Did AI Search Before Recommending Businesses? Reanalyzing 8,516 Answers
Published
This article was first published in Korean by Intrix Lab on October 11, 2026.
GPT did not name businesses without searching, while Gemini generated six in ten answers that named businesses without searching. These are the results, as of 2026-10-08, of rereading 8,516 AI answers Intrix collected from August 20 through October 7. Across three GPT models, 20% of all responses were generated without searching, but 0 of those responses named businesses. Across two Gemini models, 28% of all responses named businesses without searching.
What did we count?
Intrix's AI answer studies record two things for each response: whether a web search was used to generate it, and which businesses or institutions were named in its text. Combining these two records reveals responses that "named businesses but did not search."
These figures cannot be obtained simply by looking at the AI interface. The interface shows answers and sources, but not whether a search took place. We reread the existing responses from 16 completed studies without making new AI requests.
How much did the models differ?
| Model | Responses | Answered without searching | Named businesses without searching | No search among responses naming businesses |
|---|---|---|---|---|
| GPT Luna | 2,088 | 10.6% | 0 cases | 0 cases |
| GPT Terra | 1,278 | 9.2% | 0 cases | 0 cases |
| GPT Sol | 1,540 | 42.1% | 0 cases | 0 cases |
| Gemini 3.7 | 2,196 | 74.1% | 34.0% | 60.9% |
| Gemini 3.8 | 1,186 | 86.9% | 17.9% | 58.7% |
GPT Sol answered four in ten responses without searching, but those responses provided only general explanations and named no businesses. This means a business must appear in search results for GPT to name it. For Gemini, by contrast, 60% of responses naming businesses were unrelated to search, and responses without search actually contained more names. Gemini 3.7 included an average of 8.5 names per response without search, compared with 7.1 in responses with search.
Which questions led AI to skip searching?
The difference between the two AI systems becomes clearer by question type. GPT almost always searched for recommendation questions asking "which is good" and questions specifying a location. The proportions answered without searching were 0.7% and 0.9%, respectively. For questions asking about issues or describing situations, 20–26% were answered with general explanations without searching.
For Gemini, recommendation questions had the highest proportion of responses naming businesses without searching, at 50.3%. Situation questions followed at 41.3%, then issue questions at 25.5%, cost questions at 22.0%, and location questions at 20.8%. For questions checking whether a business was the right one, the figure was only 4.7%. When customers ask for recommendations, half the names Gemini supplies do not come from search results retrieved at that moment.
What does this difference change?
The two AI systems recommend businesses through different routes. For GPT, the entire process is appearing in search results and then being selected from them. Gemini also supplies names the model already knows, separately from that route. Among responses naming the businesses under study, 51 of Gemini's 75 responses were generated without searching, while all 7 GPT responses involved a search.
That is why "AI visibility" should not be combined into a single number. GPT's figures can be changed by improving the site and its search results. Names Gemini supplies without searching do not change immediately when the site is updated. Intrix reports the two figures separately and checks whether names supplied without searching recur in the next round. We first examine which route surfaces your business through a free website audit and a query study.
How far can these results be applied?
The evidence used to determine whether a search took place differs by model. Comparisons between models should therefore be treated as directional only. The studies cover six industry groups: legal services, branding, design for business support programs, children's clothing, GEO agencies, and tax accountants. All questions were in Korean.
Even within branding, the proportion of responses naming businesses without searching was 58.1% for Gemini 3.7 and 15.0% for Gemini 3.8. The two studies were conducted a month apart, so the effects of model generation and timing cannot be separated. Names of the target businesses were matched conservatively, which may have resulted in undercounting.
Frequently asked questions (FAQ)
Q. Does this mean improving a site has no effect on Gemini?
No. The other four in ten Gemini responses naming businesses involved a search, and that portion depends on the site and search results. The point is that the portion generated without searching does not change immediately with site updates.
Q. What were the 20% of GPT answers generated without searching?
They explained procedures or criteria without naming businesses. Such answers were common for issue and situation questions, and rare for recommendation questions.
Q. What about Perplexity?
Among 180 responses in a legal services study, 3.9% were generated without searching. It searched almost every time.