How AI search describes dental clinics: a reproducible India test of local treatment queries, September 2026

The answer has to be two numbers, not one. First, X of Y local dental query runs in which a clinic's own domain was cited, with the location, the query set and the date range stated. Second, how often the description the assistant gave of a named clinic was correct across five checkable fields. We have not run this test, so no fraction for dental clinics appears on this page. The design is published in full: ten treatments crossed with six cities, which is 60 queries, on three assistants with three repetitions, which is 540 observations per wave. The nearest real fractions we hold are from other sectors entirely, and they are named as such: 0 of 6 on Perplexity in September 2026 on our own domain, and top source on 1 of 18 on ChatGPT on 18 August 2026 for a legal technology company.

The answer, first

Four things are specific to this sector and change the whole test.

Assistants are cautious about health, and that is a result rather than an obstacle. Many dental queries come back as general information with a recommendation to see a dentist, and no clinic named. Count that outcome in its own column. A category where no clinic is ever named is a finding you can act on, and it is completely invisible if you only count your own citations.

Being described wrongly is worse than not being described. A clinic named with the wrong area, the wrong treatments or a price it does not charge will lose the patient at the phone call. So the second number on this page exists, and it is scored across five fields.

Advertising by dentists is regulated. In India the Dentists Act 1948 and the code of ethics regulations made under it restrict advertising and canvassing, and state dental councils apply them. Read the current text from the Dental Council or the relevant state council, record the date you read it, and take borderline questions there rather than to a marketing agency.

Patient material needs consent, in law. Before and after photographs, case descriptions and reviews involve personal and health data, which the Digital Personal Data Protection Act 2023 governs. Publish nothing about a patient without written consent, and keep the consent.

How it was measured

The grid: ten treatments crossed with six cities, which is 60 queries. Both lists are fixed in a dated file before any run.

• Ten treatments, in the words patients use: root canal treatment, dental implant, braces or clear aligners, teeth whitening, wisdom tooth removal, crown or cap, dentures, scaling and cleaning, a child's filling, and full mouth rehabilitation.
• Six cities: Mumbai, Delhi, Bengaluru, Pune, Hyderabad, Ahmedabad. Use your own market instead, keep the count, and state the substitution.

Three query shapes, one per purpose, and each treatment and city pair gets the shape that matches what patients actually type.

• Find: "Where can I get [treatment] in [city]."
• Cost: "How much does [treatment] cost in [city]."
• Decide: "Do I need [treatment], and how do I choose a clinic in [city]."

The blind rule. The clinic's name never appears in any prompt, wrapper, system instruction or earlier turn. Our record shows why plainly: on 27 July 2026, in the clawlaw.in programme, the same 78 questions run twice in one day gave a brand first place on almost every question when the wrapper named it, and second by breadth with complete absence from the most valuable questions when nothing was named. The branded run was discarded.

The arithmetic. 60 queries times 3 assistants times 3 repetitions is 540 observations per wave. Publish that so any count on the page can be rebuilt.

Conditions recorded per observation. Fresh session, memory and personalisation off. The location the session reports and how it was set. The interface used. Whether a live search was visible and how many sources were listed. The language of the query, with a separate grid if you also run Hindi or a regional language. The UTC timestamp.

Six outcome scores.

• Own domain cited. A URL on a clinic's own website appears in the sources. This is the numerator of the first fraction.
• Any clinic named. Whether a specific clinic is named at all.
• Directory or map source cited. Whether a listing site or map data supplied the answer instead of a clinic.
• Health guardrail. Whether the answer declined to recommend and advised seeing a dentist, with the wording recorded.
• Price stated. Whether a cost or range was given, and whether a source was attached to it.
• Official or professional source cited. Whether a council, hospital or professional body was used.

The five field description accuracy check, scored only where a clinic was named. For each named clinic, mark each field correct, wrong or absent, and copy out anything wrong.

• The treatments the answer says the clinic offers.
• The area or locality it places the clinic in.
• Any price or range it attributes to the clinic.
• The dentist's qualification or specialisation as stated.
• The opening hours or availability as stated.

Report accuracy as a fraction of the fields that were checkable, not of five, because an answer that says nothing about hours has not made a mistake about them.

Recording the search claim. Log the visible sources, then ask the assistant afterwards how many queries it searched for, and keep that as a separate claim field. In three separate batches of our September 2026 audit the assistant withdrew its own earlier statement, saying it could not honestly substantiate the claim that it had run a live search for each question.

What the numbers were

Dental clinic runs completed: none, as of 29 September 2026. The 60 query grid has not been run, so there is no X of Y, no description accuracy count and no date range for this sector. That is the plain state of it.

