How ChatGPT, Perplexity, Google AI Mode, and Claude Each Decide What to Cite

Last updated: 2026-10-13

TL;DR

• ChatGPT answers from a mix of parametric knowledge and live web search, and cites Wikipedia at 12.1 percent of its sources in the Analyze AI 83,670-citation study [1][2].
• Perplexity retrieves live for almost every answer, numbers its citations inline, and cites Reddit at about 6.6 percent of its total citations [3][2].
• Google AI Mode uses “query fan-out,” splitting a question into many sub-queries, and YouTube is the single most-cited domain at roughly 23 percent of AI Overview citations in Surfer SEO’s 2025 analysis of 36 million AI Overviews [4][5].
• Claude’s web search tool always includes citations when it searches, and LinkedIn appeared zero times across the Analyze AI study period [6][2].
• The engine, not the query, drives most of the variance in who gets cited. A single answer can look completely different across the four engines on the same prompt.

If a buyer, investor, or candidate ran your company name through ChatGPT, Perplexity, Google AI Mode, and Claude this morning, they would get four answers that disagree with each other. In our August 2026 audit of past TwoSetAI guests, 95 percent appeared on a Google surface and 86 percent were cited by at least one AI engine. The engines did not cite in parallel: Google AI Mode and Perplexity leaned on the video and the watch page; When ChatGPT did not cite a YouTube video, it reached guests through third-party authoritative posts (founder profiles, twosetai.com pages, Medium posts) [7]. Same guests, same episodes, four different pictures.

This is the question underneath every citation gap: each engine has its own retrieval design, its own training mix, and its own default sources. If you don’t know which engine favors what, you’ll optimize for the wrong one. Here is how each of the four main engines picks what to cite, drawn from the vendors’ own documentation and from the largest citation studies published to date. Then I’ll close with what each engine means for your site, and what has to live outside it. For the broader thesis these four engines sit inside, see the parent pillar on the Citation Gap.

How does ChatGPT decide what to cite?

ChatGPT holds two kinds of knowledge: what’s baked into the model’s weights from training (parametric), and what it fetches live when it searches. OpenAI’s October 2024 launch of ChatGPT search described it as “answers with links to relevant web sources” using third-party search providers and partner content [1]. In practice, that means ChatGPT only cites when it searches. If it answers from memory alone, there are no sources attached, and you have no idea where the claim came from.

The sources it does cite skew toward encyclopedic and authoritative references. Rodrigo Stockebrand’s O’Reilly book on answer engine optimization, published July 2026, pulls the headline from Analyze AI’s analysis of 83,670 AI citations: Wikipedia at 12.1 percent of ChatGPT citations is the single largest share [2]. LinkedIn is a surprising second: Semrush’s March 2026 study of 89,000 LinkedIn URLs cited by AI search tools found that ChatGPT Search cited LinkedIn in 14.3 percent of responses, the highest rate of any engine studied [8].

Two implications for a founder. First, if your Wikipedia entity is thin or not disambiguated, you lose a measurable share of ChatGPT real estate on brand-adjacent questions. Second, long-form LinkedIn posts of 500 to 2,000 words make up the bulk of cited LinkedIn URLs in Semrush’s data [8], and ChatGPT reads them. A founder’s authored articles on LinkedIn are not vanity. The engines are reading them, and that is one of the mechanics behind the Founder-Voice Moat.

The honest limit: ChatGPT does not always search. A Profound analysis cited in a 2026 industry write-up reported Claude invoking live search for 36.6 percent of tested prompts [9], and the analogous ChatGPT number is not published cleanly. When ChatGPT doesn’t search, you get the parametric answer and no citation. That is the “dark funnel” lane Stockebrand describes, where AI influences a buying decision but never leaves a trace in your analytics [2].

How does Perplexity decide what to cite?

Perplexity is the most retrieval-anchored of the four. Its own help center describes it as a system that searches the web in real time and returns responses grounded in current information with numbered citations [3]. Vendor analyses in 2026 describe a three-stage pipeline: live retrieval against a freshness-focused index, candidate ranking by relevance and recency, then synthesis with citations attached to each claim [10]. I treat the pipeline specifics as directional rather than definitive, since Perplexity has not published a full architecture paper.

