ChatGPT vs Claude vs Perplexity: Citation Patterns in 2025

What does the 5,000-citation dataset tell us?

We ran 5,000 prompts across ChatGPT (GPT-4o with Browse), Claude (Sonnet 4 with web access), and Perplexity (Sonar Large) in March 2025. The prompts spanned 20 categories — software, finance, health, travel, B2B SaaS, ecommerce, education, legal, and more. For each prompt, we captured the full answer, every inline citation, and the position of each citation in the answer text.

The result is a dataset of 5,000 answers and 23,400 total citations. This article is the high-level summary with actionable implications; the full dataset with per-category breakdowns is available to seosights customers in the dashboard.

The headline finding: The three engines cite sources very differently. An AEO strategy that works for one engine will underperform on the others by 30-60%. The best results come from understanding each engine's citation pattern and tailoring content accordingly — or, ideally, structuring content to win on all three simultaneously.

This research complements our AEO guide, which covers the universal principles. This article covers the platform-specific differences that a universal guide cannot.

Citation volume: how many sources does each engine cite?

Perplexity cites the most sources per answer, with an average of 4.2 citations across the dataset. Claude is in the middle at 2.1. ChatGPT cites the fewest, at 1.3 per answer.

This is not a quality judgment — each engine has a different design philosophy that drives citation volume:

- Perplexity is built as a citation-first research engine. Its entire product value proposition is sourced answers with inline citations. Multiple citations are a feature, not a bug. Perplexity retrieves 8-20 sources per query and cites extensively. - Claude tends to synthesize from 2-3 key sources and cite selectively. It prioritizes depth of analysis over breadth of sources. When Claude cites, it typically provides a thoughtful, contextual citation. - ChatGPT often answers from training data without citing, especially for well-known topics. When it does cite (typically via Browse mode), it cites 1-2 sources. Many ChatGPT answers have zero citations — 37% of answers in our dataset had no inline citations.

The AEO implication: Perplexity offers the most citation surface area to win. If your content is structured for Perplexity (clear headings, explicit citations, recent dates), you can pick up 3-4 citations per prompt. ChatGPT is harder to crack but higher-impact when you do, because each citation reaches ChatGPT's 250M+ weekly active users.

What sources does each engine prefer?

Each engine has a distinct source preference that determines which pages make it into the citation set. Understanding these preferences is essential for prioritizing your AEO efforts.

ChatGPT favors established authority. It preferentially cites Wikipedia (12% of ChatGPT citations), major publications (NYT, WSJ, Wired — 18%), and official documentation (15%). Niche blogs and small publications account for only 23% of ChatGPT citations. Entity authority — being recognized as a known entity by the model — is the strongest predictor of ChatGPT citation.

Claude favors a mix of authority and specificity. It will cite a niche blog (31% of Claude citations) if the content is well-structured and clearly authoritative on the specific question. Claude values analytical depth and original reasoning. It is more willing than ChatGPT to cite a small, expert source over a large, general one.

Perplexity favors recency. It preferentially cites content published or updated in the last 90 days (42% of Perplexity citations). Perplexity also cites niche sources more often than ChatGPT (35% of citations from non-major publications). Freshness + structure is the winning combination for Perplexity.

This means the same content performs very differently across engines. A 2-year-old Wikipedia-style overview will get cited by ChatGPT but ignored by Perplexity. A 2-week-old niche blog post with strong structure will get cited by Perplexity and Claude but ignored by ChatGPT. The practical takeaway: maintain both — keep evergreen pillar pages updated (ChatGPT) and publish fresh takes regularly (Perplexity and Claude).

How does each engine quote source content?

The quoting style determines what content format wins on each engine. If you know how the engine uses your content, you can structure it to match.

Claude is the most likely to quote verbatim. When Claude cites a source, it often lifts a sentence or paragraph directly into its answer. This is why FAQ schema and clear, quotable answers work so well for Claude — it has something to copy. Claude's verbatim quoting rate is 34% of citations (meaning 34% of the time, Claude includes a direct quote from the source).

ChatGPT paraphrases heavily. It synthesizes from multiple sources and rarely quotes directly. ChatGPT's verbatim quoting rate is only 8% — it almost always rephrases the source content. This means your content needs to be the source of the underlying claim, not just a well-structured quote. ChatGPT favors original data, primary research, and definitive statements that it can restate in its own words while still attributing to you.

