Who do ChatGPT and Perplexity cite when people ask about health and supplements?

Which kinds of sites get cited?

Each cited domain was sorted into one of seven types. Perplexity leaned harder on medical and academic sources; ChatGPT search cited fewer sources overall and more small blogs.

The mix depends on the question. On the 30 informational questions (for example, does zinc help with colds), 47% of citations were medical, academic or government sources. On the 16 buying questions (for example, what is the best magnesium for sleep), health publishers took 30%, general media 25%, review and ranking sites 16% and brand sites 13%. Medical sources fell to 7%.

Do AI answers cite supplement brands' own websites?

Rarely, and almost never the brands they recommend. Supplement brands were named in 19 of the 100 answers, including 12 of the 32 answers to buying questions. The sources behind those recommendations were mostly publishers, magazines and review sites. One leading practitioner supplement brand was named six times across five questions; its website did not appear as a source once. Of all the brands named in answers, only two had their own site cited in the same answer.

We also asked about four brands by name: a probiotic, a greens powder, a hair-growth supplement and a lab-testing membership. For the probiotic and the greens powder, neither engine cited the brand's own website; the answers drew on review sites, magazines and Healthline. For the hair-growth brand, both engines cited its site, and ChatGPT linked the brand's own pages about its clinical studies. For the lab-testing membership, Perplexity cited the brand's site and ChatGPT did not.

When ChatGPT did link to a brand's own site, the page was an article, buying guide, study summary or press release, never a product page. The only product pages it linked were reseller listings. The most-cited brand site in the study, a long-established supplement maker, was cited four times; both times ChatGPT linked it, the page came from the brand's article library.

Across all 100 answers, only four of the 50 well-known health brands we checked in a separate homepage scan had their own website cited at all, for eight citations in total.

What should a health brand take from this?

Three things. First, the product page is not what gets cited. Answer-first articles, guides and study summaries on the brand's own domain are. Second, most of a brand's AI visibility is decided on other people's sites: the publishers, magazines and review sites above. Being named there is how a brand ends up in the answer even when its own site is not a source. Third, the engines disagree. A brand that only checks ChatGPT, or only Perplexity, sees part of the picture.

None of this means brand sites cannot be cited. They can, for narrow questions that the big publishers have not answered well. That is how CalmGut, a new brand, began appearing in Google's AI features. See how AI engines choose sources and GEO vs SEO for what that work involves.

How was this measured?

We wrote 50 questions a consumer might ask, across gut health, sleep, energy and focus, women's health, fitness, longevity, immunity, lab testing and telehealth, and buying supplements: 30 informational, 16 buying and 4 about a named brand. Each question ran once through the same pipeline as the live check on our homepage: OpenAI's API with the web search tool (model gpt-4o-mini) and Perplexity's Agent API (fast preset), with a short instruction to answer as a careful health assistant would and cite its sources. We recorded every cited domain, then classified each one by hand.

What are the limits?

This is one run on one evening. AI answers change from run to run and day to day, so treat the shares as a snapshot, not a constant. The API versions of ChatGPT and Perplexity are close to, but not the same as, the consumer apps people use. The source types are our own classification, and some sites could reasonably sit in another group. The 50 questions are ours; a different set would give different numbers. Brand names in answers were matched against a fixed list of supplement and health brands. We describe individual brands by category rather than by name; in the dataset, brand names in the four brand questions are replaced the same way.