What makes Perplexity different is also what makes it predictable. Its citation set overlaps more with Google’s organic top 10 than any other AI engine does. The Ahrefs Brand Radar study of 15,000 long-tail queries, reported in Stockebrand’s book, found about 28 percent URL overlap between Perplexity and Google top 10, compared with 12 percent for ChatGPT, Gemini, and Copilot [2]. If you rank for a query in Google, you have a better shot at a Perplexity citation than at any other engine.

Perplexity’s platform mix also tilts toward communities. Analyze AI’s 83,670-citation data put Reddit at about 6.6 percent of total Perplexity citations [2]. Semrush found Perplexity cited LinkedIn in only 5.3 percent of its responses, the lowest among the three engines Semrush tested, and that company pages made up as much as 59 percent of its LinkedIn citations when it did [8]. The pattern is consistent: Perplexity rewards third-party corroboration, review-site facts, and community discussion over opinion content.

How does Google AI Mode decide what to cite?

Google AI Mode and the AI Overview feature in classic Search both use a technique Google calls “query fan-out.” Google’s own developer documentation describes AI Mode as issuing multiple related searches across subtopics and data sources, then synthesizing an answer from the results and attaching the pages that supported it [4]. Each sub-search runs its own retrieval. A page can be cited for a sub-question the user never literally typed [11].

The platform mix skews heavily to video. Surfer SEO’s analysis of 36 million AI Overviews and 46 million citations, reported in Stockebrand’s Chapter 6, found YouTube led at roughly 23 percent of citations, Wikipedia followed at about 18 percent, and Reddit, LinkedIn, and community forums made up a meaningful remainder [2][5]. In Semrush’s 89K-URL study, Google AI Mode cited LinkedIn in 13.5 percent of responses [8]. For a sibling deep-dive on the YouTube share, see How AI Cites YouTube Videos.

For a founder, Google AI Mode is both the biggest opportunity and the hardest to game. Because fan-out breaks a question into many sub-queries, a page that cleanly answers one specific sub-question (for example, “how do real-time retrieval systems handle outdated data” rather than just “what is RAG”) can earn a citation even when its site has no ranking power on the main query. Google says there is no special markup required: the page just needs to be indexed and snippet-eligible [4]. Schema is not required, but Microsoft’s October 2025 guidance confirms it helps AI systems understand content, and controlled experiments across ChatGPT, Claude, Gemini, Perplexity, and Google AI Mode show pages with schema plus matching visible content get noticed and extracted at measurably higher rates than visible-content-only pages. The mechanism and the confidence levels on each engine are unpacked in How Schema Markup Might Actually Work in AI Search [19]. That is the shape of the opportunity, and why we spend time on the difference between being indexed and being cited by AI.

How does Claude decide what to cite?

Claude is the most conservative of the four. Anthropic’s web search tool documentation describes a model that “automatically generates search queries, retrieves relevant results, and incorporates them into its responses with citations” when a request depends on information that is current, changing, or outside its training data [6]. Citations are not optional. If Claude searches, Claude cites. If Claude doesn’t need to search, Claude answers from training and does not fabricate a citation.

The platform mix shows the conservatism. In Analyze AI’s 83,670-citation study, LinkedIn appeared zero times in Claude’s citations across the study period, and Wikipedia was only 0.1 percent of citations [2]. One Profound analysis reported Claude invoking live web search for just 36.6 percent of tested prompts [9]. The practical implication: Claude cites fewer sources, searches less often, and shows a strong preference for primary documentation (official docs, academic references, canonical category explainers) over social or community content.

For a founder, this cuts two ways. You cannot buy your way into Claude’s answers with LinkedIn posts. You can earn citations by having canonical, structured, primary documentation of your product, category, and claims where Claude will find them. If your docs page, your pricing page, or your category primer is the clearest statement of what you do, Claude is more likely to repeat it verbatim than any other engine.