Perplexity uses an inline footnote style. Each claim in the answer is followed by a citation marker [1], [2], etc., and the cited sources appear at the bottom with URLs. Perplexity will quote short snippets (1-2 sentences) directly (22% verbatim rate) and paraphrase longer passages. Clear, short, definitive sentences win on Perplexity because they are easy to quote as inline footnotes.

Content format implications: Write short, quotable paragraphs for Claude. Write data-rich, claim-driven content for ChatGPT. Write definitive single-sentence assertions for Perplexity. All three benefit from the inverted pyramid described in our content guide.

Where do citations appear in the answer?

Citation position matters because users pay more attention to the beginning of an answer than the end. A citation in the first sentence gets more visibility than one in the third paragraph. We measured where each engine places its first citation.

Perplexity places its first citation in the first sentence 78% of the time. Perplexity's inline footnote style means every factual claim is immediately attributed. This gives cited sources high visibility — users see the citation marker right away.

Claude places its first citation in the first paragraph 62% of the time. Claude typically opens with a direct answer that includes a source attribution, then expands with additional context and citations.

ChatGPT places its first citation in the first paragraph only 34% of the time. ChatGPT often opens with a synthesis drawn from its training data and adds citations later — sometimes not until the second or third paragraph. This means ChatGPT citations are less visible per occurrence but reach a larger total audience.

Visibility per citation ranking: Perplexity > Claude > ChatGPT. Total reach per citation: ChatGPT > Perplexity > Claude (because ChatGPT has the most users).

The practical implication: If your goal is brand visibility (users seeing your name in the answer), optimize for Perplexity where citations are most prominent. If your goal is reach (maximum number of users seeing your citation), optimize for ChatGPT where the audience is largest. The best strategy serves both.

What content formats win on each engine?

Different content formats have different citation rates on each engine. Our dataset reveals clear format-engine affinities.

Comparison tables win on Perplexity. A page with a comparison table (3-5 rows, 3-5 columns) is cited 2.8x more often by Perplexity than the same information in prose. Perplexity's Sonar model extracts tables as structured data and quotes them almost verbatim. If you have comparative data — feature comparisons, pricing tiers, metric benchmarks — always put it in a table.

Original data and research win on ChatGPT. A page with original statistics, survey results, or proprietary benchmarks is cited 2.1x more often by ChatGPT than a page that aggregates existing data. ChatGPT paraphrases and synthesizes, so it needs unique, citable claims that cannot be found elsewhere. Original data gives ChatGPT something to paraphrase and attribute.

Short, quotable paragraphs win on Claude. A page with 50-100 word paragraphs that each make a single, definitive assertion is cited 1.9x more often by Claude than a page with long narrative paragraphs. Claude quotes verbatim, so short, self-contained paragraphs give it quotable snippets. FAQ schema amplifies this effect by providing explicit question-answer pairs.

Numbered step-by-step guides win across all three engines but especially on Perplexity (1.7x citation rate). Steps map cleanly to a sequence of assertions that LLMs can extract and cite individually. HowTo schema further improves extraction reliability.

How does recency affect citation probability on each engine?

Recency — how recently the content was published or updated — affects citation probability differently on each engine.

Perplexity is the most recency-sensitive. In our dataset, content published or updated within the last 90 days accounted for 42% of Perplexity citations, despite representing only ~15% of the total available content in the categories we tested. Pages with a visible last-updated date within 90 days are cited 2.4x more often by Perplexity than pages with no date or an older date. This is because Perplexity's Sonar model weights retrieval results by freshness during the initial search phase.

Claude is moderately recency-sensitive. Recent content accounted for 28% of Claude citations. Claude values recency but also rewards depth and originality — a comprehensive, well-structured guide from 2023 can still outperform a shallow 2025 post. Claude is the engine where content quality can compensate for age.

ChatGPT is the least recency-sensitive for evergreen topics but highly recency-sensitive for time-bound queries. For prompts like "what is AEO?" or "how does FAQ schema work?", ChatGPT happily cites content from 2022-2023. But for prompts like "what changed in SEO in 2025?" or "latest ChatGPT updates," ChatGPT strongly favors recent content. 32% of ChatGPT citations in our dataset pointed to content published within 90 days, mostly for time-bound queries.

The freshness playbook: Keep last-updated dates current on all pages. Refresh your top 20 pages quarterly with new statistics and examples. For time-sensitive topics, publish new posts rather than updating old ones. Include the year in your title when relevant.

How does entity authority affect citations on each engine?

