How to write content that Claude will quote

TL;DR: Claude quotes content that answers the question in the first sentence of each section, uses concrete numbers with named sources, structures data in tables and lists, and earns links from authoritative pages. Write for extraction, not engagement. These patterns hold across Claude, ChatGPT, and Perplexity because all three retrieve on similar heuristics.

What does Claude actually look for when it decides to quote a source?

Claude quotes passages that answer the user's question at the sentence level. That is the whole game. Claude is a large language model from Anthropic, and while it does not browse the open web by default, Claude.ai's web search mode and Anthropic's API integrations retrieve live pages and then pick which chunks to surface. The chunk it picks is almost always the one that opens with a complete answer.

The most useful public research here is a 2024 study from Seer Interactive that looked at over 5,000 AI citations across ChatGPT, Perplexity, and Bing Copilot. Cited pages ran longer (around 1,400 words versus about 900 for non-cited), had more referring domains, and answered questions in the first paragraph of each section instead of building toward the answer [1]. Claude is not in that dataset. But Anthropic has said Claude's training rewards "helpful, harmless, and honest" responses, which pushes the model toward passages that are direct and factually grounded [2].

Here is the practical version. Claude wants extractable chunks. A chunk is roughly a paragraph or short section that stands on its own as a full answer. If your paragraph only makes sense after reading the three before it, it will not get quoted. If it leads with the answer and follows with support, it will.

A 2023 paper on retrieval-augmented generation from researchers at Princeton and MIT found that the top retrieved chunks almost always put the answer in the first sentence, used concrete entities (numbers, proper nouns, named sources), and ran between 100 and 250 words [3]. Those findings map cleanly onto how Claude handles retrieved content in web search mode.

Does Claude quote differently than ChatGPT or Perplexity?

Yes, and the differences matter. Perplexity is built on retrieval-augmented generation and shows citations up front, so its signals look most like traditional search: domain authority, fresh content, direct keyword match [9]. ChatGPT with web browsing behaves close to Perplexity in its citation patterns.

Claude is pickier. Anthropic trained it to steer away from unreliable sources, so Claude leans toward pages that themselves cite primary sources, government data, or peer-reviewed research. A page that says "studies show" without naming the study gets quoted less than one that says "a 2023 meta-analysis in JAMA found X." Claude seems to pass credibility down the chain: cite credible sources, and Claude reads your page as more credible too.

The table below sorts the main behavioral differences, drawn from the available research and Anthropic's published guidance.

| Model | Citation display | Favors | Weakness | |---|---|---|---| | Claude (web search) | Inline links, selective | Pages citing primary sources, structured answers | Less transparent about what it retrieved | | ChatGPT (web) | Inline numbered refs | High-authority domains, recent content | Sometimes cites the same domain repeatedly | | Perplexity | Numbered sidebar sources | Direct keyword match, fresh pages | Can over-cite thin content on exact matches | | Gemini (web) | Inline links | Google-indexed authority signals | Strong bias toward well-known brands |

The generative engine optimization guide breaks down these retrieval architectures in more detail.

One difference is worth memorizing. Claude has a stronger preference than GPT-4 for content that hedges honestly. A line like "nobody has good data on this; the closest estimate is X from source Y" gets quoted by Claude more readily than a confident claim with no source. That tracks Anthropic's stated preference for calibrated uncertainty in Claude's outputs [2].

What content structure makes Claude most likely to cite your page?

Structure decides whether a retrieval system can lift an answer-sized chunk off your page. It is not cosmetic. The pattern that works best is the inverted pyramid applied at the section level, more than the article level. Every H2 should open with a sentence that fully answers the question the heading asks, then follow with support.

Claude's retrieval reads the top of each section heavily. Bury the answer in paragraph four and Claude may retrieve the chunk without the answer in it. That chunk is dead weight.

A few structural moves that raise citation odds:

Question-format headings. H2s phrased as questions get retrieved more often because the heading text matches a user's query. A 2024 Brightedge analysis found 58% of Google AI Overviews pulled from pages where the H2 matched query intent almost verbatim [5]. The same logic drives Claude's retrieval.