What does each engine cite, by platform?

The clearest way to see the differences is side by side. These are the dominant sources per engine, drawn from the Analyze AI 83,670-citation dataset, Surfer SEO’s 36M-AIO analysis, and Semrush’s 89K-URL LinkedIn study, as compiled in Stockebrand’s O’Reilly book, Chapter 6 [2][5][8]:

Engine	Dominant platform	Share of citations	Notable gap
Google AI Mode / AI Overviews	YouTube	~23% (Surfer SEO, 2025)	Heavy fan-out means long-tail sub-questions matter
ChatGPT	Wikipedia	12.1% (Analyze AI)	LinkedIn appears in 14.3% of responses (Semrush)
Perplexity	Reddit	~6.6% (Analyze AI)	~28% URL overlap with Google top 10
Claude	Primary docs and canonical sources	LinkedIn 0% (Analyze AI); Wikipedia 0.1%	Searches for only ~37% of prompts (Profound)

Three cross-engine findings sit behind this table. First, the engines overlap less with each other than most SEO frameworks would predict: only about 12 percent of URLs cited by ChatGPT, Gemini, and Copilot also appear in Google’s traditional top 10 for the same query, with Perplexity the live-retrieval outlier at 28 percent [2]. Second, the citation set is drifty: SparkToro’s 2026 study of 2,961 prompt runs across ChatGPT, Claude, and Google’s AI found the chance of two identical brand lists for the same question was under 1 in 100 for ChatGPT and Google [12]. Third, these are directional industry snapshots, not laws. Methodologies vary, several studies come from vendors that sell AI-visibility products, and the engines retune often.

What does this mean for your site?

The practical cut, per engine, for a pre-Series B founder:

• For ChatGPT: make sure your Wikipedia entity is accurate (or exists), and that your LinkedIn Articles tell your category story in 500 to 2,000 words [8]. Your own site is maybe 5 to 10 percent of what ChatGPT retrieves, per McKinsey, so this is largely an off-site job [13].
• For Perplexity: rank for your branded and category queries in Google first, since Perplexity’s overlap with Google top 10 is highest [2]. Earn third-party corroboration on Reddit, review sites, and your own company page, because that is what Perplexity reads most heavily.
• For Google AI Mode: publish long-form video with transcripts and chapters, since YouTube is 23 percent of Overview citations [5]. On your own site, write pages that answer specific sub-questions clearly. Query fan-out rewards pages that stand alone on a narrow sub-topic rather than one monolithic page.
• For Claude: invest in canonical primary documentation: your docs, your pricing, a clear “what is X” page for your category, and dated fact sheets. Claude searches less, cites conservatively, and prefers primary sources when it does.

Across all four, the single mechanism that compounds is the one Stockebrand calls “wide beats deep” [2]: engines place more weight on a claim when multiple independent sources agree with it. The practical consequence is that a founder saying the same specific thing across a podcast, a LinkedIn article, a Medium essay, and a company page is corroboration the engines can see, which is why the tier an engine puts you in is set by what it finds across all four, not by any one page. For more on that tier, see the Buyer Bucket.

What should you do this week?

Pick one engine where you are not yet cited, pull five queries a counterparty would type, and run them. Note which URLs that engine cited, and which ones were not yours. Those non-owned URLs are the fastest leverage: they are the pages the engine already trusts in your category. For the full version of this exercise, I walk through how to audit your Citation Gap in 30 minutes.

The engines will keep changing. The underlying mechanism, that each one assembles an answer from a thin slice of retrieved pages and that your site is a small share of them [13], is what makes the per-engine work worth doing. When your buyers ask AI, be the answer.

Who is TwoSetAI?

TwoSetAI is one of the top AI podcast and video studios in Silicon Valley, and the team behind this article, hosted by Angelina Yang. The show has featured 100+ AI founders and leaders on how real AI products get built, sold, and trusted. Alongside the podcast, TwoSetAI runs the Founder Visibility Program that turns each conversation into the articles, clips, and pages that make its guests easy to find. Watch full episodes on YouTube at youtube.com/@TwoSetAI, and find every guest and episode at twosetai.com.

