Does AI Judge a Company by Its Name? A 240-Call Replication with Korean Dermatology Clinics

This article was first published in Korean by Intrix Lab on August 31, 2026.

Peec AI, a German GEO analytics tool, reported on its blog that AI assistants judge companies by their names and transfer words from those names into company descriptions. We attempted to replicate this finding on the same day using a Korean-language dermatology clinic sample measured on 2026-08-31. The results did not support the original claim, but produced another finding that deserved more attention.

What did Peec claim?

Peec's finding was that words in a company's name carry over into AI-generated descriptions of the company. It measured whether AI would describe a company as upscale when its name contained a word with an upscale meaning.

If this finding holds, the business name itself becomes a GEO variable. The less information AI has about a company, the more likely it may be to use the name as material for its description.

How did we test this with 20 Korean dermatology clinics?

We fixed the hypotheses, sample, and assessment rules before seeing the results. On 2026-08-31, the day we found the external claim, we completed preregistration before collecting data.

The sample consisted of 20 dermatology clinics selected nationwide, divided into those with abundant information on the web and those with little information.

We queried two models under conditions with web search enabled and disabled, collecting a total of 240 answers.

For each answer, we checked whether words from the name carried over into the clinic description, whether it included statements asserted as verifiable facts, and whether it said "I don't know." Below, we refer to these as answers with name-word transfer into descriptions, factual claims, and answers primarily saying "I don't know," respectively.

Results: Name-word transfer was not replicated

The differences between strata were small. With search disabled, model B showed 13.3% (4/30) for the information-poor group and 10.0% (3/30) for the information-rich group. Model A showed 0% in both groups.

The transfer signal was concentrated entirely in Cheongdam. All 7 answers classified as showing transfer involved clinics with "Cheongdam" in their names. There were 0 instances of transfer for ordinary words such as miso (smile), goun (beautiful), bit (light), and malgeun (clear).

Cheongdam alone cannot establish a name effect. Cheongdam is both a place name and a term with upscale connotations, so this design cannot separate a name effect from regional stereotypes.

With search enabled, the rate of name-word transfer into descriptions fell to 0%. This was consistent with the direction of the second hypothesis: that access to web information reduces the influence of words in the name.

An unexpected finding: No answer said it did not know

The result that deserved more attention was that neither model stayed silent when information was unavailable. This was an exploratory finding outside the preregistration that drew our attention after we saw the results.

Answers primarily saying "I don't know" accounted for 0/240. All 240 answers included factual claims.

Even without information available on the web to check, AI generated specific details. It confidently generated lists of medical services even for clinics with 0 pages indexed by Google.

However, classifying an answer as containing factual claims does not mean those claims were verified as true. This design cannot measure the accuracy of individual claims, so we do not report an error rate.

What can we conclude for Korea?

Peec's name-word transfer was not replicated in this Korean-language sample. There is not yet evidence that names in Korean business names change AI descriptions.

However, terms such as Cheongdam, which are both place names and words with connotations, showed an exception signal. They merit separate testing with a design that distinguishes name effects from regional stereotypes.

The more practical conclusion is that Korean clinics should be concerned less about their names than about the content AI invents when information is missing. AI did not stay silent even when information was unavailable on the web.

How far can these results be trusted?

These results should be treated as a pilot rather than a full study. The sample was 20 dermatology clinics, with n of 10 per stratum, so we cannot attach confidence intervals to the differences between strata.

Other constraints also affect interpretation. Information from clinics with the same name in other regions may have been mixed in for up to 6/20 clinics, so we invalidated the region-related metrics.

This observational study cannot establish causation, and its results cannot be generalized beyond a single industry and a single country. Intrix does not guarantee rankings, citations, or visibility.

What should clinics check first?

First, check how AI describes your clinic. Start with claims that can be compared against official materials, such as medical services, addresses, and clinical staff.

Assess both the answer's overall stance and its specific claims. Intrix's free SEO and GEO audit checks site indexing and AI access conditions through actual requests. The accuracy of answer content needs to be measured separately using a sample of questions and comparisons against official materials.

Frequently asked questions (FAQ)

Q. Was Peec's name effect confirmed in Korea?
No. In this sample of 20 dermatology clinics, there were 0 instances of transfer for ordinary words, and we could not separate the Cheongdam signal from regional stereotypes.

Q. Does this mean all 240 answers were wrong?
No. All 240 answers contained factual claims, but we did not determine whether each claim was correct or incorrect. This must not be read as an error rate.

Q. Can I apply these results to my clinic?
They cannot be generalized directly. This was a pilot in a single industry and a single country, so AI answers for each clinic need to be checked separately against official materials.