A Census of Taiwan's Dental Clinic Websites: AI Won't Risk Citing 9 Out of 10 Claims

IDAEO Data Column・AI Visibility Field-Test Series № 003｜Data Snapshot 2026-07-28 (JST)｜IDAEO Data Column Editorial Team (AI-assisted・Human-reviewed)

Of the 7,002 National Health Insurance-contracted dental clinics in Taiwan, 543 have their own website — the other 6,459 have none at all, so AI has no material to work with.

We checked those 543 websites sentence by sentence. Only about 1 in ten claims has evidence behind it; for the remaining 9 out of ten, AI can read them but will not risk citing them. This does not mean those 9 out of ten are lies. It means they are written in a form AI cannot cite.

Three headline findings:

• About 1 in ten—the share of website claims that can be supported by dental literature
• 43 times—the difference in daily AI reads between a clinic with properly structured data and the average for the strongest 8 clinics in the directory
• 94%—the share of dental clinics across Taiwan whose public ranking entries remain blank

Definitions for the raw logs, the scoring methodology, and links to the full charts appear under “Data Definitions” and in the ranking link at the end of this article. Every figure can be traced back and audited. No estimated or forecast values are used anywhere in this article.

01 Patients Are Already Asking AI

Twenty years ago, patients relied on two things to find a dentist: recommendations from family and friends, and signs in the neighborhood. Ten years ago, the process moved to Google: search for “recommended dentist,” read review scores, and compare clinic websites one by one. Clinics learned to rank, buy keywords, and manage reviews.

Now the behavior has changed again. Patients pick up their phones and ask ChatGPT directly: “Which Taipei clinic is good for dental implants? Don't show me ads.” “My child is terrified of pain. Find a pediatric dentist who is patient.” “Find a dentist open on holidays that takes credit cards.”

AI does not hand them ten links and wait for them to compare—it gives them an answer, with one to three clinics at a time.

Here is the biggest difference: you could see when you lost before—your ranking fell to a later page, or no one clicked your ad. Now you lose somewhere you cannot see. It is not that the patient compared you and chose someone else. The patient never knew you existed.

02 Census Result: Among the 543 Clinics With a Website, AI Can Confidently Use Only 1 in Ten Claims

What do those 9 out of ten claims look like? They generally take forms like these:

• “Professional team”
• “Advanced equipment”
• “Patients come first”
• “Pain-free treatment”

There is no source to check, no matching literature, and not enough specificity to verify the claim. A person may nod along; AI will not risk using it. ("Will not risk using" is a mechanism-based inference — AI citation requires checkable sources — not a sentence-by-sentence measurement of AI behavior; what this article measured is the literature-support rate.)

AI has to stand behind any sentence it cites. It will not answer a patient's question with “our technology is industry-leading.” It needs information with a traceable source.

A Verifiable Digital Identity

In plain language: other people can trace what you say back to a source. The clinic name, dentist names, and public credentials match across pages, and each has its own source.

“Verifiable” means a field includes a source that can be checked. It does not mean identity certification by a competent authority or an assessment of healthcare quality.

03 The Difference Between SEO and AEO in One Sentence

• SEO: the assignment for the Google era—get yourself near the top of search results.
• AEO: the assignment for the AI era—give AI information about you that it can find, access, and reconcile when answering.

Even if you rank first, patients will not see you if you are absent from the AI's answer.

04 Taiwan's Dental AI Ranking: Which Tier Is Your Clinic In?

We turned the census results into a public ranking. All 7,002 National Health Insurance-contracted dental clinics in Taiwan are listed: if a clinic is in the registry, it is on the ranking. Anyone can look up their own clinic or a peer.

Current progress:

• 7,002 clinics—the nationwide dental registry, all of which are on the ranking
• 543 clinics—have their own website (the other 6,459 have none, leaving AI no material)
• 429 clinics (6%)—already in the scoring pipeline
• 339 clinics (5%)—verifiability scoring completed
• 94%—still blank on the ranking; this is not a low score, but the absence of any data to score

Every AI visit on the ranking comes from a raw log recorded request by request at the network edge by Cloudflare and stamped with its official verification—not an estimate and not a score we calculated ourselves.

