Why Istanbul dental clinic ratings pile up at the top — and what to look at instead

The finding: the score is a ceiling effect, not a quality signal

Across six Istanbul dental clinics we collected 2,115 publicly visible reviews and counted the stars ourselves, one by one. Five stars appeared on 93.85 percent of the reviews in the whole set.

At that level of saturation the average stops carrying information. When 93.85 percent of the ratings sit at the maximum, the arithmetic mean across the corpus as a whole is penned into a narrow band regardless of what the remaining reviews say. In our set that band runs from about 4.75 to 4.94 - and the low end of that range assumes every single non-five-star review is a one-star review, which is not the case anywhere we looked. The band is a property of the corpus taken together, not a rule each clinic obeys; applied clinic by clinic it fails three times out of six, and not by equal margins. One clinic sits far below the band at 4.20. The two above it clear the line by much smaller margins: 4.97 a little further out, 4.95 only barely. Both of those gaps are thin, and it is worth saying why they still count: we recomputed every average from the individual reviews, so the rounding here is in how the figures are printed, not in how they were measured.

We recalculated each clinic's average from the individual reviews and set it beside the figure the platform displays. Across all six the gap never exceeds 0.052; at the platform's one-decimal display, five of the six land on the same figure and the sixth does not — ours drops a notch there, 4.9 against a displayed 5, and that clinic is exactly where the largest gap sits. We cannot say whether that is an error or two snapshots taken at different moments, because we do not know when the platform computed the figure it shows. What these figures describe is a distribution in which the high scores stop separating anything: five of the six clinics sit between 4.81 and 4.97, a spread of 0.16, while the sixth sits alone at 4.20. Moving a high average takes more than a bad week - if every other review were a five, one one-star in twenty would bring it down to 4.8, and where the middle of the scale gets used at all, fewer will do it. So the gaps between the clinics at the top carry far less information than they appear to.

Why the distribution looks like this: timing

Rather than counting them, we read 50 one-star reviews by hand — every one-star review in the corpus with enough text to describe what happened. Two patterns came up repeatedly.

At least seven of the fifty described a delayed failure: a bridge that came down eighteen months later, teeth that fractured one after another over a period of years, an implant with bone loss around it discovered well after the work was finished. At least seven described a communication breakdown once the patient had flown home - contact with the clinic broke down after the return; we recorded the pattern, not what caused it.

Both patterns we read share a property that matters here, and it is one we could not put a date on. We cannot date the reviews we collected: 43.9 per cent of the dates came back as a year only, so we do not know when a satisfied review is written, and we have no measure of when one normally is. The failures we read about became visible months or years later, by which time the patient was no longer in the review-writing moment and often no longer in contact with the clinic at all. Where in the life of a treatment a star average is taken is therefore something our data cannot tell us.

What we looked for and did not find

We went looking for fabrication signatures, and we did not find them. In our set the median five-star review runs 187 characters; we did not compare that against any external benchmark for template text. Templated phrasing accounted for 6.3 percent of the text. A reviewer-farm pattern — one account reviewing more than one of the six clinics — appeared in 14 of 2,085 reviewer profiles.

We are stating this plainly because it cuts against the assumption people arrive with: on this evidence the claim that these clinics are buying fake reviews is not supported, and we do not make it. What we found is not a problem of invention. What it is instead we cannot say from this data, because the one axis that would answer it — when each review was written — is the axis we could not measure.

What we could not measure

We wanted to test whether reviews arrive in bursts - clusters in time that would suggest coordinated posting. We could not run that test. Our collection tool capped out at roughly 490 reviews per clinic, and 43.9 percent of the dates it returned were year-only, with no month or day. That resolution is not enough to detect clustering.

So we are not saying there is no burst pattern; we are saying our design could not measure it. Anyone repeating this work with full timestamps would be answering a question we left open.

How the sample was chosen, and why not the biggest names

We chose the six clinics as a stratified sample, and we deliberately turned down the obvious alternative of taking the five clinics with the most reviews. We did not measure what any clinic spends on anything, so we make no claim about how marketing-intensive that top stratum is, and our own data gives no support to the idea that reviews are bought. Stratifying across the range yields a less flattering but more representative picture.

We are not publishing the raw data; in this analysis we do not name the clinics, we do not quote reviews verbatim, and no reviewer's name appears anywhere in this work. Leaving the clinics unnamed here is not a promise that they cannot be identified: the reviews are public, matching them against other records could point to a particular clinic, and we do not guarantee that it cannot be done.

What is checkable when the score is not

We could not verify a rating against any source we hold. The things listed below can at least be asked for, though we have not confirmed against Turkish law or any register that they must be provided.

Rather than accepting it from a website, a dentist's registration is the kind of thing a public register would carry, though we have not verified which Turkish register holds it or what it shows. A written treatment plan, an itemised cost breakdown, and the specification of the materials used - brand, model, batch - are documents to ask for, though we have not verified what a clinic in Turkey is obliged to hand over. What happens if something fails after you have flown home is a question to settle in writing before you travel, and we have not established what the law provides when nothing is written. Where a complaint would be filed, and to which authority, is something we have not traced to a published procedure.

One more thing turned out to be legible in the data. Of the 2,115 reviews we collected, 1,039 carried a translation, and for 1,034 of those we were able to read Google's own declared source language, which Google prints inside the review itself. Every count in this paragraph is recounted from those collected review files rather than carried over from a table, and the five translated reviews that declared no source language are counted nowhere below. Across those 1,034, Arabic leads at 455 and French follows at 390, ahead of Russian at 51 and German at 42, and French is present at all six clinics without exception. These are counts of reviews, not of patients: anyone who was treated and wrote nothing is counted nowhere, so what we are describing is a share of what was written, not a measure of who was treated. Reviews written in French are also not spread evenly across the six clinics, so the share of reviews written in French observed here holds only for this set of reviews: it does not describe the mix of review languages at any individual clinic or the languages of its patients.

What this page does not have yet

A named editor has now been appointed, and the foot of this page names the dentist who wrote it and the medical editor who checked it. The sign-off behind that name is written and dated, and an approval line runs page by page as well: a page is read by the medical editor, and a dated signature record is then either written against it or withheld. This page carries no clinical proposition, so it holds no such record; the 41 pages that do carry one have been signed. What a reader can ask to see for this page is still the collective sign-off rather than a dated review of its own, and on the pages that hold a record the date sits in the record rather than printed on the page. Our corrections log is published and only grows. Our evidence audit page explains how claims are counted and presents the recorded results. We are stating these gaps here rather than waiting to be asked about them.