A new study, AI’s Sources of Truth: How Chatbots Cite Health Information, analyzed 5,472 citations generated by the four leading web-enabled models: ChatGPT, Claude, Gemini, and Perplexity. The findings show both encouraging signs of reliability and some concerning blind spots. More importantly, they suggest how our relationship with healthcare information is being rewritten by AI systems.
The Concentrated Core of AI’s Health Sources
When chatbots answer health questions, their citations are surprisingly concentrated. The most frequently cited domain across all models was PubMed Central, a free archive of biomedical research, which appeared 385 times in the sample. AI systems currently lean heavily on peer-reviewed research that’s openly available.| Rank | Website | Total mentions |
|---|---|---|
| 1 | pmc.ncbi.nlm.nih.gov | 385 |
| 2 | my.clevelandclinic.org | 174 |
| 3 | www.mayoclinic.org | 163 |
| 4 | www.ncbi.nlm.nih.gov | 150 |
| 5 | www.sciencedirect.com | 93 |
Close behind were some of the internet’s most trusted health websites. The Cleveland Clinic’s patient information portal was cited 174 times, and the Mayo Clinic’s site 163 times. Another top source was the NIH’s National Center for Biotechnology Information (NCBI) site, with 150 mentions. These four show that chatbots gravitate toward established, credible medical knowledge.
Overall, nearly one in three citations (30.7%) in the study came from health media sites. About 23% of references were traced to commercial or affiliate sites (like corporate blogs, product pages, or other pages with a marketing slant). Another roughly 23% were from academic research sources. The chatbots as a group seem to favor accessible, consumer-friendly explanations of health topics. Traditional news articles made up only about 3.7% of citations, and social media or user-generated content only 1.6%. Mainstream journalism and personal anecdotes thus barely register in the bots’ answers.
Fresh, Up-to-Date Information in Answers
When it comes to how current the information is, the chatbots show a strong bias toward recent material. Nearly two-thirds of all cited sources were published in either 2024 or 2025. In fact, the single most common publication year among the citations was 2025, accounting for about 40% of all references. After 2025, the number of citations from older years drops off dramatically.This recency bias likely reflects both the design of the bots (some have browsing enabled to find current info) and a built-in preference for newer, more relevant data. If you ask about a medical treatment or emerging health issue, the chatbots are inclined to cite something from the last year or two, rather than a decades-old paper. It is a reassuring habit given how quickly medical consensus can change.
Different Chatbots, Different Source Preferences
The most interesting insight from the study is how each AI model has its own style in sourcing information. While all four chatbots broadly favored authoritative, recent, open-access material, the mix of sources varied by platform.
For example, ChatGPT and Claude showed the strongest preference for highly authoritative domains. Around 68% of all citations from ChatGPT came from domains with the highest domain authority rankings (like DR 81–100 on Ahrefs), and Claude was similar at 67.4%. In comparison, Google’s Gemini and Perplexity were a bit less top-heavy: about 56–58% of their citations were from these elite top-rated sites. Gemini and Perplexity dipped more into mid-tier sources (for instance, websites that are reputable but not the absolute top of the internet’s authority food chain), and Perplexity in particular ventured the furthest down the credibility ladder. The study notes that Perplexity cited the largest share of low-authority websites (3.3% of its sources were from domains in the lowest credibility tier).
Looking at content categories: ChatGPT tended to cite health media outlets the most, with 35.8% of its references coming from sites like Mayo Clinic, WebMD, Cleveland Clinic, etc. About 23% of ChatGPT’s citations were academic papers or journals, meaning it still included a fair amount of hard science but leaned more toward those consumer health explainers. Claude, by contrast, was more evenly split, roughly 29.7% health media and 28.9% academic sources, essentially balancing between easy-to-read guides and original research.
Gemini stood out by citing government and NGO sources far more than the others. Nearly a quarter (24.9%) of Gemini’s citations were from official public health sites or nonprofit health organizations. Meanwhile, Perplexity was the real outlier. It’s the only model where commercial content was the number-one source category, making up 30.5% of its citations. Perplexity also cited social or user-generated content more than any other bot. This chatbot is a bit more likely to throw in a Reddit thread, a Quora answer, or a YouTube video as part of an answer.
The Future of Health Search
The shift from Google-style search to AI-powered health assistants is behavioral. Instead of wading through a swamp of links, users now get tailored explanations, neatly cited, with bias toward accessibility and recency.- Trust is being redefined. People may start trusting AI models as much as, if not more than, traditional search engines. Yet each model’s sourcing bias means users could receive subtly different “truths.”
- Paywalled research is at risk of invisibility. If LLMs overwhelmingly favor open-access content, cutting-edge but gated science could be sidelined from public discourse.
- Media narratives may shape science. With 59% of citations coming from summaries and health media, the interpreters of science could become more influential than the researchers themselves.
- Transparency matters. LLMs cite live, working links is a step toward accountability, but users must still validate the credibility and intent of those sources.
by Irfan Ahmad via Digital Information World
