Are AI-Readable Clinic Websites Cited More? Reanalyzing 450 Dermatology Practices
Published
This article was first published in Korean by Intrix Lab on October 8, 2026.
They were cited more, but the difference was smaller than a research paper reported. As of 2026-10-07, we reanalyzed AI citation data for 450 dermatology practices. When asked directly about a practice by name, AI cited the practice's own domain in 26.1% of citations for AI-readable websites, compared with 13.9% for other websites. The share of responses citing the practice's own domain at least once was 1.77 times as high. The direction matches the paper's claim that "78% of answers for agent-readable websites come from the company's own pages," but the magnitude is much smaller.
What did the paper claim?
A paper posted on arXiv in September 2026 proposed AX (agent experience) as the next standard for AI answer optimization. It argued that, as AI agents read websites directly to build answers, how easily agents can read a website becomes a competitive advantage. Analyzing 1,056 companies, it reported that 78% of answers for agent-friendly websites were based on the companies' own pages and that they received 1.9 times as many recommendations.
Intrix checked whether a difference in the same direction appeared for Korean-language medical practice websites. Without making new AI calls, we recombined dermatology citation data collected in August with website inspection results.
How did we define a readable website?
We checked whether the website loaded normally, did not block major AI bots, and included sufficient body content in the HTML sent by the server. We classified websites meeting all criteria as "readable." This group contained 166 practices.
The "poorly readable" group contained 239 practices: 134 with website errors, 28 blocking bots, 75 with insufficient body content, and 2 that could not be reached. We excluded 29 intermediate cases that could not clearly be assigned to either group. We fixed this definition before calculating the results.
Does asking for a practice by name lead to citations of its own website?
| Practice-name queries, repeated 3 times | Readable | Poorly readable |
|---|---|---|
| Own-domain share of citations | 26.1% (155 practices, 4,320 citations) | 13.9% (210 practices, 5,563 citations) |
| Responses citing the practice's own domain | 66.3% | 37.5% |
The difference in own-domain citation share was 12.2 percentage points, with a 95% interval of 7.3–16.9 percentage points that did not cross 0. The share of responses citing the practice's own domain was 1.77 times as high (interval: 1.46–2.17 times). Even when people searched by name, AI used content from elsewhere instead of the practice's website much more often for poorly readable websites.
Does simply having a website that loads make a difference?
That accounted for a substantial part of the difference. When we excluded websites with errors from the poorly readable group and compared only websites that loaded, the gap narrowed to +8.2 percentage points and 1.48 times. Conversely, restricting the analysis to clinic-level practices yielded 28.7% versus 11.8%, with a ratio of 2.25 times. We interpret this as website condition having a larger effect for small clinics because larger hospitals have more news articles and portal documents available for citation outside their own websites.
Was the pattern the same for local recommendation queries?
It differed by model. These data came from asking 3 types of "recommend a dermatology practice" questions 10 times each across 16 areas.
| Local recommendation queries | Readable | Poorly readable | Ratio |
|---|---|---|---|
| Model A | 3.07% | 1.45% | 2.12 times (interval: 1.00–4.98) |
| Model B | 7.03% | 6.53% | 1.08 times (interval: 0.61–1.89) |
Model A's ratio was similar in size to the paper's 1.9 times, but the lower end of its interval touched 1.00. Model B showed no difference. This means that, for questions from customers who do not know a practice's name, some models are influenced more by factors other than website condition.
To check whether AI can read your practice's website, you can request a free website audit. To examine who gets cited in local recommendation queries and why, continue with a first measurement.
How far can these results be trusted?
These are observational data, not a randomized experiment. Practices with readable websites tend to be larger, longer-established, and better known. Looking separately at clinic-level practices is only a rough adjustment, so we cannot say that fixing a website would increase citations by the same amount.
Our definition is narrower than the paper's. The paper used a composite score that included llms.txt and alternative Markdown documents, while we examined only four signals. The outcome measures also differ. The paper used evaluators to assign the share of answer sentences originating from different sources; we measured citation links. Practice-name queries were collected from one model in a single day, and dermatology was the only industry covered. We could not verify the claims of a 64% increase in costs and 41% higher accuracy with our data. This analysis does not guarantee any particular citation or ranking.
Frequently asked questions (FAQ)
Q. What should we check first on our practice's website?
Start by checking whether the website loads normally, whether robots.txt blocks AI bots, and whether the body content is included in the HTML. In these data, a substantial part of the gap came from whether the website loaded.
Q. Why is this so different from the paper's 78%?
The measurements differ. The paper used evaluators to assign the share of answer content originating from the company's own pages, while we counted the own-domain share of citation links. The industry and language also differ. The comparison extends only to the direction of the results.
Q. Will fixing our website get us into local recommendations?
These data cannot establish that. For local recommendation queries, only Model A showed a difference; Model B showed none. Name queries and recommendation queries need to be checked separately.