Tables and structured lists. Tabular data is easy to extract. Claude can quote a single cell. The same numbers written out in prose are harder to chunk cleanly. If you have comparison data, put it in a table.

Fact density. Aim for a concrete number, named entity, or cited source roughly every 150 words. Thin paragraphs full of hedged generalities do not get quoted, not because Claude punishes vagueness, but because there is nothing there to extract.

Short definitions up top. If a section introduces a term, define it in one sentence before anything else. Claude quotes clean definitions verbatim all the time.

Keep paragraphs between 80 and 200 words. Under 80, the chunk may lack context. Over 200, Claude may pull only part of it, and the part it pulls might miss your point.

Average word count: cited vs non-cited pages in AI responses

| | | |---|---| | Cited pages (avg words) | 1,400 | | Non-cited pages (avg words) | 900 | | Google AI Overview cited pages (Authoritas) | 1,447 | | AI Overview threshold (under-performing) | 700 |

Source: Seer Interactive, AI Citation Study 2024

How long should your content be for Claude to trust it?

Length is a proxy for depth, not a target on its own. The data is fairly consistent though: cited pages run longer than non-cited pages. Seer Interactive's 2024 analysis put cited pages at around 1,400 words against about 900 for non-cited pages in the same topics [1]. A separate 2024 Authoritas analysis of Google AI Overviews found cited pages averaged 1,447 words, and content under 700 words got cited at less than half the rate of content over 1,200 words [4].

Nobody has clean numbers on Claude's length preferences specifically, because Anthropic has not published them. The reason length correlates with citation is coverage. Longer pages tend to answer more sub-questions, so they match more retrieval queries. Coverage signal, not a word count rule.

My practical target for a page you want Claude to quote: 1,200 to 2,500 words. Under 1,000 and you are probably missing angles. Over 3,000 and you dilute answer density with padding, which hurts extraction.

If a page is already doing well in ai search, bolt on a dedicated FAQ section (10 to 14 questions, 50 to 80 words each, self-contained answers). Each FAQ is its own extractable chunk. A page with 12 solid FAQs has 12 more shots at matching a Claude query.

What role do external links and domain authority play in Claude citations?

This is the part most content teams underrate. Claude does more than judge your page alone. It appears to weigh whether credible sources cite your domain in the first place.

Anthropic has not published a domain authority score the way Google talks about PageRank, but the indirect evidence is strong. The Seer Interactive study found pages with more than 50 referring domains were cited at roughly 3.4 times the rate of pages with fewer than 10, and that relationship held even after controlling for content quality scores [1].

So here is the hard truth. If you are a new domain or a brand site with few inbound links, perfectly structured content may not be enough on its own. You need the authority baseline first. Building it for AI citation works the same way it does for traditional SEO: earn links from relevant, high-authority pages by publishing original data, real research, or free tools people actually reference.

One move that pays off with Claude specifically: get cited by Wikipedia and the sources Wikipedia cites. Claude's training data leans heavily on Wikipedia and its references. A link from an academic site, a government page, or a major news outlet gives your page a real authority lift for citation purposes. There is no shortcut. It takes months, not days.

To check whether your authority gains are turning into actual citations, tools in the ai visibility tool and ai seo tools categories can read citation frequency across models.

Which types of claims does Claude prefer to quote verbatim?

Claude quotes some claim types far more than others. Based on patterns across documented AI citation behavior and Anthropic's stated priorities around accuracy, these are the ones that show up most:

Numerical findings with a named source. "A 2024 study by [organization] found that X% of [population] did Y" is the single most quotable sentence structure there is. Specific, falsifiable, attributed. Claude can repeat it without distortion.

Definitions, especially of technical or contested terms. Define "generative engine optimization" clearly, with a date or originating source, and Claude will quote that line again and again.

Statutory or regulatory thresholds. Name a specific rule, law, number, or deadline from an official source and that passage turns highly extractable. This matters most for legal, financial, and health content, but the pattern travels.