Glossary

• Query fan-out (also called: multi-query retrieval, sub-query expansion). Google’s technique in AI Mode of breaking a question into subtopics and running many searches at once, then synthesizing an answer from the results.
• Parametric knowledge (also called: in-weights knowledge, training-time knowledge). What a language model knows from training data, held in its weights, rather than what it retrieves live from the web.
• Live retrieval (also called: real-time retrieval, retrieval-augmented generation, RAG). The pattern where an engine fetches current web pages at query time, then grounds its answer in them.
• Citation share (also called: platform share, source share). The percent of an engine’s cited URLs that come from a specific domain or platform.
• Dark funnel (also called: zero-click AI influence). Stockebrand’s term for AI-influenced buying journeys that never show up in your analytics because the citation never gets clicked.

About Angelina Yang

Angelina Yang is VP of AI, a fast.ai fellow, and the host of the TwoSetAI podcast, where she has interviewed 100+ AI founders and leaders about how real AI products get built, sold, and trusted. Find her on LinkedIn, talk to her AI clone on Delphi, watch the podcast on YouTube, or browse twosetai.com.

Frequently Asked Questions

How does ChatGPT decide which websites to cite? When ChatGPT searches, it fetches live web pages via third-party search providers and partner content, then writes an answer with links [1]. In Analyze AI’s 83,670-citation dataset reported by Stockebrand, Wikipedia was 12.1 percent of ChatGPT citations, the single largest share [2]. ChatGPT does not always search, and when it answers from training alone there is no citation attached.

Why does Perplexity give different sources than Google? Perplexity retrieves live and ranks for recency and relevance. Its URL overlap with Google’s top 10 is about 28 percent on long-tail queries, which is higher than other AI engines but still means most of its citations are different URLs [2]. Reddit and company pages are over-represented in its mix.

What is Google AI Mode’s query fan-out? It is Google’s own term for splitting a question into subtopics and running many related searches at once. The answer is then written from those retrievals, and the citations are the pages that supported each sub-search [4]. A page can be cited for a sub-question the user never literally typed, which is why writing pages that cleanly answer narrow questions matters.

Does Claude always cite its sources? When Claude uses the web search tool, yes. Anthropic’s documentation says Claude incorporates web results into responses with citations [6]. When Claude answers from training alone, no citations are attached. One industry analysis found Claude searched for only 36.6 percent of tested prompts [9].

Why do these engines all give different answers to the same question? They have different training, different retrieval designs, and different default sources. In SparkToro’s 2026 study of 2,961 runs across ChatGPT, Claude, and Google’s AI, the chance of the same brand list twice for ChatGPT and Google was under 1 in 100 [12]. Expect per-engine variance, and audit all four.

Does ranking in Google help me show up in AI answers? It helps most with Perplexity, which shares roughly 28 percent URL overlap with Google top 10. It helps less with ChatGPT, Gemini, and Copilot, where overlap is around 12 percent [2]. Treat classic ranking as a starting point, not a sufficient condition.

Does Claude cite LinkedIn or Wikipedia? In Analyze AI’s dataset reported by Stockebrand, Claude cited LinkedIn zero times across the study period, and Wikipedia at 0.1 percent of citations [2]. Claude’s citation mix skews toward primary documentation and canonical sources.

References

[1] OpenAI (2024). “Introducing ChatGPT search.” October 31, 2024. https://openai.com/index/introducing-chatgpt-search/

[2] Stockebrand, R. (2026). Answer Engine Optimization. O’Reilly Media, July 2026. https://learning.oreilly.com/library/view/answer-engine-optimization/0642572275549/

[3] Perplexity (2026). “How does Perplexity work?” Perplexity Help Center, accessed October 13, 2026. https://www.perplexity.ai/help-center/en/articles/10352895-how-does-perplexity-work

[4] Google (2026). “AI features and your website.” Google Search Central, accessed October 13, 2026. https://developers.google.com/search/docs/appearance/ai-features

[5] Surfer SEO (2025). AI Overview citations analysis (36 million AI Overviews / 46 million citations). Summarized in Stockebrand (2026), Chapter 6.

