AI assistants rarely quote your page word for word. ChatGPT, Perplexity, Google’s AI features and Claude pick out short passages from the pages they retrieve, restate them in their own words and attach a citation to the sentence that used them. Several are built to avoid long verbatim copying. So the goal isn’t a paragraph an engine will paste. It’s sentences that each carry one complete, checkable claim, so that when the engine paraphrases them the claim stays accurate and the citation still points to you.

This page covers what the vendors document about how answers use sources, what published studies found about which passages get picked, the writing rules that hold up, and the popular rules that don’t. For the GEO research paper in full, see What is Generative Engine Optimization? For page-level fixes, see the AI Citation Readiness Checklist.

Do AI assistants actually quote verbatim?

Mostly not, and that is by design. Here is what each vendor’s own material says.

Perplexity describes its answers as summaries. Its help center says Perplexity searches the web and “distills this information into a clear, concise summary,” with numbered citations linking to the original sources (How does Perplexity work?).

Claude keeps its own wording separate from your exact text. In Anthropic’s Citations API documentation, the answer reads “the grass is green” while the citation’s cited_text field holds the source sentence, “The grass is green.” Plain-text documents are split into sentences, which set “the minimum granularity of possible citations”: Claude can cite one sentence or several consecutive ones. When Claude searches the web through the API, each citation carries “up to 150 characters of the cited content” (web search tool docs). Anthropic’s published system prompts leave out the search instructions, but a leaked copy of the Claude 4 version, analyzed by Simon Willison in May 2025, told Claude to include at most “ONE very short quote from original sources per response,” under 15 words. The current wording may differ; the intent, paraphrase over reproduction, is plain.

OpenAI says ChatGPT can search the web to answer “with current information and links to relevant sources,” and that responses using search “may include citations” (ChatGPT search help). In its API, a citation is an annotation on the model’s own answer that records “the URL, title and location of the cited source” (web search guide). OpenAI’s Model Spec (October 29, 2026 version) also says the model “must respect creators, their work, and their intellectual property rights.”

Research measures attribution, not copying. The GEO paper (Aggarwal et al., KDD 2024) measured a source’s visibility by the number of words in answer sentences that cite it, not by how much of its text appears word for word. And when Kevin Indig traced ChatGPT citations back to pages (more below), his team matched answers to source sentences with sentence-embedding models, which compare meaning rather than exact wording, according to Search Engine Land’s write-up.

The practical upshot: exact wording survives in short pieces at most, such as a name, a number, a defined term or a phrase. Everything else gets restated. Write so the restatement is still right.

What engines take from a page: passages, not pages

Engines retrieve whole pages, but they hand the model selected pieces of them.

Perplexity says its index splits documents into “self-contained spans, each of which can be individually retrieved and ranked at query time,” and that its pipeline is “designed to surface the most atomic units possible to the model” (Perplexity engineering post, September 25, 2025). A span that only makes sense next to the previous paragraph is a weaker candidate. More in How to Rank in Perplexity AI.

Google ties its AI features to ordinary snippets. To appear as a supporting link in AI Overviews or AI Mode, a page “must be indexed and eligible to be shown in Google Search with a snippet,” and the same nosnippet, data-nosnippet and max-snippet controls limit what Google shows from your pages (AI features and your website). Google also says “there’s no requirement to break your content into tiny pieces,” that its systems can understand several topics on one page “and show the relevant piece to users,” and that “there’s no ideal page length” (Google’s guide to generative AI features). Clear sections help; chopping a page into fragments doesn’t.

An outside look at Gemini’s grounding. Dan Petrovic of the SEO firm DEJAN logged the snippets that Google’s search grounding tool passed to Gemini through its API for 7,060 queries in his clients’ industries, and compared them with 2,275 full pages (DEJAN, December 20, 2025). The selected chunks averaged 15.5 words and were extracted from the page text, not rewritten. Each query got a median of about 1,900 words of grounding in total, and the median source contributed 377 words. Longer pages got only a little more selected, and selection levelled off at about 540 words: pages under 1,000 words averaged 370 grounding words (61% of the page), pages over 3,000 words 544 (13%). The dataset isn’t public, it covers the Gemini API rather than AI Overviews, and other factors weren’t controlled, so read it as a description of one pipeline, not a rule. It fits the picture above: the model sees extracts, then writes its own answer.

