ChatGPT cites roughly 10.4 sources per response in 2026, against 21.9 for Perplexity, according to citation data from Discovered Labs and Whitehat SEO. Measured differently, by unique domains rather than total citations, an Attrifast study of 1,200 buyer intent prompts run across April and May 2026 found median citation density of 6.4 unique domains on Perplexity, 3.6 on Claude, 3.1 on ChatGPT, and 2.4 on Gemini. Both measurements point at the same conclusion: the set of sources any single AI answer draws from is small, it is dominated by a handful of recognizable domains, and getting into it is a different problem from ranking on Google.
Two more numbers reframe the difficulty. Research from 5W Public Relations found Wikipedia at 13.15 percent and Reddit at 11.97 percent of all ChatGPT citations in the United States, together more than a quarter of everything the model cites, while the Wall Street Journal, New York Times, and Bloomberg did not appear in the top 20 at all. And only about 11 percent of domains are cited by both ChatGPT and Perplexity, which means a single content strategy aimed at “AI search” as one surface is aimed at something that does not exist.
How many sources does each AI engine cite per answer?
Each engine has a distinct citation budget, and knowing yours changes where you spend effort. Here are the 2026 figures by engine, using both measurement methods, since the two count different things.
1. Perplexity: roughly 21.9 citations, 6.4 unique domains
Perplexity is the most generous citer by a wide margin, roughly double ChatGPT on total citations and nearly double on unique domains. It runs retrieval on almost every query and surfaces its source list openly. That makes it the easiest engine to break into and the best place to measure whether your content is retrievable at all.
2. ChatGPT: roughly 10.4 citations, 3.1 unique domains
ChatGPT cites about half as often as Perplexity and concentrates heavily. It also does not run retrieval on every query, answering many from parametric knowledge with no sources at all. When it does cite, Wikipedia and Reddit take a disproportionate share.
3. Claude: 3.6 unique domains
Claude sits between ChatGPT and Gemini on citation density and tends toward fewer, more substantive sources per answer rather than long lists.
4. Gemini: 2.4 unique domains
Gemini is the most concentrated of the four, citing the fewest unique domains per answer. Google AI Overviews and AI Mode behave similarly, leaning on a small set of sources that already rank well in Google’s index.
Knowing the citation math is only useful if you know where you currently sit in it. Get your free AI visibility audit and see which prompts already cite you and which ones name a competitor.
Which domains take most of the citations?
A small set of platforms absorbs a disproportionate share, and they are not the ones most brands assume. Wikipedia accounts for an estimated 47.9 percent of ChatGPT’s top-10 source share, while Reddit accounts for roughly 46.7 percent of Perplexity’s top-10 share. Across engines overall, community platforms including Reddit and Quora capture about 52.5 percent of citations against 47.5 percent for brand owned domains.
The recurring names are consistent across studies: Wikipedia, Reddit, YouTube, LinkedIn, Forbes, Quora, and G2 or comparable category review platforms. Major newspapers are conspicuously underrepresented, largely because of licensing arrangements and crawler blocking rather than quality.
That distribution tells you something uncomfortable. If more than half of citations go to community and reference platforms, then a strategy consisting only of publishing on your own domain is competing for less than half the available slots. Presence on the platforms engines already trust is not a supplement to owned content. It is the larger share. Our guides on using Reddit for AI visibility and Wikipedia for AI visibility cover the two biggest ones specifically.
Why is the overlap between engines only 11 percent?
Because each engine runs a different retrieval stack over a different index with different source preferences, so “getting cited by AI” is four or five separate problems. Roughly 11 percent domain overlap between ChatGPT and Perplexity is the measured reality, and it explains why brands report being highly visible in one engine and absent from another.
The mechanics differ concretely. ChatGPT search grounds against Bing’s index plus OpenAI’s own crawling, which makes Bing indexation a prerequisite rather than an afterthought. Perplexity runs its own crawler and index with heavy weighting toward forums and recent content. Google AI Overviews and AI Mode ground in Google’s index, so classic ranking still governs eligibility there. Copilot draws on Bing. Claude uses its own retrieval with a preference for substantive primary sources.
The practical consequence is that the diagnostic step comes before the content step. Confirm you are indexed in Bing, confirm your robots.txt does not block GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot, or Google-Extended if you want to be eligible, and confirm your pages render server side. Plenty of brands invisible in ChatGPT have no content problem at all. They have a Bing indexation problem. Our how to get indexed on Bing guide covers the fix.
What makes a page one of the sources an engine picks?
Engines pick pages that contain a directly extractable answer near the top, attached to a recognized entity, corroborated elsewhere. Those three properties, extractability, entity strength, and corroboration, explain most of the variance between pages that get cited and pages that do not.
