What do AI assistants cite? 19,713 citations from ChatGPT, Claude, Gemini and Perplexity

What are the key findings?

• Citing is not a given. Perplexity and Gemini cited in practically every answer; Claude in 9 answers out of 10; ChatGPT in 55%, and only when it had actually run a web search. When ChatGPT answered from memory, it cited nothing.
• The volume differs by an order of magnitude. Perplexity returned 21.7 citations per answer, Gemini 11.7, Claude 9.7, ChatGPT 2.5.
• The engines read different webs. Of the 4,850 question-and-domain pairs, 77.2% were cited by a single engine and 0.9% by all four. ChatGPT's citations overlapped with any other engine's on about 5% of domains per question.
• News sites are the most shared family. Between 23% (ChatGPT) and 74% (Perplexity) of answers cited at least one news site. Wikipedia appeared in 47.5% of Claude's answers and in 5.6% of ChatGPT's.
• Reddit and YouTube depend on the engine. Gemini cited Reddit in 21.9% of answers and YouTube in 23.4%; Perplexity cited Reddit in 13.6% and YouTube never; ChatGPT and Claude cited Reddit in none of 850 answers.
• The subject's own website is cited more by Gemini and Perplexity (in 38.9% and 35.6% of answers) than by Claude and ChatGPT (16.9% and 13.5%).
• The long tail is long. 2,417 distinct domains were cited; 861 of them once. The ten most cited domains took 38% of ChatGPT's citations and 19% of Perplexity's.

How often does each engine cite a source?

Each engine answered the same questions, through its official API, with web search or grounding enabled. A citation is a link the API returned attached to the answer. Rates come with 95% confidence intervals (Wilson); means with a 95% interval under the normal approximation.

ChatGPT's rate has a simple explanation: its API used the web search tool in 236 of 431 answers, exactly the 54.8% that carried citations. The other 195 answers came from the model's own knowledge and cited nothing. When ChatGPT did search, it cited 4.6 pages on average.

Which kinds of pages do AI assistants cite?

Every cited page was classified by Tonecast's page families: news and blogs, Wikipedia, the subject's own website, review platforms, Reddit, YouTube, forums and Q&A sites, high-citation editorial domains, and other web pages. The table gives the share of answers that cited at least one page of each family, with the 95% interval.

Counted as a share of citations rather than of answers, the picture is flatter: pages outside the named families take two thirds of every engine's citations (66.8% for ChatGPT, 69.3% for Claude, 65.7% for Gemini, 68.3% for Perplexity). News sites follow, with 10.5% to 19.6%. Wikipedia is 11.3% of Claude's citations and 1.4% of Perplexity's. The subject's own site takes between 3.2% (Claude) and 10.3% (Gemini) of citations.

Reddit is the clearest split. Gemini and Perplexity cite it; ChatGPT and Claude did not cite it once in 850 answers. Whether an engine can reach a platform's pages at all, and whether it chooses to, matters more than the platform's popularity.

Do the four engines cite the same pages?

For every question, the study took the set of domains each engine cited across all its answers, and compared the sets. The 81 questions produced 4,850 question-and-domain pairs.

The overlap between two engines, measured as the share of domains they had in common for the same question (Jaccard index, averaged over questions), was 16.3% for Claude and Gemini, 15.8% for Claude and Perplexity, 15.7% for Gemini and Perplexity. ChatGPT sits apart: 5.6% with Claude, 4.8% with Gemini, 4.9% with Perplexity. Part of this is volume, since ChatGPT cites far fewer pages; part is retrieval, since each engine runs its own searches.

Which domains are cited most?

The fifteen most cited domains, the subjects' own websites excluded. Counts are citations, across all engines, with the split per engine. "Subjects" says for how many of the seven subjects the domain was cited at all.

Only three domains were cited for every subject: Wikipedia, Reddit and YouTube, and even those by different engines. Several national news sites were cited by one engine alone: lefigaro.fr and corriere.it almost only by Perplexity, factually.co only by Claude, facebook.com almost only by Gemini. The family labels are Tonecast's: a newspaper classified as "Web" is one its connector catalogue had not yet filed under news.

The subjects' own websites, taken together, received 1,575 citations, 8.0% of the total. The brand's site was the single most cited domain of the study (935 citations), ahead of Wikipedia, because the questions were about that brand. The study does not name the subjects; the dataset replaces their domains with their subject type.

How concentrated are citations?

The fewer pages an engine cites, the more it concentrates on a short list. ChatGPT's ten most cited domains took 38% of its citations; Perplexity's, 19%. For a brand, this means that the pages to care about differ by engine: a short, stable list for ChatGPT, a long and changing one for Perplexity.

Does the language of the question matter?

The questions were asked in English, French, Italian, German and Spanish, each with a country. For Gemini and Perplexity, the citation rate stayed at or near 100% in every language. The two engines that do not always cite varied with the language, with small samples outside English.