Infrastructure for AI Citation and Recommendation

First, put the data in a form that AI can find, access, and reconcile. Whether AI will use it in an answer or recommend it is a separate matter.

This infrastructure handles data publication, retrieval, and source reconciliation. It is only one of the upstream conditions that make citation possible.

05 How the Ranking Is Scored

This is not a ranking of clinical skill. It measures how much of what your website says can withstand scrutiny. The three scores are independent:

• Verifiability—whether what your website says can withstand scrutiny. Each sentence is classified as substantiated / partially supported / pending verification / contradicted
• Completeness—whether you have supplied the information AI seeks. An 11-item readiness check shows exactly what is missing
• AI Score—whether AI has actually come to read the material: how many kinds of AI have visited and how many times

The most painful result in the verifiability column is not a low score but “contradicted”: the website says one thing, and the evidence says the opposite.

The current leader scores below 70. That is the standard on this track today, which is precisely why now is the most valuable time to enter.

06 The Real Gap Appears Only After You Enter the Ranking

On the same day, with the same AI and the same measurement method, read volume differed sharply among clinics in the directory:

• 1,829 reads—the clinic onboarded to IDAEO (986 completed page fetches)
• 41.9 reads—peers without AEO, averaged across the 8 most-read clinics in the directory
• 55 reads—the strongest peer
• 0 records—a clinic not even in the directory

43 times. The control group was not a casual sample. It comprised the 8 most-read clinics in the directory. The IDAEO clinic was compared with the industry's strongest group and still recorded 43 times as many reads. This gap includes a page-count effect (986 pages on the showcase site vs a few dozen on a typical clinic website; per page, the showcase site drew about 1.9 reads per page, while the control group's per-page basis was not archived and cannot be computed precisely).

The difference did not come from clinic size, name recognition, or years in business. It came from whether the data was structured so AI could read and verify it.

To draw the line clearly: these are AI read records, not exposure, not rankings, and certainly not patient counts. But reading is a prerequisite for everything downstream: if AI has never read your information, you will not appear in its answer.

07 How AI Chooses: Three Patterns in the Logs

Pattern one: AI does not read adjectives; it recognizes facts it can verify.

Where did AI direct its effort when reading the IDAEO clinic?

• Nearly 9 out of ten—sentence-level evidence pages with verifiable claims
• About 1 in ten—professional article pages
• 1%—the branding homepage

The branding homepage you spent the most money building received only 1%. This is the other side of the earlier finding that AI will not risk using “9 out of ten”: if nine out of ten claims on your website have no supporting evidence, AI will not devote nine-tenths of its effort to them.

AI-Readable Professional Content

Write credentials, treatment explanations, limitations, frequently asked questions, and sources clearly and separately. Do not pack them into one large block of advertising copy.

“Readable” means the page is publicly available in a form that permits source checking. Logs cannot prove that AI read every word, understood it, or accepted it as true.

Pattern two: AI fetches the full set of your data, rather than selecting a single page.

Logs from the same day show requests covering every version of this clinic's material—different languages and different pages—at nearly identical frequencies, without favoring one over another.

For clinics in Taiwan, the significance is not “more international patients.” It is this: the more complete your data, the more AI can take away. If you have written only three pages, AI has only those three pages available when answering patients. Add dentist credentials, treatment explanations, and frequently asked questions, and each return visit gives it more material to retrieve.

Patients can ask the same thing in dozens of ways: “My tooth hurts so much—what should I do?” “Do I need a root canal for a cavity?” “Are implants expensive?” “My child won't open their mouth.” The more complete your data, the more of these phrasings it can address.

Foreign residents in Taiwan and medical tourism are an extra opportunity that complete data can naturally support, but the main arena will always be local patients in Taiwan. The language-version figures above are request records from a single-day demonstration case. They prove only that AI came to read; they do not mean that each request read every word of the entire site, nor do they equal conversational citations, recommendations, patient acquisition, or patient counts. “Dozens of ways” is a natural-language scenario, not a count of queries measured from these logs.