Extractability is the mechanical part. A retrieval system chunks your page, embeds the chunks, and matches them against the query. A chunk that reads “The exam preparation market was valued at $74.2 billion in 2026” is a self contained fact that can be lifted into an answer. A chunk that reads “Our approach is tailored to your unique needs” is not extractable at any length. This is why lists, tables, defined terms, stated numbers, and question format headings outperform narrative prose so consistently.
Entity strength is whether the engine has a resolved concept of who you are. That comes from consistent naming across your site, schema markup that declares your organization and authors, a Wikipedia or Wikidata presence where warranted, and repeated co-occurrence of your brand name with your topic across independent sources.
Corroboration is whether your claim appears anywhere other than your own site. Engines discount unverified self referential claims. The same fact appearing in a trade publication, a review platform, a Reddit thread, and a YouTube description is treated very differently from the same fact appearing once on your own homepage.
How many citations do you actually need to win?
Fewer than most people expect, because the slot count is small and most competitors are not trying. If ChatGPT cites about three unique domains for a given prompt and Gemini cites about two, then being one of five to eight recognized sources on a specific topic is sufficient to appear regularly.
That reframes the goal usefully. The target is not “rank for a keyword” but “be one of a handful of sources an engine considers authoritative on a narrow question.” That is achievable on specific questions and near impossible on broad ones. No small brand becomes a top source on “marketing.” Plenty become a top source on “how much does a hail damage roof inspection cost in Kansas” or “what evidence does a hair relaxer claim require.”
Measure it accordingly. Track share of prompts where you appear rather than average position, run the same prompt set monthly across ChatGPT, Perplexity, Gemini, Claude, and Copilot, and record which competitor names appear alongside or instead of yours. Tools including Otterly.ai, Profound, and Peec AI do this at scale, and a spreadsheet of 30 prompts run manually does it well enough to start. Our AI share of voice guide covers the measurement framework.
Frequently asked questions
How many sources does ChatGPT cite per answer?
ChatGPT cites roughly 10.4 sources per response according to 2026 data from Discovered Labs and Whitehat SEO, or about 3.1 unique domains per answer using the median citation density method from Attrifast’s 1,200 prompt study. The difference reflects what is counted: total citation links versus distinct domains. ChatGPT also does not run retrieval on every query, answering many from parametric knowledge with no citations at all.
Which AI engine cites the most sources?
Perplexity, by a wide margin. It averages about 21.9 citations per response against ChatGPT’s 10.4, and a median of 6.4 unique domains against 3.1 for ChatGPT, 3.6 for Claude, and 2.4 for Gemini. Perplexity runs retrieval on nearly every query and displays its source list openly, which makes it both the easiest engine to get cited by and the best one to use for diagnosing whether your content is retrievable.
What websites does ChatGPT cite most often?
Wikipedia and Reddit dominate. Research from 5W Public Relations found Wikipedia at 13.15 percent and Reddit at 11.97 percent of all US ChatGPT citations, together over a quarter of the total, with Wikipedia estimated at 47.9 percent of top-10 source share. YouTube, LinkedIn, Forbes, and Quora also recur. Notably, the Wall Street Journal, New York Times, and Bloomberg did not appear in the top 20, largely due to licensing and crawler access rather than quality.
Why am I cited by Perplexity but not by ChatGPT?
Because only about 11 percent of domains are cited by both. Each engine runs a different retrieval stack over a different index. ChatGPT grounds substantially against Bing plus OpenAI’s own crawling, so if you are not well indexed in Bing you are effectively ineligible regardless of content quality. Perplexity runs its own crawler with heavy weighting toward forums and recent content. Check Bing indexation and crawler permissions before assuming it is a content problem.
How many sources do I need to be to appear in AI answers?
Being one of roughly five to eight recognized sources on a narrow topic is usually enough to appear regularly, since engines cite only two to six unique domains per answer. The practical implication is to compete on specific questions rather than broad ones. No small brand becomes a top source on a category level term, but many become a top source on a precise question with a checkable numeric answer that nobody else has bothered to publish.
Does blocking AI crawlers affect whether you get cited?
Directly. Blocking GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot, or Google-Extended in robots.txt removes you from the eligible set for those engines. Many sites block these by default through a CDN setting or a security plugin without anyone deciding to, then conclude their content is not competitive. Auditing robots.txt and server logs for these user agents is the cheapest first diagnostic in any AI visibility project.
Citation counts are small, which means the gap between appearing and not appearing is narrower than it looks. Run a free AI visibility audit and see exactly how many of your target prompts return your name today.
The citation math is the whole strategy in miniature. When an engine assembles an answer from two to six domains, visibility is not a gradient. You are in the set or you are not, and the set is chosen from sources that are extractable, entity-resolved, and corroborated somewhere other than your own website. Most brands lose that contest on mechanics rather than merit: blocked crawlers, missing Bing indexation, and pages full of language no retrieval system can lift. Fix those, publish a checkable answer to a question nobody else answered precisely, and the small size of the citation set stops being a barrier and starts being the advantage.
Tagged