What Does AI Say About a Nonexistent Clinic? GPT Denies It, Gemini Recommends Alternatives, and Perplexity Invents Reviews in 15% of Responses
Published
This article was first published in Korean by Intrix Lab on September 30, 2026.
AI models respond differently when asked about businesses that do not exist. As of 2026-09-14, we asked three models reputation questions and questions about location, prices, and hours for 9 fictional businesses in Busan, including medical clinics, law firms, rental companies, dental clinics, Korean medicine clinics, tax accountants, and private academies. We repeated each question 3 times per model, for 162 responses. GPT answered that it could not verify the business in 74% of responses. Gemini suggested other real businesses in the same area instead of denying the premise in 80%. Perplexity denied the premise in 56%, but presented reviews, addresses, and phone numbers for nonexistent businesses as facts in 15%. GPT and Gemini produced 0 fabrications. A paper featured in GEO Watch proposed measuring AI refusal policies by including questions outside the knowledge base. We replicated that approach with Korean-language questions about local businesses.
Why ask about nonexistent businesses?
An AI mention of our business alone does not tell us whether that mention has supporting evidence. To assess what a mention means, we need to know what AI does when evidence is absent. When asked about a name that does not exist, there is one correct answer: "I cannot verify it." How a model departs from that answer reveals its tendencies.
How did we measure this?
We created fictional business names across different industries. For each name, we asked reputation questions and questions about location, prices, and business hours 3 times each across three models: GPT-5.6 Terra, Gemini 3.8 Flash, and Perplexity Sonar. We classified responses into four categories: denial (saying the business could not be verified), substitution (suggesting other real businesses or institutions), fabrication (presenting addresses, hours, prices, or reviews as facts), and mixed (providing some specific facts while saying the business could not be verified). We excluded one of the ten names initially created because a similar real brand existed.
What were the results?
| Model (54 responses each) | Denial | Substitution | Fabrication | Mixed | Responses with citations |
|---|---|---|---|---|---|
| GPT-5.6 Terra | 74% | 26% | 0 | 0 | 41/54 |
| Gemini 3.8 Flash | 20% | 80% | 0 | 0 | 49/54 |
| Perplexity Sonar | 56% | 30% | 9% | 6% | 0/54 |
Results differed by question type. For questions about location, prices, and hours, Gemini used substitution in 93% of responses. GPT used denial in 67%, while Perplexity used substitution in 52% and denial in 44%. For reputation questions, Perplexity's fabrication and mixed responses rose to 26%. Examples included assigning the address "1133 Jungang-daero, Yeonje-gu" and office hours for each day of the week to a nonexistent Korean medicine clinic, or claiming that a nonexistent dermatology clinic had "about 19 verified-visit reviews."
What does Gemini's "substitution" mean?
When given a nonexistent name, Gemini recommended real businesses in the same area and industry instead of saying it could not verify the name. That may be helpful to the user, but it has a different meaning for GEO. When AI cannot find our business, the result is not "no mention" but "a competitor recommendation." Failing at the discovery stage does not leave an empty spot. Someone else fills it.
Why did Perplexity's answers have no citations?
In this collection, Perplexity Sonar included no citations in its API responses across all 162 cases. We therefore had no way to check the source of statements such as "the reviews found repeatedly describe the staff as friendly." Because those statements concerned businesses that did not exist, we classified them as fabrications. This experiment directly shows that mentions without citations leave no way to check the supporting evidence.
What do we do differently as a result?
First, when counting AI mentions, we count mentions with citations separately from those without them. Second, for questions where our business does not appear, we also check who gets recommended instead. Those are the actual competitors.
Limitations
This study covers 9 fictional names and 162 responses. It does not represent the refusal policies of the models as a whole. Classification was automated, so there are borderline cases between denial and substitution. We do not use these results as evidence that a particular model is superior or that citations are guaranteed.
Frequently asked questions (FAQ)
Q. AI described reviews of our clinic. Can we trust them?
If citations are included, you can open those pages and check. Without citations, even nonexistent businesses can be assigned reviews, as in this experiment, so we do not use those statements as evidence.
Q. If AI cannot find our business, does it simply say it does not know?
It depends on the model. GPT generally says it does not know, while Gemini recommends other businesses. Failure at the discovery stage therefore means exposure for competitors.
Q. Can we run this experiment with our brand?
Real brands require a different design. This method measures behavior "when there is no evidence." For a real brand, the question is which pages serve as evidence when evidence exists.