ChatGPT searched, and therefore cited, most often in English and least often in German; with 52 German answers the interval is wide (20 to 44%). Claude did the opposite: it cited in nearly every French and German answer and least in Spanish. The language of the question changes not only which pages are cited but whether anything is.

Does the type of question matter?

Each question belongs to one of six types: discovery ("which companies…"), comparison ("X or Y"), evaluation ("is X any good?"), pricing, support ("how do I…") and current events. For ChatGPT, the type decided whether it searched at all.

ChatGPT searched for prices and news, and never for a comparison: all 69 comparison answers came without a single citation. Claude and Gemini did the opposite on comparisons, which drew their highest number of citations per answer (14.5 and 18.3). Perplexity cited about the same volume whatever the question. For a brand, the practical consequence is that "X or Y" questions are answered by ChatGPT from what it learned in training, where no page can be cited or corrected, and by Gemini from a wide set of fresh pages.

Is this stable from week to week?

The same questions were asked in four weekly cycles, after a small pilot on 7 September. The citation rates barely moved: ChatGPT cited in 53.8%, 54.7%, 53.8% and 54.7% of answers, Claude in 90.2%, 90.2%, 90.2% and 90.6%. The number of citations per answer drifted for one engine.

Perplexity went from 19 to 24 citations per answer in four weeks, and Claude from 8.4 to 10.5. Four points are not a trend, but they are a reminder that an engine's citing behaviour is a moving target: a number measured once has no error bar and no shelf life.

How was this measured?

• Subjects. Seven subjects tracked for Tonecast's own validation: two brands, one consumer product, one public figure and three public-affairs topics. The study is about the sources, not the subjects, which are not named.
• Questions. 81 questions in English, French, Italian, German and Spanish, each with a country (United States, United Kingdom, Ireland, France, Italy, Germany, Spain), across the six question types above.
• Engines and models. OpenAI gpt-4o-mini with the web search tool; Anthropic claude-haiku-4-5 with web search; Google gemini-2.5-flash with Google Search grounding; Perplexity sonar. All through the official APIs, with the country passed where the API allows it. No consumer app was scraped.
• Sampling. One or two answers per question, engine and cycle; four weekly cycles from 13 September to 4 October 2026, plus a pilot of 28 answers on 7 September. 1,836 calls, of which 1,730 returned an answer; the 106 failed calls (5.8%, 83 of them Gemini on 13 September, re-run the next day) are excluded.
• Citations. The links the API returned with the answer: annotations for OpenAI and Anthropic, grounding chunks for Gemini, the citations list for Perplexity. 19,713 in total. Each link was reduced to its registered domain (so every Wikipedia language edition counts as wikipedia.org), and classified by family with Tonecast's connector catalogue; "another web page" is every page not in a named family.
• Statistics. Rates carry 95% Wilson intervals; means carry a normal 95% interval. Overlap is the Jaccard index of the domain sets per question, averaged over questions. The AI visibility calculator computes the same intervals.

What are the limits?

• Models. These are the API models Tonecast samples, at the mid-size tier. Larger models, the consumer apps and future versions can cite differently; the week-to-week drift above shows that they do.
• Subjects. Seven subjects in a handful of categories. The most cited domains reflect those categories (smartphones, air travel, European politics); the engine-level patterns are the transferable part.
• Classification. Two thirds of citations fall in "another web page", including some newspapers. Finer families would change the family table, not the engine comparison.
• Small cells. Below about 60 answers the intervals are wide; the language and question-type tables say how many answers each cell has.

Where is the data?

The tables behind this page are published as CSV files under the Creative Commons Attribution 4.0 licence. They contain counts only: no answer text, no question text, no URL. The subjects' own websites are replaced by their subject type.

• by_engine.csv: answers, citation rate with interval, citations per answer, distinct domains, per engine.
• by_engine_language.csv, by_engine_campaign_type.csv, by_engine_prompt_category.csv, by_engine_cycle.csv: the same measures per language, subject type, question type and weekly cycle.
• source_types_by_engine.csv and first_citation_type_by_engine.csv: page families per engine, and the family of the first citation of each answer.
• overlap.csv: how many engines cited each question-and-domain pair.
• domains.csv: every domain cited at least twice, with its family, total and per-engine citations, the share of each engine's answers citing it, and the number of subjects.
• README.txt: column descriptions and licence.

The same files are archived on Zenodo with a permanent DOI, 10.5281/zenodo.23613190, and mirrored on Hugging Face (with a table viewer) and Kaggle.

To cite: Tonecast (2026). What AI assistants cite: citations in ChatGPT, Claude, Gemini and Perplexity answers. https://tonecast.io/research/what-ai-assistants-cite. doi:10.5281/zenodo.23613190. Questions about the data: support@tonecast.io. The study will be refreshed as the weekly cycles continue; the date at the top is the date of the data.

Questions about this study