Pattern three: patients' questions are already driving AI to retrieve data.

The platform has recorded an instance in which someone asked a question in ChatGPT and AI came to read the data on the spot. It was not only ChatGPT: Microsoft Copilot's data source and other search AI systems read it during the same period.

The loop is already turning.

08 AI Was Observed Returning About Every 8 Days

The observed cadence was this: AI conducted a trial read the day after the data went live, completed full-batch indexing on day 10, and then returned about every 8 days. New content was picked up on roughly a weekly timescale—an observed cadence that can inform planning, not a fixed schedule.

09 Enter Half a Year Late and Your Rival Has Run 22 Rounds; You Are Still at Round 0

At the observed cadence, AI returns to read about every 8 days. That sounds technical, but laid out across a half-year timeline, the meaning is simple:

• Enter 1 month late—a peer already in the system has completed about 3.8 rounds; you have completed 0
• Enter 3 months late—a peer already in the system has completed about 11.3 rounds; you are still at 0
• Enter half a year late—a peer already in the system has completed 22.5 rounds; you remain at 0

This is not being a little slower. It is being absent from the scoreboard.

In this conditional scenario, those 22 rounds cannot be made up. If you enter late, AI does not go back and rerun that missed half year for you. It begins the sequence again only from the day your data goes live: “trial read the next day, full-batch indexing on day 10.” By the time you complete those 10 days, your competitor has run another round.

Here is today's starting line: of 7,002 clinics across Taiwan, only 429 have entered scoring and 339 have complete verifiability scores; 94% remain blank. That means two things. Enter now, and the early-mover advantage comes mainly from the time difference. But blank entries will not remain blank forever.

And you will receive no notification. No missed call, no canceled appointment, no negative review—nothing will happen. You will not know that a question was asked, and you will not know whether you appeared in the answer.

In the AI era, being passed over shows up in no report — which is why you have to check for yourself.

10 Three Things You Can Check Right Now

1. Look up your clinic on the ranking. All 7,002 clinics are listed. Search your name and see whether you have a score, how many claims are substantiated, and how many are contradicted—all on one page. You do not need to leave any information or ask anyone.
2. Ask ChatGPT, “Do you know ○○ Dental Clinic?” This tests whether AI has information about you, not whether it will recommend you. If it cannot answer or cobbles together a patchwork response, the material available to AI is inadequate and too weak for it to use.
3. Open your website, choose 10 statements, and ask yourself: which one has evidence behind it? Can a dentist's credentials be traced to a certification number? Are treatment outcomes supported by literature or cases? Can reviews be traced to a source? This is how to test that “9 out of ten” finding yourself.

11 Three Conclusions and Five Layers of Assets

① This is a race against time. 94% of clinics still have no score, and the leader remains below 70. The IDAEO clinic was read 43 times as often as the average for the top eight peers in the directory. That gap was not created by slogans. It accumulated one read round at a time.

② This is an asset, not an advertisement. An ad disappears when spending stops. Well-structured, verifiable data remains in place, and AI has been observed returning to read it about every 8 days as the data continues to be updated.

③ The barrier is not budget. It is the willingness to make every claim verifiable. 9 out of ten claims fail the test, not because those clinics are worse, but because no one told them that a website must be written this way in the AI era.

In plain language, the five layers are:

• Foundation｜A verifiable digital identity—names, people, and public information can be reconciled
• First layer｜AI-readable professional content—content has a clear structure and retains traceable sources
• Second layer｜Infrastructure for AI citation and recommendation—provides the prerequisite data in an accessible, reconcilable, and updatable form
• Third layer｜Brand-authority building in the AI world—replaces self-assertion with continuously verifiable professional data
• Roof｜A long-term digital asset—data and source records that a clinic can maintain, correct, and reuse

Boundary: these are layers of work, not an outcome funnel. Reaching any layer does not mean that an AI citation, recommendation, or patient result has occurred.