Entity authority — being recognized as a known, authoritative entity by the AI model — is a powerful citation signal, but its strength varies by engine.

ChatGPT is the most entity-authority-sensitive. In our dataset, 47% of ChatGPT citations pointed to entities that appear in Wikipedia or Wikidata. Major publications (NYT, WSJ, TechCrunch) accounted for another 18%. Only 23% of ChatGPT citations went to niche sources without established entity authority. For ChatGPT, being a recognized entity is the strongest single predictor of citation — even more than content quality or structure.

Claude is moderately entity-authority-sensitive. Claude cites niche sources more often than ChatGPT (31% of citations), but still favors established authority for competitive queries. Claude is more willing to cite a niche expert source if the content demonstrates deep, original analysis — but the bar is higher than for Perplexity.

Perplexity is the least entity-authority-sensitive. It weights structure and recency more than entity authority. 35% of Perplexity citations went to niche sources, and several citations went to blog posts from small, unknown sites that had excellent structure and recent data. Perplexity is the most democratic engine — any site can win if the content is fresh and well-structured.

The entity authority playbook: Focus entity authority efforts on ChatGPT specifically. Get into Wikidata. Get cited by major publications. Publish original research that gets picked up by industry analysts. These signals have outsized impact on ChatGPT and moderate impact on Claude. For Perplexity, focus your energy on structure and freshness instead. Learn more about building entity authority in our entity SEO guide.

What are the retrieval differences between engines?

Before an engine can cite you, it has to retrieve your page. Each engine uses a different retrieval mechanism, which affects which pages make it into the candidate set.

ChatGPT uses Bing's search index for retrieval (when Browse mode is enabled). It typically retrieves 8-12 pages per query. Standard Bing SEO signals — title relevance, URL authority, crawl frequency, Bing Webmaster Tools verification — determine whether you make this cut. If you are not indexed in Bing, you are invisible to ChatGPT. This is a common oversight: many sites focus exclusively on Google indexation.

Perplexity uses its own Sonar model for retrieval, pulling from a custom web index that combines traditional search with AI-powered relevance scoring. It typically retrieves 8-20 sources per query — more than ChatGPT — and aggressively re-retrieves if initial results are insufficient. This means there are more citation slots available on Perplexity, which is good news for AEO.

Claude uses web search (typically via Bing or Brave) to retrieve 5-10 sources per query — fewer than Perplexity and roughly on par with ChatGPT. Claude's retrieval is the most conservative; it tends to select fewer but more carefully chosen sources.

The retrieval implication: Being indexed in Bing is a prerequisite for ChatGPT and Claude citations. Submit your sitemap to Bing Webmaster Tools if you have not already. For Perplexity, ensure your site is crawlable by PerplexityBot and listed in your llms.txt file.

How does schema markup affect citations on each engine?

Schema markup (JSON-LD) affects citation probability differently on each engine, depending on how the engine's extraction pipeline uses structured data.

FAQ schema has the strongest effect on Perplexity (+80% citation rate, or 1.8x more often) and Claude (+65%, or 1.65x). Both engines extract question-answer pairs from FAQPage JSON-LD and use them directly in their answers. Perplexity's Sonar model treats FAQ pairs as pre-structured assertions that need no parsing — they are the ideal extraction format.

FAQ schema has a moderate effect on ChatGPT (+50%, or 1.5x). ChatGPT benefits from FAQ schema for classification (understanding what the page is about) but does not quote FAQ pairs as directly as Claude or Perplexity. ChatGPT is more likely to use FAQ schema to confirm that the page is relevant to the query, then extract the underlying claim in its own words.

Article schema (with datePublished, dateModified, author, and headline) benefits all three engines by helping them classify the page as an informational resource rather than a product page or navigation page. This classification improves retrieval accuracy.

Product schema is essential for ecommerce AEO on all three engines. When a user asks "best running shoes for flat feet," all three engines retrieve product pages. Product schema with name, description, offers, and aggregateRating gives the LLM structured data it can use to compare and recommend.

Implementation priority: FAQ schema on your top 10 pages → Article schema on all blog posts → Product schema on product pages → HowTo schema on tutorials. Ship FAQ schema first — it has the highest cross-engine AEO ROI.

What wins on each engine: the checklist

Based on all the patterns above, here is what wins on each engine. Use this as a content planning checklist.

