significantly, independent measurement in August 2026 found free and paying ChatGPT users being served from materially different retrieval backends, which means the pool of sources available to cite is not the same for both.

Do AI search engines pay the sites they cite?

Some do, selectively. OpenAI launched ChatGPT search with licensing agreements naming AP, Axel Springer, Condé Nast, the Financial Times, Le Monde, News Corp, Reuters and The Atlantic among others. Perplexity runs a revenue-sharing programme with cited publishers while defending copyright suits. Yahoo joined Microsoft's Publisher Content Marketplace pilot in January 2026. Most cited sites are paid nothing.

Which AI search products still show ordinary links alongside the answer?

Google AI Overviews sit above the ranked results, which remain on the page. Copilot Search in Bing, launched April 2025, deliberately shows the generated answer and conventional web results together. Brave Search puts its Summarizer above results and allows it to be switched off. Pure answer engines such as Perplexity have no ranked list at all.

How AI search engines cite sources