Do AI engines cite whole pages or specific sentences?

AI engines don't share a citation format. Each one decides what a citation points at, and the answer ranges from a short quote to a whole page.

So Claude can tell you which sentence it leaned on. Perplexity tells you which page. OpenAI tells you which URL, and separately which URLs it read but didn't cite. If you want the platform detail, see how Claude picks sources and cites them and how Perplexity's citation IDs work.

What this means for you: the question "did the AI cite us?" has a different shape on each platform. On Claude it can be answered at the sentence level. On Perplexity and ChatGPT it's answered at the page or URL level.

A citation from Claude says "this 150-character span supported this claim." A citation from Perplexity says "source number 3 supported this sentence." Those are different measurements. One counts quotes. The other counts documents.

Adding them together is like adding page views to paragraphs read. You get a number, but it doesn't describe anything real. Three problems stack up:

• Different units. One page can produce several Claude span citations in a single answer, where Perplexity would credit that page once.
• Different surfaces. Google says AI Overviews and AI Mode "may use different models and techniques" and can return different links, even inside one company.
• Different retrieval states. Retrieval is discretionary per query on every platform. An answer given without live search has no citations at all, so any blended score averages answers with and without retrieval.

The practical move is to keep each engine's numbers separate and label the unit. A rise in Claude span citations and a rise in Perplexity source citations are two findings, not one. For a closer look at the formats side by side, read why Claude and Perplexity cite sources differently.

Citation counts are a noisy proxy for influence. The evidence points to two reasons.

First, engines read more than they cite. OpenAI's web search returns the full list of sources the model consulted separately from its citations, and the sources list is usually longer. A page can shape an answer without appearing as an inline citation.

Second, the act of citing is partly separate from the act of using a source. A 2026 mechanistic study of attribution in retrieval-augmented generation (arXiv:2606.28358) repaired more than 90% of missed citations by amplifying specific attention components, without changing the sources at all. The model already had the information. Whether it emitted a citation was a separate step.

This matters when you read any citation report. A missing citation does not prove a page went unread. A present citation does not prove the page drove the answer. Treat citations as evidence of presence, not as a measure of weight.

Sentence-level citation makes one thing concrete: the engine needs a sentence it can lift. If your price, your founding year, or your one-line product definition only exists as an implication spread across three paragraphs, there is no clean span to quote.

The evidence here is narrow but positive. In a 252,000-trial experiment on competitive citation (arXiv:2605.25517), explicit price information and a recent timestamp helped consistently, while formatting-only edits had little impact. Topical relevance and list position were the biggest drivers. That study ran on documents already placed in the model's context, so it says nothing about whether a page gets retrieved in the first place.

Two boundaries keep this honest:

• This is about facts, not style. Google states you don't need to write in a specific way for generative AI search. Rewriting a page to sound "AI-friendly" has no evidence behind it, and citation-oriented rewrites have been found to impair retrieval.
• Access comes first. A quotable sentence that only appears after JavaScript runs may be invisible to most AI crawlers, which fetch raw HTML. Put load-bearing facts in the server-rendered HTML.

Say you run a hypothetical accounting firm. "Monthly bookkeeping starts at $300" is a sentence an engine can quote. "Flexible pricing for every business" is not. For more on this, see what makes a sentence easy for AI to quote.

You control access and the clarity of your facts. You don't control which spans or sources an engine chooses. Work in that order.

1. Confirm crawlers can reach the page. Check that robots.txt allows the search bots (OAI-SearchBot, PerplexityBot, Claude-SearchBot) and that your firewall doesn't return 403 to them.
2. Check that key facts are in the raw HTML. View the page source with JavaScript off. If the price or definition is missing, it's missing for most AI crawlers too.
3. State each important fact once, plainly, and consistently. Contradictions between your pricing page and your about page give engines two answers to choose from.
4. Find who else gets cited. When engines cite third-party pages instead of yours, being named on those pages is the association the evidence supports. The playbook on finding the pages AI engines cite walks through it.
5. Measure per engine, with the unit labelled. The free AI citation checker shows which pages get cited, so you can compare like with like rather than trusting one blended number.

What is cited_text in Claude's web search? cited_text is the field Anthropic's web search tool returns with each citation. It holds the exact span of source text, up to 150 characters, that supports a claim in Claude's answer, alongside the URL and page title. As of October 2026, Anthropic doesn't bill these citation fields as tokens. It lets you see which sentence on a page was used, not just which page.

How does Perplexity's citation index work? Perplexity's Agent API returns a list of search results, each with an id, URL, title, snippet and dates. The answer text then references sources by integer index, such as [1] or [3]. The citation points at the whole source, not a specific sentence. Note that web search is opt-in on the Agent API, so an integration without it returns answers with no citations.

Can I combine citation counts from ChatGPT, Claude and Perplexity? You can add the numbers, but the total won't mean anything. Claude counts text spans, Perplexity counts whole sources, and OpenAI counts URLs. One page can earn several Claude span citations where Perplexity credits it once. Report each engine separately, name the unit, and say whether live search was on for the answers you measured.

Does adding schema or FAQ markup make sentences easier to cite? The evidence doesn't support that. Google says structured data isn't required for generative AI search, and an Ahrefs matched-control study found AI Overview citations fell 4.6% after pages added schema. Existing markup does no harm, but it isn't a lever for AI answers. A plainly stated fact in server-rendered HTML is the more defensible move.

Why does an AI answer use my content without citing me? Engines retrieve more pages than they cite, and OpenAI exposes the two lists separately. Research also shows the citation step is partly separate from using a source. The answer may also come from the model's training data with no live search, in which case nothing is cited. A missing citation is not proof your page went unread.