Hedged summaries of uncertain evidence. Counterintuitive, but true. Claude quotes "the evidence here is mixed; the best available study found X but was limited by Y" more readily than a confident overclaim. Anthropic trains Claude to be calibrated, so it favors sources that model that calibration [2].

What Claude almost never quotes: marketing language, abstract benefit statements, unsourced comparisons, and anything that reads like a product page. "Our solution delivers best-in-class results" will never land in a Claude response. "Independent testing by [lab] found the product reduced X by 34% under Y conditions" might.

For how these patterns play out across platforms, ai-powered search features covers the retrieval mechanics.

Does publishing original research or data make Claude more likely to cite you?

Yes. This is probably the highest-leverage thing you can do.

Original data means Claude has no other source to quote for that finding. Survey 500 customers and publish the results, and that data point lives only on your page. Every time a user asks Claude a question your data answers, your page is the only source in the room. That is a monopoly on a citation.

The bar for "original research" is lower than it sounds. You do not need a peer-reviewed methodology. Useful original data includes a survey of your own customer base with a stated sample size and method, an analysis of public data you ran yourself and have not seen elsewhere, a benchmark test comparing tools with documented steps, or an annual trends report built on data you collected.

The requirements are simple. State the methodology, the sample size, and the date, and make the findings scannable in a table or numbered list. Findings buried in prose get quoted less than the same findings in a labeled table.

Spawned's own review of pages that show up often in AI citations finds original data pages earn roughly 3 to 5 times as many AI citations per month as opinion or guide content of the same length. That is internal observation, not a published study, so treat it as directional.

To measure whether this is working, ai search visibility metrics kpis walks through the metrics worth tracking.

How does freshness affect Claude citation likelihood?

Freshness matters less for Claude than for Perplexity, and less than most people assume. Claude's base model has a training cutoff (Anthropic updates it periodically; the cutoff for Claude 3.5 Sonnet is early 2024 [2]). For queries where Claude is not using web search, freshness is irrelevant. What gets into training is chosen by quality and authority, not publish date.

With web search on, freshness matters for time-sensitive queries (current events, recent product releases, live pricing) and matters much less for evergreen ones. A well-structured 2022 article explaining a concept will often beat a thin 2025 article on the same topic, because the older page has more inbound links and has been indexed longer.

The one freshness signal worth caring about is the last-updated date. Pages that show a clear "last updated" date and have actually refreshed their content (new data, updated statistics) do better than pages that are visibly stale. Updating your statistics once a year and marking the date is worth the hour.

For topics touching recent events or recent AI behavior, Perplexity and Bing Copilot are more citation-competitive than Claude because they index harder and faster. If your topic is time-sensitive, optimize for those first and treat Claude as the secondary target.

What technical and on-page factors affect Claude citation rates?

Most technical factors that matter for Claude are the same ones that matter for traditional SEO, because the pages Claude retrieves in web search are the same pages Google indexes. A few specifics:

Page speed. A 2024 Brightedge analysis found pages cited in AI Overviews loaded in under 2.5 seconds on mobile at the 75th percentile [5]. Slow pages get de-prioritized in Google's index, which cuts their odds of showing up in retrieval at all.

Schema markup. FAQ schema and HowTo schema make content easier for retrieval systems to parse. Pages with FAQ schema get their Q&A pairs surfaced more directly in AI responses. Adding FAQ schema to question-structured pages is low effort and worth doing.

Mobile friendliness. Googlebot crawls mobile-first. Pages that render badly on mobile get indexed worse, which hurts retrieval. Table stakes, not a differentiator, but ignoring it is a real problem.

Canonicalization and crawlability. If your best content sits behind a login, a paywall, or a robots.txt block, Claude cannot retrieve it through web search. Obvious, and still constantly overlooked by B2B teams who gate too much.

Image alt text. Claude does not read images in retrieved web pages the way it reads text. Data that lives only in an infographic is invisible to retrieval. Always put the key numbers from an infographic into text on the same page.

For a full technical audit, the walkthrough at ai seo covers the checklist.

How should you write FAQs to maximize Claude citations?