[6] Anthropic (2026). “Web search tool.” Claude API documentation, accessed October 13, 2026. https://platform.claude.com/docs/en/agents-and-tools/tool-use/web-search-tool

[7] TwoSetAI (2026). Full-catalog guest search audit, checked August 10 to 25, 2026, incognito protocol. Internal data; protocol in Methodology note.

[8] Loktionova, M. / Semrush (2026). “We Analyzed 89K LinkedIn URLs Cited in AI Search: Here’s What Drives Visibility.” Semrush Blog, March 10, 2026. https://www.semrush.com/blog/linkedin-ai-visibility-study/

[9] Something Inc. (2026). “AI Content Strategy: Claude vs ChatGPT,” summarizing a Profound analysis of Claude live-search invocation rate. 2026. https://somethinginc.com/blog/ai-content-strategy-claude-vs-chatgpt/

[10] LLMPulse (2026). “How Does Perplexity Work? How It Finds, Ranks and Cites Sources.” Accessed October 13, 2026. https://llmpulse.ai/blog/how-perplexity-works/

[11] Google (2025). “AI Mode in Search: our most powerful AI search yet.” Google Blog, May 2025. https://blog.google/products/search/ai-mode-search/

[12] Fishkin, R. and O’Donnell, P. (2026). “New Research: AIs are highly inconsistent when recommending brands or products.” SparkToro, January 28, 2026. https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/

[13] McKinsey & Company (2025). “New front door to the internet: Winning in the age of AI search.” October 2025. https://www.mckinsey.com/capabilities/growth-marketing-and-sales/our-insights/new-front-door-to-the-internet-winning-in-the-age-of-ai-search

[14] Lewis, P. et al. (2020). “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.” NeurIPS 2020; arXiv:2005.11401. https://arxiv.org/abs/2005.11401

[15] Muck Rack (2026). “What Is AI Reading? New Insights May 2026.” https://muckrack.com/blog/what-is-ai-reading-new-insights-may

[16] Liu, N. F., Zhang, T., and Liang, P. (2023). “Evaluating Verifiability in Generative Search Engines.” Stanford; EMNLP 2023; arXiv:2304.09848. https://arxiv.org/abs/2304.09848

[17] Aggarwal, P. et al. (2024). “GEO: Generative Engine Optimization.” KDD 2024; arXiv:2311.09735. https://arxiv.org/abs/2311.09735

[18] OtterlyAI (2026). “YouTube AI Citation Study 2026.” Updated March 2, 2026. https://otterly.ai/blog/youtube-ai-citation-study-2026/

[19] Yang, A. (2025). “How Schema Markup Might Actually Work in AI Search.” Medium, December 17, 2025 (republished at twosetai.com/blogs/how-schema-markup-might-actually-work-in-ai-search). https://angelina-yang.medium.com/how-schema-markup-might-actually-work-in-ai-search-d017b1d2ef83

Methodology note. Last verified 2026-10-13. Per-engine behavior comes from each vendor’s own documentation, dated above. The citation-share numbers come from third-party aggregate studies (Analyze AI 83,670 citations; Surfer SEO 36 million AI Overviews / 46 million citations; Semrush 89,000 LinkedIn URLs; SparkToro 2,961 prompt runs), as compiled in Rodrigo Stockebrand’s July 2026 O’Reilly book on answer engine optimization, Chapter 6. Several of these studies come from vendors that sell AI-visibility products, so treat the specific percentages as directional rather than causal. The TwoSetAI audit [7] covered a cohort of past guests, not the full catalog, with about seven queries per guest run in incognito windows between August 10 and 25, 2026. “Citation share,” “parametric knowledge,” and “live retrieval” are standard industry terms. “Dark funnel” is Stockebrand’s own term. Nothing here guarantees a citation, a ranking, or a business outcome.