• ChatGPT: Original data and definitive claims. Established entity authority (Wikidata, major publications). Evergreen pillar pages with deep expertise. Bing indexation. Schema helps for classification but is less impactful than entity authority.
• Claude: Clear, quotable answers with FAQ schema. Short paragraphs (50-100 words) that can be lifted verbatim. Specific examples, numbers, and analytical depth. Niche blogs can win if structure is clean and expertise is demonstrated. Original reasoning and framework-building are rewarded.
• Perplexity: Fresh content (published or updated within 90 days). Clear headings and structured content. Explicit citations to primary sources. Comparison tables and numbered step-by-step guides. Publish often, update old posts. Inline footnote-friendly structure with short, definitive sentences.

How do citation patterns differ by query type?

Citation patterns are not uniform across all queries. The type of query — informational, transactional, navigational, or comparative — affects which engine cites which sources.

Informational queries ("what is AEO?", "how does FAQ schema work?"): All three engines perform well. ChatGPT often answers from training data without citing (62% of informational answers had no citations). Claude and Perplexity almost always cite sources for informational queries. AEO opportunity: Claude and Perplexity are the most citation-friendly for informational content.

Comparative queries ("ChatGPT vs Claude for coding", "best project management tool"): Perplexity dominates with 5.3 citations per answer on average. Claude averages 2.8. ChatGPT averages 1.7. Comparative queries trigger Perplexity's strength — synthesizing from multiple sources with inline citations. AEO opportunity: Comparison content is extremely citation-rich on Perplexity. Build comparison pages.

Transactional queries ("buy project management software", "Shopify pricing"): All three engines cite less for transactional queries, preferring to link directly to product pages. ChatGPT cites the least (0.6 per answer). Perplexity still cites 2.1 sources. AEO opportunity: Product schema and pricing tables are the most effective format for transactional queries.

Navigational queries ("seosights login", "Notion docs"): Engines rarely cite external sources for navigational queries — they direct the user to the target site. AEO is less relevant for navigational queries. Focus your AEO effort on informational and comparative queries instead.

What is the unified AEO strategy for all three engines?

The best AEO strategy does not pick one engine — it serves all three. The unified playbook is:

1. Maintain evergreen pillar pages with original data and definitive claims. These win on ChatGPT (entity authority + original claims) and provide the foundation for Claude and Perplexity citations too. Update them quarterly to keep last-updated dates fresh.

2. Add [FAQ schema](/blog/faq-schema-the-underrated-ai-citation-signal) and short, quotable answer blocks to those pillar pages. FAQ schema benefits all three engines but especially Claude and Perplexity. The quotable paragraphs serve Claude's verbatim quoting style. Each page should have 3-5 FAQ pairs in FAQPage JSON-LD.

3. Publish fresh takes and updates regularly, with clear headings and primary source citations. New content (or refreshed content with a recent last-updated date) wins on Perplexity. Freshness also helps Claude for time-sensitive queries and signals to ChatGPT that the page is actively maintained.

4. Build entity authority in parallel. Get into Wikidata. Get cited by industry publications. Publish original research. These signals compound over time and are essential for ChatGPT citations, which weight entity authority heavily. See our entity SEO guide for the full playbook.

5. Track per-engine citation share. Use the AI Visibility Checker monthly. Identify which engine you are weakest on and adjust your content strategy accordingly. If ChatGPT is lagging, invest more in entity authority. If Perplexity is lagging, invest more in freshness and structure.

Do all five and you will see citation share lift across all engines. The underlying content principles — directness, structure, entity clarity, recency, schema, primary sources — are universal, as described in our AEO guide. Get the fundamentals right and you will win on all three.

How has ChatGPT's citation behavior changed in 2025?

ChatGPT's citation behavior has evolved significantly between 2024 and 2025. Understanding these changes is important for AEO strategies that were calibrated on 2024 data.

Browse mode is now default. In 2024, ChatGPT only accessed the web when users explicitly enabled Browse with Bing. In 2025, GPT-4o automatically searches the web for queries where training data is insufficient. This means ChatGPT now cites more often — our 2025 dataset shows 1.3 citations per answer, up from an estimated 0.6 in 2024.

Citation formatting has improved. Earlier versions of ChatGPT Browse cited sources as vague references ("according to a source..."). Current ChatGPT includes inline links and names the source more often. This makes ChatGPT citations more valuable for brand visibility than they were in 2024.

Search quality has improved. ChatGPT's integration with Bing has improved, and it now retrieves more relevant results. This is good news for AEO — better retrieval means more opportunities to be cited. It also means that standard Bing SEO signals (title tags, meta descriptions, URL structure) matter more for ChatGPT than they did in 2024.