FAQs are the highest-density citation surface on a page. Each one is an independent retrievable chunk, and Claude pulls from FAQ sections often because the format mirrors how people phrase questions to AI assistants in the first place.

A strong FAQ for Claude has four properties. First, the question reads exactly as a user would ask it, including natural phrasing and common variants. "What is the difference between GEO and SEO?" beats "FAQ: GEO vs SEO."

Second, the answer opens with the answer. No preamble, no "great question." First sentence answers it. Everything after is support.

Third, each answer stands alone. Claude may extract just the FAQ with none of the surrounding page. An answer that says "as mentioned above" breaks the moment it is lifted out.

Fourth, put at least one concrete fact in each answer: a number, a date, a named source. Purely abstract answers do not get quoted. Answers with a specific, attributable data point do.

The sweet spot for Claude extraction is 50 to 90 words per answer. Under 50 and there often is not enough for a useful response. Over 100 and the answer loses focus, which drags down extraction quality.

Target 10 to 14 FAQs per major page. Cover the core question, the common follow-ups, the "how much / how long" questions, and at least one "what should I do" question. Answer the follow-on queries a user would fire off after their first one.

Are there content formats Claude will never quote, no matter how well written?

Yes. Some formats are structurally incompatible with how Claude retrieval works.

Pure opinion with no supporting evidence. Claude can engage with an opinion, but it does not quote unsupported assertions. If every claim in your article is your personal take with no data or source, none of it gets extracted.

Content that needs prior context to make sense. If a section's key claim only lands after the reader has read three earlier sections, Claude cannot quote it cleanly. It needs self-contained passages.

Content behind authentication or paywalls. If web search cannot reach the page, Claude cannot quote it. Academic content locked behind JSTOR, B2B content behind lead forms, and premium newsletter archives are all invisible to AI retrieval.

Content in non-text formats. Data in PDFs without proper text extraction, information trapped in images, video transcripts not published as text, and podcasts with no transcript all get missed. Publish text versions of every media asset.

Content on very low-authority domains with few inbound links. Not a hard never, but for practical purposes a brand-new domain with no links will not get quoted by Claude even with perfect structure. Authority is a gate you clear first.

Marketing-only pages. A page that exists purely to sell, with no informational substance, does not get cited. Claude has learned to skip marketing copy because it rarely helps a user.

How do you measure whether your content is actually being cited by Claude?

This is harder than tracking traditional search rank, and the tools are young. A few methods hold up:

Direct query testing. List the 20 to 30 questions your target pages are built to answer. Ask Claude each one, web search on, and note whether your page gets cited. Do it monthly and track the drift. Manual, but it is ground truth.

Brand mention tracking in AI responses. Several tools now track how often your brand or URL shows up in AI model responses at scale. The brandrank.ai visibility insights analysis covers one approach, and this category is growing fast.

Referral traffic from AI platforms. When Claude.ai cites pages in web search mode, it can send referral traffic. Check analytics for referrers from claude.ai and other AI assistant domains. That traffic is small for most sites today but rising. The Seer Interactive 2024 study reported AI referral traffic grew 48% quarter-over-quarter in its tracked dataset [1].

Share of voice in AI responses. In a competitive category, you want to know how often you get cited relative to competitors, more than whether you get cited. Manual testing gives you a directional read. Tools in the ai-mode-seo-tool category do it at scale.

One honest caveat. Nobody has strong longitudinal data on Claude citation rates specifically, as opposed to AI search citations in general. The closest research pools all models together. Treat model-specific attribution as approximate until measurement catches up.

Sources

1. Seer Interactive, AI Citation Study 2024
2. Anthropic, Claude Model Documentation and Usage Policies
3. Princeton and MIT, Retrieval-Augmented Generation Study 2023
4. Authoritas, AI Overviews Citation Analysis 2024
5. Brightedge, AI Search Content Analysis 2024
6. Anthropic, Responsible Scaling Policy
7. Perplexity AI, About and Documentation
8. Wikimedia Foundation, Content and Conflict-of-Interest Policies