The nearest fractions we hold, with their sectors stated.

• 0 of 6, our own domain in software. Six questions about our own category with no brand named, on Perplexity in September 2026, after which the assistant reported that it had not cited or recommended aiknowsus.com in any of the six answers.
• 1 of 18, legal technology. Eighteen blind commercial questions on ChatGPT on 18 August 2026, with clawlaw.in the top source on exactly one.
• 161 recordings of not being cited, across 24 batches and 72 conversations on our own domain in September 2026.

Two dated results that bear directly on the description accuracy half of this test. On 6 August 2026, in the clawlaw.in programme, a pricing page rendered its prices only after scripts had run, so a crawler received no prices and the prices the assistants quoted had come from an app store listing instead of the company's own site. Claude's run recorded it, and the same day ChatGPT found that the site and the listing carried different plan names and different prices for the same product and said so in its answer. For a clinic, the equivalent failure is a treatment price or a list of services that exists only in a script or only on a third party listing. The assistant will describe you from whatever it can read, and if that is a directory entry from two years ago, that is the description your patients get.

What this cannot tell you

• It cannot tell you what a patient in your area sees. Real sessions carry location and history that this test switches off to stay comparable.
• It cannot separate an assistant's answer from map data. For "near me" questions much of the answer comes from listings, and a listing citation is not a citation of your site.
• It cannot tell you anything clinical, and nothing in the query set should be read as advice about treatment.
• It cannot tell you whether a citation produced an appointment. That needs call and enquiry tracking at the clinic, which is a separate measurement.
• It cannot settle what you may publish. That is decided by the dental council rules current on the day you publish, not by this page.
• It cannot promise a position. Nobody can guarantee a place in an assistant's answer, and a clinic should treat any such promise as a warning. Disclosure: we are the vendor of AI Knows Us, which sells this kind of measurement.

Sources and change log

Sources. For what a dentist may publish: the Dentists Act 1948, the code of ethics regulations made under it, and your state dental council's current directions, read at the source with your date recorded. For patient data: the Digital Personal Data Protection Act 2023. For the dated results: the clawlaw.in programme of July to September 2026, recorded in Tier_1/GEO_BASELINE_RESULTS_2026-07-27.md and Tier_1/GEO_GAP_ANALYSIS_2026-08-18.md, and the aiknowsus.com audit of September 2026 captured to geo-audits/aiknowsus-com/. The 6 August 2026 pricing results concern public pages and listings and can be checked from outside.

Change log. 29 September 2026: first published with the ten treatment by six city grid, the three query shapes, the 540 observation arithmetic, the six outcome scores, the five field accuracy check, and the statement that no dental run has been completed. When wave one runs, this section will carry the X of Y, the accuracy counts, the cities, the date range and the raw rows, with this entry left above them.

Common questions

Assistants will not recommend a clinic. Is there anything to measure?

Yes, and the refusal is the measurement. Count how often the answer declines to name anybody, record the wording, and then count which sources it used instead. That tells you whether to compete for the naming or to become the source the cautious answer quotes for the procedure and the cost.

Should we publish treatment prices?

Publish ranges with the conditions attached: what is included, what changes the figure, whether an implant price includes the crown, and the date. Check your council's current position first. A range with conditions is quotable, and a page saying only that cost depends on the case gives an assistant nothing and leaves your price to a directory.

What is the single most useful page for a clinic?

A treatment page per treatment, in the patient's words, that states what the procedure involves, how many visits it takes, what it costs as a range with conditions, who performs it and their qualification, and what it does not suit. That is also exactly the page a cautious assistant can safely quote.

Are patient reviews and before and after photographs worth publishing?

Treat them as two separate questions. First, whether the council's rules permit it, which you must check. Second, consent: photographs and case descriptions are personal health data, so publish nothing without written consent under the 2023 data protection law and keep that consent on file.

Should we run the test in Hindi or a regional language?

If your patients ask in one, yes, and keep it as a separate grid. The sources an assistant reaches for can differ by language, and averaging them hides the difference you most want to see.

How long does one wave take?

Five hundred and forty observations scored by hand is about four days for one person, and the accuracy check adds time because each named clinic has to be verified. A reduced grid of five treatments and two cities is 10 queries and 90 observations, which is a day, and your heading should say five treatments and two cities.

What to do first

Take your three highest value treatments and your own city, run the three query shapes on two assistants with three repetitions, which is 54 observations and about half a day, and record whether any clinic was named and what was said about it. Then check the five accuracy fields for every clinic that was named, including yours, and fix whatever is wrong about you on your own site first, because that is the description the next answer will be built from.