What ATOVIA measured itself

What AI Cites for Health, Mind and Money — September 2026, 75 Korean Questions Checked

A value ATOVIA asked and tallied ourselves. It's our own in-house survey — not academic research or a third-party study. If you cite it, please note that.

More people now turn to AI instead of a search engine when something hurts or life gets hard. But what those answers are actually based on is barely known. We put 75 questions into generative AI (Gemini) — things like "I can't sleep because of my atopic dermatitis" or "I'm burned out, how do I recover on my own without counselling" — and recorded every source attached to the answers.

Snapshot

① Even with a source listed, you often can't click through

Media names under an answer were often just text, not a link. With only a name to go on, you'd have to search again yourself to find which article it meant.

This matters most for health information: having a source and being able to check it are two different things. Just because a media name shows up in an AI's answer doesn't mean it actually read that article.

② The evidence differs between asking about «an illness» and asking about «a service»

"Everything else" was mostly a company or service's own website. In other words, a service's own writing about itself got cited more often than third-party sources like academic societies or public institutions.

Ask about a symptom or condition, and academic societies and public health centres show up. Ask about a service — "where's good?" — and that service's own website and press coverage show up instead. That said, please read this as a trend, not a fact — the sample is small: some categories have single-digit counts.

When you want health information from AI, asking "what standard do you judge this by" works better than "recommend one for me" if you want an answer backed by third-party evidence.

③ A service that makes the recommendation list is one that's «been written about somewhere»

We asked again about services named in the answers — "where was this featured?" — 20 follow-up questions, and found 23 places. Press articles, official blogs and community posts all showed up fairly evenly.

In other words, an AI's recommendation list is less "best first" and more "whatever content survives online." Missing from the list doesn't mean a service is bad, and being on it doesn't mean it's verified. We don't publish individual service or media names.

④ Each AI leaned on «different evidence»

Even the same question got different kinds of evidence depending on which AI you asked. One leaned toward a company's own website and press coverage; the other leaned toward academic databases like PubMed and JAMA.

So when you get health information from AI, don't trust just one. Put the same question into two or more, and you'll see right away how different the evidence can be.

How we measured it — so you can repeat it

1. Ask without naming a brand. Name one, and you just get that brand's own description back — you can't measure what the AI reaches for first.

2. Save the whole answer word for word. Save just the link, and you lose what ① is checking for.

3. Sort sources into categories (public/academic, press, community, listicle, platform).

4. A question that got no answer doesn't get counted. Scrape a half-loaded screen and you get "0 sources" — recording that as a real zero would make the whole dataset false.

What this edition leaves out

• A comparison with Perplexity and Naver — for now it's just Gemini and GPT
• Monthly change — our first edition
• A wider range of question topics (for now: wellness, health, psychology, finance)

We didn't fill in numbers we don't have. If we measure them next time, we'll add them then.

How we grade this

• Made by: ATOVIA (the operator)
• Nature: an in-house check — not a statistically designed survey
• Scope: 75 Korean-language questions (Gemini 62 · GPT 13)
• We don't publish individual media or service names
• No claims about numbers, benefits or timeframes
• This is not medical advice. If you have symptoms, please see a medical professional.