Where on the page cited passages come from

The most repeated advice is “put the answer in the first 40, 60 or 80 words.” We found no study that supports a word count. The best public evidence on position comes from Kevin Indig’s analysis of 18,012 ChatGPT citations matched back to specific source sentences, published in his Growth Memo newsletter (February 16, 2026, paywalled) and summarized by Search Engine Land on February 18, 2026:

Where the cited sentence satShare of citations
First 30% of the page44.2%
Middle (30–70%)31.1%
Final 30%24.7%
Within its paragraph: first sentence24.5%
Within its paragraph: middle sentences53%
Within its paragraph: last sentence22.5%

So the top of the page matters, but more than half of citations came from further down, and most came from the middle of paragraphs rather than their opening lines. Indig also reported that cited passages were nearly twice as likely to use definitional phrasing (“X is,” “X refers to”), that 78.4% of question-linked citations came from headings, and that cited text was dense with named entities (20.6% proper nouns, against 5–8% in typical English).

Treat these as correlations from one analyst, on one engine, using semantic matching, with the full method behind a paywall. They support “answer early, and keep every paragraph substantive.” They don’t support any fixed word limit.

What content changes have been tested

Two benchmarks have tested rewriting a source to win more of an AI answer, with different results.

The GEO paper (KDD 2024) rewrote one source per query nine ways and had GPT-3.5-turbo answer from five sources. Adding quotations, statistics or citations to sources did best: the authors report 30–40% gains on their main visibility metric. Keyword stuffing did worse than no change. The method and full results table are in What is Generative Engine Optimization?

C-SEO Bench (NeurIPS Datasets & Benchmarks 2025) retested ten such methods, including the GEO paper’s rewrites and an llms.txt-style summary, across six domains and four models (gpt-4o-mini, claude-3-5-haiku, o3 and o4-mini), and with several competing pages rewritten at once (Puerto et al., arXiv:2506.11097). The authors found most methods “largely ineffective” and often harmful to a page’s position, while “traditional SEO strategies, those aiming to improve the ranking of the source in the LLM context, are significantly more effective.” Gains shrank as more competitors used the same methods.

Read together: specific, sourced writing can help a page that is already among the sources, but surface rewriting is a weak lever, and being retrieved and ranked comes first. That is ordinary search work: crawl access, indexing and pages that deserve to rank.

Citations often go wrong, another reason to write clearly

Engines don’t always attach the right source to the right sentence.

  • Liu, Zhang and Liang (Findings of EMNLP 2023) audited Bing Chat, NeevaAI, Perplexity and YouChat and found that only 51.5% of generated sentences were fully supported by their citations, and only 74.5% of citations supported the sentence they were attached to.
  • The Tow Center for Digital Journalism (Jaźwińska and Chandrasekar, Columbia Journalism Review, March 6, 2025) gave eight AI search tools verbatim excerpts from 200 news articles (1,600 queries in all) and asked them to name the article, publisher and URL. Collectively they answered more than 60% incorrectly, from 37% for Perplexity to 94% for Grok 3. Several cited syndicated copies on other sites instead of the original.

Both studies predate today’s models, and you can’t fix an engine’s attribution. You can make it easier to get right: one claim per sentence, the subject named in that sentence, your brand and the original source named in the text, and one canonical page for each piece of original data.

How to write passages that survive paraphrase

Each rule below follows from the evidence above. None needs special markup.

Put one complete claim in each sentence. Citations attach to sentences (Claude’s API cites sentence by sentence) and Perplexity ranks self-contained spans. Name the subject instead of writing “it” or “this,” and spell out acronyms where they matter. A sentence that depends on the one before it can’t stand alone as a citation.

Keep the number, the source and the date in the same sentence as the claim. Compare “The index is huge” with “Perplexity’s search index tracks over 200 billion unique URLs, according to the company’s September 2025 engineering post.” If an engine restates the second sentence, the figure and its origin travel with it. Name the source in the text itself, not only in the link.

Answer early, then keep every paragraph substantive. Put the direct answer in the first sentence or two of the page and of each section, because position correlates with citation. But most cited sentences in Indig’s data sat mid-paragraph, so filler after the opening is wasted. A quick self-check: read only the first sentence of each section and see whether it answers the heading.

Use headings that say what the section answers. Question-style or descriptive H2s work like the prompt that the next paragraph answers. Google’s guide also notes that people appreciate pages “organized by paragraphs and sections, along with headings that provide a clear structure.”

Add real quotations, statistics and sources. This is the GEO paper’s best-supported result. Quote named people from public sources, link the original study, and give its sample and date. Never invent figures or quotes; readers and other sites will check them, and a wrong number gets repeated under your name.

