Ask ChatGPT, Gemini and Perplexity the same question and they will often cite different pages. That much is well established. What is less clear from the growing number of published comparisons is how the engines differ, and whether the differences are stable enough to plan around. We reviewed the cross-engine studies we could find that state a method and a sample, opened each one, and compared what they actually measured.

The studies

Industry datasets, all from companies that sell search, PR or AI-visibility software, none with public raw data:

  • Profound: 680 million citations from ChatGPT, Google AI Overviews and Perplexity, August 2024 to June 2025.
  • Semrush: weekly citation snapshots for more than 230,000 prompts on ChatGPT search, Google AI Mode and Perplexity, July 14 to October 12, 2025.
  • Ahrefs, three studies: overlap with Google and Bing rankings for 15,000 long-tail prompts on ChatGPT, Gemini, Copilot and Perplexity (published August 2025); the age of about 17 million cited URLs across those assistants and AI Overviews (July 2025); and AI Mode versus AI Overviews on 540,000 query pairs from September 2025.
  • Muck Rack: more than 25 million links from ChatGPT, Claude and Gemini across 17 industries (May 2026 edition).
  • Yext: 6.8 million citations from ChatGPT, Gemini and Perplexity for location-based queries in four industries, July and August 2025.

Academic and journalistic audits:

  • Chen et al., University of Toronto (preprint, September 2025): the API versions of GPT-4o search, Perplexity's Sonar Pro, Claude 3.5 Sonnet with web search and Gemini 2.5 Flash with Google Search grounding, compared with Google on product-ranking prompts across verticals and languages.
  • Grossman et al. (preprint, April 2026): 11,500 queries run through Google Search, AI Overviews and Gemini 2.5 Flash.
  • Nguyen et al. (preprint, August 2026): 15,942 citations from ChatGPT, Perplexity and AI Overviews for mental health questions in seven languages.
  • The Tow Center for Digital Journalism (March 2025): 1,600 queries testing whether eight AI search tools could correctly attribute news excerpts.
  • The EBU and BBC (October 2025): 2,709 core responses from the free versions of ChatGPT, Copilot, Gemini and Perplexity, assessed by journalists at 22 public service media organizations.

Where the studies agree

The engines rarely cite the same pages. Ahrefs found that AI Mode and AI Overviews, both Google products, cited the same URLs only 13.7% of the time for the same query, although their answers were 86% semantically similar. Grossman et al. found average Jaccard similarity below 0.2 between the sources used by Google Search, AI Overviews and Gemini. Chen et al. found GPT-4o search reused almost none of the same domains when a question was asked in different languages, while Claude reused far more.

Perplexity sits closest to classic search. In Ahrefs' 15,000-prompt test, 28.6% of Perplexity's citations ranked in Google's top 10 for the same query, against about 8% for ChatGPT, Gemini and Copilot. Chen et al. found Perplexity drew on a broader mix than GPT-4o search or Claude, with more brand and social sources, and returned somewhat fresher links for automotive queries. It also leans on community sites: in Profound's data, Reddit accounted for 46.7% of citations among Perplexity's ten most-cited domains.

ChatGPT leans on reference and editorial sources, and shifts abruptly. Profound found Wikipedia made up 47.9% of citations among ChatGPT's top ten domains. Muck Rack reports that ChatGPT cited sources in 96% of the responses it analyzed and leaned on Wikipedia. Ahrefs found ChatGPT cited the newest content of any assistant, 458 days newer on average than Google's organic results. Semrush documented the volatility: ChatGPT cited Reddit in close to 60% of responses in early August 2025 and around 10% by mid-September, with a similar fall for Wikipedia, while AI Mode and Perplexity stayed far more stable.

Gemini scores worst on sourcing in accuracy audits. In the EBU and BBC study, 72% of Gemini responses had significant sourcing issues, against 24% for ChatGPT and 15% each for Copilot and Perplexity. The Tow Center found more than half of the responses from Gemini and Grok 3 cited fabricated or broken URLs. Both audits used news questions, where sourcing standards are strict.

Where they disagree

Reddit and community content. Profound and Semrush put Reddit at or near the top for Perplexity and Google's AI surfaces, and Muck Rack says Gemini leans on Reddit. Yet Chen et al. found almost no social content in AI engines' answers to product-ranking prompts, and Yext found forums accounted for about 2% of citations. The likeliest explanations are prompt type (Chen's "top 10 brands" prompts and Yext's location queries invite editorial and directory answers), interface (Chen queried APIs, not consumer apps) and timing, given the shifts Semrush recorded.

Earned media versus brand-owned pages. Chen et al. and Muck Rack find earned media dominant, with Muck Rack putting it at 84% of citations. Yext finds 86% of citations in brand-managed websites and listings, and reports engine differences within that: Gemini leaned on first-party websites (52.1% of its citations) and ChatGPT on listings (48.7%). Questions about where to go and questions about what to buy draw on different parts of the web.

Freshness. Ahrefs found ChatGPT the freshest assistant and Perplexity among the older-leaning ones on average age; Chen et al. found Claude and GPT-4o search returned older automotive content than Perplexity. The measures differ: Ahrefs used last-published dates from its own index across millions of URLs, while Chen et al. crawled and dated a much smaller set of links in two verticals.

How often each engine cites at all. Muck Rack's figures, citations in 96% of ChatGPT responses, 82% for Gemini and 55% for Claude, come from prompts the firm chose and does not describe in its post. Similarweb has reported web citations in 6.8% of ChatGPT conversations in May 2026 (US, desktop). The gap shows how far test prompts can sit from ordinary use.

Why method matters more than headlines

Five choices explain most of the divergence:

  • Interface. API models with search tools are not the consumer products, which add their own retrieval, personalization and presentation.
  • Prompts. Synthetic ranking prompts, location queries, news questions and real user conversations produce different citation mixes.
  • Counting unit. Share of all citations, share of responses citing a domain, and share within a top-10 list are different numbers. Profound's 47.9% for Wikipedia on ChatGPT is a top-10 share; Wikipedia's share of all ChatGPT citations in the same data was 7.8%.
  • Time. Engines change retrieval without notice, as Semrush's September 2025 data shows.
  • Repetition. SparkToro found fewer than one in 100 repeated runs of the same prompt returned the same list of brands, and Ahrefs reports that about 45% of AI Overview citations change between generations. A single snapshot understates the variance.

Most industry studies are produced by firms whose products depend on the answer, and their data cannot be independently checked. The academic studies are smaller and often run on APIs, but they describe their methods fully, and some, such as the mental health audit by Nguyen and colleagues, release their annotated data.

How to read the next comparison

Before acting on a cross-engine headline, check which interface was tested, where the prompts came from and what they asked, what unit the percentage describes, when the data was collected, and whether each prompt was run more than once. A finding that survives those questions is worth planning around. Most published comparisons today answer only some of them.