Same Questions, Two AIs: Clinic Citation Rates of 9% and 37%
Published
This article was first published in Korean by Intrix Lab on August 13, 2026.
We asked two AIs the same questions on the same day, and the results differed fourfold. Across 450 dermatology clinics nationwide in South Korea, we asked each model 480 questions seeking options, such as "good dermatology clinics in [location]." Official domains owned by clinics accounted for 9.1% of the sources Gemini cited and 37.5% of those GPT-5 cited (measured on 2026-08-13).
This article examines where that gap comes from and whether it answers the question, "Will improving a clinic's website increase its citations?"
First, we withdrew our previous article
On the morning of 2026-08-13, we published a column on the same topic and took it down that day. The figure in its title was wrong.
We had reported "9 official clinic sites out of 1,674 sources, or 0.5%." The actual figure was 9.8%. The error occurred during aggregation, rather than measurement. Clinic-owned domains were not being identified correctly, understating the share by a factor of about 18.
What made this defect troubling was its silence. A smaller directory simply produces fewer matches without triggering an error. With no warning on screen, it returns a plausible figure suggesting "almost no citations." Because that figure happens to be interesting, it can even reduce the motivation to verify it. We probably would have caught it sooner if it had contradicted our expectations.
What did the two models cite?
The citation shares differed, and so did the sources themselves.
For Gemini, booking and comparison platforms dominated the top positions: my-doctor.io with 541 citations, doctornow with 479, tistory with 322, and cashdoc with 220.
For GPT-5, professional societies and official clinic websites ranked highest. The Korean Dermatological Association (derma.or.kr) had the most citations at 436, followed by the Association of Korean Dermatologists (akd.or.kr) at 178. Individual clinic websites came next.
The lists of clinics appearing in answers also differed. Gemini cited 63 clinics and GPT-5 cited 91, with 44 appearing in both. Of the combined set of 110 clinics, only 40% overlapped.
Will improving a website increase its citations?
To measure this, we separately checked the websites of 450 clinics. The important point is that we selected the pages to measure without looking at the citation data. Measuring cited pages to explain citations would be circular.
We statistically examined the relationship between those pages' condition and citation counts.
GPT-5 showed associations, but the details were unexpected. The statistically significant factors were basic infrastructure rather than content quality: whether the site responded normally, whether it had an h1 tag, and whether it had a robots.txt file. Those were the three. Medical schema (JSON-LD) was not statistically significant.
For Gemini, none of the factors was statistically significant. Apart from a weak association with body text length, website condition explained virtually nothing. This was consistent with most citations going to platforms.
Correction: measuring indexing changed the conclusion (added 2026-08-13)
After publishing the analysis above, we directly measured Google index counts that same day. We are correcting the interpretation in the preceding section.
In the initial analysis, we had not measured indexing, so we used "whether the site loads" as a proxy. That led us to write that "website condition explains almost nothing for Gemini." But when we actually measured indexing, it was the strongest explanatory variable for Gemini.
We counted indexed pages for each clinic and checked whether its main treatment page was indexed. Before using the measurements, we checked their error against measurements from our own site.
Here are the results: the citation-count multipliers when each variable was entered on its own.
Gemini: published page count, 1.43 times; indexed page count, 3.67 times; main treatment page indexed, 9.53 times.
GPT-5: published pages, 1.11 times; indexed pages, 1.87 times; main page indexed, 2.48 times.
The picture becomes clearer when published and indexed page counts are entered together. For GPT-5, the effect of publishing volume disappears (1.06 times, not statistically significant). What appeared to be a publishing-volume effect was a reflection of indexing. The association with body text character count also disappears. For Gemini, indexing is the main pathway, but publishing volume also retains an independent association.
This changes the order of priorities. Ensuring that one existing treatment page is indexed comes before creating more pages.
Here is the current situation. Of the 450 clinics, 146 (32%) had no indexed pages at all, and 216 (48%) had their main treatment page indexed. More than half had not gotten the very page that should be cited into the index.
There are limitations. We counted indexed pages only up to a set cap, so we do not know the slope above that range. This is also observational. Well-indexed sites tend to be well maintained, so this design cannot separate the effect of indexing itself from overall site maintenance.
The strongest signal was not optimization
Of the 450 clinics, 134 (29.8%) had websites that could not even be accessed. There were 104 with no response, 15 returning 404, 9 returning 403, and 4 returning 400.
Excluding these clinics changes the results. For GPT-5, even the sign of the coefficient for body text length reverses. In other words, much of the association seen for GPT concerns "whether the site is functioning." Among functioning sites, those with less body text were actually cited more often.
The practical order is clear. Before schema or content strategy, check whether the site loads.
How far can we trust these results?
A few points need to be clear.
The population is narrow. It is a complete census of 450 facilities nationwide in South Korea that met all these conditions: dermatology as the listed specialty under Health Insurance Review and Assessment Service classifications, at least 1 board-certified dermatologist, and an official website. There is no guarantee that other medical specialties or industries would produce the same results.
Each model was measured in 1 round. Results also fluctuate between rounds. When we ran the same survey twice, 4 days apart, the clinic-owned domain shares were reproduced at 9.80% and 9.40%, but only 60% of the actual cited clinic lists overlapped. A single measurement is not enough to say, "Our clinic gets cited by AI." The aggregate is stable; the individual lists are not.
Naver was excluded. We could not include Naver AI Briefing because it has no official API. The absence of Naver remains a limitation of research on the Korean market.
This is an observational study, not a causal one. Larger institutions may have both functioning websites and more third-party exposure. This design cannot separate those three factors.
We changed the design twice after collecting the data. One preregistered main explanatory variable, pages that are empty shells without JavaScript, applied to only 5 clinics. That was too few for estimation, so we excluded it. We also included the 134 inaccessible sites as having 0 body text characters rather than discarding them as missing data. Both changes and their reasons are recorded in the preregistration document's change history.
Frequently asked questions (FAQ)
Q. Which AI should we optimize for?
The recommendations differ. For GPT, there are areas you can address by improving your own website. For Gemini, the key is how you appear on booking and comparison platforms. This study's conclusion is that "AI optimization" cannot be treated as one thing.
Q. Will adding schema increase citations?
In this dataset, medical schema (JSON-LD) was not statistically significant for either model. That does not mean you should avoid it. It means this study did not establish an effect. Measuring causality requires a randomized controlled experiment, which we are running separately.
Q. Can you check whether our clinic is cited?
We do not recommend reporting a conclusion from a single measurement. As noted above, 40% of the list changed in 4 days. Repeated measurements are needed. Even then, what we can say is limited to "it was cited this many times during this period."
GEO and SEO results are affected by external factors, including search and AI service policies, competition, and website condition. Intrix does not guarantee specific rankings or citations.