Prefer plain, definitional statements. “Generative engine optimization is the work of…” restates cleanly. A sentence full of hedges and asides doesn’t. You don’t need to repeat the exact query wording: Google says its AI systems “can understand synonyms and general meanings.”

Keep pages focused, without padding or chopping. Google says there’s no ideal length, and DEJAN’s data suggests length beyond about 2,000 words buys little extra grounding. Cover what the reader needs, then stop.

Publish your original data where you control it. If your numbers live in a gated PDF, engines that can’t read it will cite whoever repeated them. Put key findings on a public, indexable page, link to it from syndicated versions, and keep the facts current. The content refresh playbook covers upkeep.

Rules to drop

These circulate widely, and none rests on published evidence:

  • Fixed word counts (“answers of 40–80 words,” “quotable claims of 15–50 words,” “answer cards of 60–80 words”). No vendor or study sets one. The documented units are sentences, self-contained spans and, in one study, extracts averaging about 15 words.
  • Writing for verbatim reuse. Assistants are built to paraphrase, and the leaked Claude prompt capped quotes at under 15 words. Aim to be credited, not copied.
  • Chunking pages into tiny sections for Google. Google says to ignore it.
  • FAQ schema as a quoting signal. Google stopped showing FAQ rich results on May 7, 2026, and says no special schema is needed for its AI features. A visible Q&A section can still help readers. See the structured data guide.
  • Exact-match keyword phrasing. Google says its AI systems understand synonyms, and the GEO paper found keyword stuffing lowered visibility.

How to check whether your passages are used

Run a fixed set of real buyer questions on each engine you care about every month, more than once each, because answers vary. Log your appearance rate (the share of answers that name you), your citation rate (the share that link to your site) and your share of AI voice (your share of brand mentions across the competitors you track). Then put the answer sentence that cites you next to your own sentence. If the engine restated your claim correctly, the passage works. If it credited your figure to someone else, or garbled it, tighten that sentence. The full method is in How to Measure AI Citations; engine-specific steps are in the guides for ChatGPT, Claude and Google AI Overviews.

FAQ

Do AI assistants quote content word for word?

Rarely, beyond short fragments. Perplexity describes its answers as summaries, Anthropic’s citation format pairs Claude’s own wording with the exact source sentence, and a leaked 2025 Claude system prompt limited quotes to one under 15 words. Expect a paraphrase with a citation, and write so the paraphrase stays accurate.

How long should a quotable passage be?

No vendor or study gives a word count. What is documented is the unit: Claude’s API cites whole sentences, Perplexity ranks self-contained spans, and one study of Gemini’s grounding found extracts averaging about 15 words. Make each sentence complete and specific, and make each section answer its heading.

Does putting the answer in the first paragraph get me cited?

It helps, but it isn’t enough. In Kevin Indig’s analysis of 18,012 ChatGPT citations, 44.2% came from the first 30% of the page, so most came from further down, and over half came from the middle of paragraphs. Answer early, and keep the rest of the page substantive.

Do statistics and quotations really make content more citable?

In the GEO paper’s benchmark, adding them raised a source’s share of the answer by 30–40% on the main metric. A later benchmark, C-SEO Bench, found such rewrites largely ineffective once several competing pages used them, and found ranking mattered more. Use real, sourced facts because they make claims checkable, not as a trick.

Can AI cite content that sits behind a login or in a gated PDF?

Usually not through search, because crawlers can’t get past forms and logins, though some publishers license content to AI companies and agent browsers can act in a user’s signed-in session. If you want a finding cited, publish it on a public, indexable page.


Want an outside look? The free AI Visibility Snapshot tests 10 prompts on ChatGPT, Google AI Overviews and Perplexity, with a short note on where you’re cited and where you’re not. The Citation Audit tests 30–50 locked prompts across six surfaces (ChatGPT, Google AI Overviews, Perplexity, Gemini, Claude and the OpenAI Responses API) and delivers competitor gaps, schema and content fixes and a prioritized 60–90 day roadmap in 5 business days. It’s $1,500, fully credited toward an Implementation Sprint or retainer signed within 30 days. We don’t guarantee citations or rankings. See pricing.

Request a free Snapshot or a Citation Audit →

Sources (checked November 11, 2026):

Related reading: What is Generative Engine Optimization?, How to Rank in Perplexity AI, How to Measure AI Citations, The AI Citation Readiness Checklist, The AEO Content Refresh Playbook.