12 Data Definitions (All Figures Are Traceable and Auditable)

Census and ranking: observation window from 2026-07-18 to 2026-07-26 (the ranking uses a 30-day rolling window, updated daily; this article's snapshot covers the 2026-07-18 to 07-26 segment). The nationwide National Health Insurance-contracted dental registry contains 7,002 clinics; 429 clinics (6%) have entered the scoring pipeline, 339 (5%) have completed verifiability scoring, and 94% have no score on the ranking. A total of 7,431 clinic pages have been built. The website census extracted 18,731 claims, of which 1,913 (10.2%, described in the article as “about 1 in ten”) can be supported by dental literature; verification cited 745 PubMed papers. During the observation window, AI made 12,943 observed visits covering 3,937 clinic pages. Another 117 clinics already had verifiability scores but had not yet received an AI visit. The ranking is continuously updated; consult the ranking page for the latest figures. Sentence-level verdicts fall into four categories (provable / partially supported / pending review / contradicted); this release archived only the count of the "provable" category (1,913 claims), and the distribution of the other three categories was not archived with the statistics and cannot be reconstructed. The precise meaning of "about 1 in ten claims has evidence" is: among claims that completed literature matching, about 1 in ten met the provable standard. The 18,731 claims were extracted from clinics whose websites could be fully crawled (between 339 and 543 clinics; the exact number of sampled clinics was not archived with the statistics); the 6,459 clinics without websites are outside the extraction scope.

Scoring methodology: verifiability = (substantiated + 0.5×partially supported) ÷ checkable sentences; no score is issued when the sample is insufficient. Completeness = the weighted total of the 11 readiness items, totaling 1.00. AI Score = AI crawler types (breadth, capped at 40) + logarithmically compressed visit count (depth, capped at 60). Since 2026-07-22, rank order has been determined by AI Score. The leader is not necessarily the most verifiable: verifiability and completeness are displayed alongside the ranking but do not determine rank.

Single-day deep-dive case: the IDAEO clinic recorded 1,829 reads across 986 pages in one day; page-type distribution was evidence pages 1,610 reads 88% / article pages 204 reads 11% / homepage 15 reads 1%; language-version counts were Traditional Chinese 474 / Japanese 464 / English 449 / Simplified Chinese 442; and the daily full-batch index comprised 1,939 pages, with return visits about every 8 days. These figures came from a direct query of Cloudflare GraphQL Analytics API edge-request logs for the idaeo.ai domain, executed on 2026-07-28. The measurement period was 2026-07-21 to 07-28, with a one-day peak on 2026-07-26; the sample was an illustrative clinic from the platform's first onboarded cohort. This series uses a different definition from the ranking's AI Score: the former counts requests across all language routes, while the latter scores visits to Embassy pages within the observation window. The two cannot be compared directly. The control group comprised the 8 clinics with the highest AI read volume in the national dental AI directory (average 41.9 reads, maximum 55 reads), not a random sample.

AI identity determination: Cloudflare's official verified-bot determination is used, based on reverse IP lookup + behavioral verification, rather than a self-declared User-Agent alone. “AI came to read” in this article means requests from this class of verified bots, led primarily by OpenAI GPTBot. Other engines during the same period included Microsoft bingbot (a Copilot data source) 448 times, YandexBot 307 times, OAI-SearchBot 11 times, and user-triggered ChatGPT-User 2 times.

Cadence note: “returns every 8 days” is the cadence observed during the period above. This article does not extrapolate it into a fixed cycle or guarantee a number of visits across months. Any result a reader calculates independently is not a claim made by this article.

Honesty statement: this article reports observed results for “AI engines indexing and reading clinic data” and a census of the literature verifiability of website claims. Indexing is a prerequisite for AI citation, but indexing itself does not mean the clinic has been cited in an AI conversation, much less that it represents a patient count. Independent tracking for the latter has been built and data is accumulating. Low verifiability does not mean poor healthcare quality; it refers only to how verifiable the website's language is. No estimated or forecast values are used anywhere in this article. The figure 6,459 is the simple subtraction 7,002−543.

Disclosure of interest: the showcase clinic behind the 1,829 reads is an onboarded IDAEO client; IDAEO provides related services. Readers weighing this article's conclusions should be aware of this interest.