What AI assistants cite: citations in ChatGPT, Claude, Gemini and Perplexity answers

Citation data from 1,730 answers that ChatGPT (gpt-4o-mini with web search), Claude (claude-haiku-4-5 with web search), Gemini (gemini-2.5-flash with Google Search grounding) and Perplexity (sonar) returned through their official APIs between 7 September and 4 October 2026. The answers cover 81 questions about 7 subjects (2 brands, 1 product, 1 public figure, 3 topics) in English, French, Italian, German and Spanish, sampled in a pilot and four weekly cycles. The dataset holds 19,713 citations reduced to the registered domain, as aggregate counts: no answer text, no question text, no URLs. The subjects' own websites are replaced by a placeholder. Files give citation rates and citations per answer by engine, language, subject type, question category and cycle, with 95% intervals; the share of answers citing each source family (news, Wikipedia, the subject's own site, reviews, Reddit, YouTube, forums, editorial, other); the overlap of cited domains across engines; and every domain cited at least twice. Main findings: Perplexity cited sources in 100% of answers, Gemini in 99.6%, Claude in 90.5% and ChatGPT in 54.8% (only when it ran a web search). Perplexity returned 21.7 citations per answer, ChatGPT 2.5. For the same question, 77.2% of cited domains were cited by one engine only and 0.9% by all four. Limits: mid-size API models, not the consumer apps; seven subjects in a few categories; some language cells are small. Full method and tables: https://tonecast.io/research/what-ai-assistants-cite