Original research · dental practices in Odesa, Ukraine
What four AI assistants answer when a local customer asks for a recommendation
Two measurements on one local market, 239 answers in all. Who gets named, how little the assistants agree, what the answers are built from, and where visibility collapses. Every figure below was recomputed from the stored answers on 1 October 2026 and can be checked against them.
-
63
different businesses named across 24 answers to the same three questions
-
4
of them named by all four assistants. The rest depend on which one your customer opens
-
11.2
pages cited behind an average answer, from 57 different domains
-
0%
share for one practice on urgent questions, against 92% on general ones
Method
Two measurements, one market
One goes wide across assistants, the other deep into one. Both were run on dental practices in Odesa, Ukraine, with questions written the way customers ask them, in Russian and Ukrainian.
| Breadth | Depth | |
|---|---|---|
| Date | 8 September 2026 | 7 September 2026 |
| Assistants | ChatGPT, Gemini, Claude, Perplexity | Gemini |
| Questions | 3 | 43 |
| Repeats per question | 2 | 5 |
| Answers recorded | 24 | 215 |
| Business mentions extracted | 183 | 1836 |
Every answer was stored verbatim with its citations. Nothing was asked once: a single answer is a coin toss, and the figures below are shares across repeats.
Finding 1
Four assistants, almost no agreement
The same three questions, asked twice on each of four assistants, produced 63 different business names in 24 answers. 4 of those businesses were named by all four assistants. 46 were named by exactly one.
A single blended "AI visibility score" would average four lists that barely overlap. For a business in this market, being visible on Gemini says close to nothing about being visible on Claude — which is why every figure we report is per assistant.
Finding 2
An answer is assembled from about eleven pages
268 citations across 24 answers — 11.2 per answer, from 57 domains. The clinics' own sites account for 47% of them; directories, rating sites, booking platforms and local media for the other 53%.
| Domain | Citations | What it is |
|---|---|---|
| top20.ua | 33 | City directory |
| triodent.od.ua | 30 | A clinic's own site |
| doc.ua | 18 | Medical directory |
| likarni.com | 15 | Medical directory |
| phildent.od.ua | 13 | A clinic's own site |
| kichukdentalclinic.com.ua | 12 | A clinic's own site |
| uba.top | 8 | Rating site |
| stomatolohii.com | 8 | Dental directory |
| odesa-day.com | 6 | Local media |
| dymovdental.com | 6 | A clinic's own site |
The six most-cited domains supplied 45% of all citations. The single most-cited source was a city directory; second was a clinic's own website.
What this means in practice
Your own pages are not a minority player here — they are the largest single category of source, and the one you write yourself. The practical order follows from that: state the facts a buyer asks about on your own pages first, then make the directories and rating sites that get read alongside them agree. A contradiction between four descriptions of the same clinic is a reason for an assistant to name a fifth.
Finding 3
Visibility collapses as the question gets closer to a booking
Taking one practice from the set and reading its share by question type gives the shape we now expect to find: strong where nothing is at stake, absent where money or urgency is mentioned.
Averaged into one number this practice sits at 57% and looks healthy. The number that matters commercially — can somebody who needs a dentist today find it — is zero.
By question type
The same practice, six kinds of question
| What the customer asked | Times asked | Times named | Share |
|---|---|---|---|
| General recommendation | 25 | 23 | 92% |
| A specific service | 75 | 53 | 71% |
| Reviews and trust | 35 | 19 | 54% |
| Price | 25 | 7 | 28% |
| Near me | 25 | 5 | 20% |
| Needed right now | 20 | 0 | 0% |
Limitations
What this study does not show
It is one market, two days and 239 answers. Treat it as a worked example of a method, not as a law of nature.
-
One city, one niche
Dental practices in a single city. Other categories — regulated, national, B2B — behave differently, and we say so rather than generalising from one market.
-
Counting names is a judgement
63 is the count when a name we could not match to a known business is counted as itself. Merge strictly and the same answers give 33. We publish both, because the method changes the headline by a factor of two.
-
Three questions across assistants
The breadth measurement is deliberately small: the same three questions, twice each. It is enough to show disagreement, not enough to rank assistants against each other.
-
Answers move
These are two days in September 2026. Assistants change their retrieval and their sources, which is the reason measurement is repeated rather than done once.
A correction worth making in public
This site first published 65 businesses, three named by all four and 48 named by one. Those were the figures from the first pass over the same run. Duplicate names have since been merged, and recomputing on 1 October 2026 gives 63, 4 and 46. Every page has been updated to the new figures, and this note stays here rather than the change happening quietly — a number that moves without explanation is exactly what we tell clients not to accept.
Run this on your own market
The method is published in full and copyable. If you would rather we did it, the free check runs a short version on one assistant and emails you the answers, usually within thirty minutes.