Remaining limitation: ChatGPT still frequently answers from training data for well-known topics, especially in its default mode. 37% of answers in our dataset had zero citations. This is unlikely to change — ChatGPT will always prefer its training data for topics it "knows well." The AEO opportunity is in the long tail — specific, niche queries where ChatGPT needs to search the web.

How has Perplexity's citation behavior evolved?

Perplexity has been the most citation-friendly engine since launch, and its 2025 behavior continues to favor extensive, well-structured citations.

Sonar Large is more accurate. Perplexity's current model (Sonar Large) produces more accurate citations than the 2024 version. In our 2024 tests, ~12% of Perplexity citations pointed to pages that did not support the claimed fact (hallucinated citations). In 2025, this rate dropped to ~4%. Better citation accuracy means AEO practitioners can trust Perplexity citations more.

Citation density has increased. Perplexity now averages 4.2 citations per answer, up from an estimated 3.5 in late 2024. The increase comes from more aggressive re-retrieval — Perplexity now fetches additional sources when initial results are insufficient, rather than synthesizing from a smaller set.

Referral traffic has grown. Perplexity served 500M+ queries in 2024 and continues to grow. The referral traffic from perplexity.ai is now significant — our case study shows it as the largest single AI referral source for the startup, driving 60% of AI referral visits. Track this in GA4.

Focus API availability. Perplexity now offers an API (Sonar API) that developers can use to build applications with Perplexity's citation-rich answers. This expands the surface area for AEO — your content can be cited not just on perplexity.ai but in any application built on the Sonar API.

The AEO implication: Perplexity remains the most important engine for AEO practitioners because it offers the most citation surface area, sends the most referral traffic, and is the most responsive to content structure and freshness. Optimize for Perplexity first, then adapt for Claude and ChatGPT.

What are the limitations of this research?

No dataset is perfect. Here are the limitations of our 5,000-citation analysis and how they affect the conclusions.

Non-deterministic citation behavior. AI assistants are non-deterministic — the same prompt can produce different citations on different runs. We ran each prompt 3 times and took the median, but there is still variance. The citation rates we report are probabilities, not certainties. You are optimizing for a higher probability of being cited, not a guarantee.

Temporal snapshot. Our data was collected in March 2025. AI models are updated frequently — GPT-4o, Sonnet 4, and Sonar Large will all be replaced by newer versions. Citation patterns may shift with model updates. We recommend re-running this analysis quarterly.

Category distribution. Our 20 categories are weighted toward B2B SaaS, software, and technology topics. Citation patterns may differ for health, legal, or financial topics where answer engines apply stricter source-selection criteria (e.g., only citing medical journals for health queries). Apply these findings with caution outside the categories we tested.

Prompt design. Our prompts were designed to be neutral and representative of real user queries, but prompt wording affects citation behavior. Longer, more specific prompts tend to produce more citations. Shorter, vaguer prompts tend to produce fewer. Your actual citation share will depend on the specific prompts your audience uses.

Missing Google AI Overviews. This study covers only ChatGPT, Claude, and Perplexity — not Google AI Overviews (which is a GEO surface, not an AEO surface). Google AI Overviews has different citation patterns that we plan to analyze in a future study.

Key takeaways for cross-platform AEO

The 5,000-citation dataset is clear: engines are different, but the underlying content principles are universal. Directness, structure, entity clarity, recency, schema markup, and primary source citations improve citation probability on all three engines. The differences are in emphasis and degree, not in kind.

• Perplexity offers the most citation surface area (4.2 citations/answer). Optimize for structure and freshness to win here.
• ChatGPT has the most users (250M+ WAU). Optimize for entity authority and original data to win here.
• Claude is the most likely to quote you verbatim. Optimize for quotable paragraphs and FAQ schema to win here.
• Comparison tables are the highest-citation format — especially on Perplexity (2.8x citation rate).
• FAQ schema benefits all engines but especially Claude (+65%) and Perplexity (+80%). Ship it first.
• Entity authority compounds over time and is essential for ChatGPT citations. Start building it immediately.
• Per-engine measurement is non-negotiable. Aggregate citation share hides engine-specific gaps. Track each separately.
• The unified strategy wins. Structure every page for all three engines: evergreen depth (ChatGPT) + quotable blocks (Claude) + fresh citations (Perplexity).