AI Rewrites "Recommend Someone" into Searches for Institutional Sites: Intent Analysis of 12,809 Queries
Published
This article was first published in Korean by Intrix Lab on September 28, 2026.
AI does not search using the user's question verbatim. It even changes the search intent. As of 2026-09-14, among 387 user questions in Intrix's experimental corpus, 78% were commercial questions seeking companies or candidates and 21% were informational. Yet the 12,809 rewritten search queries AI actually used in response were 47% commercial, 43% navigational queries specifying particular institutions, associations, or sites, 9% informational, and 1.5% transactional. Nearly half of "recommend someone" questions become "let's open an institution's directory" inside AI. A paper featured on GEO Watch classified questions from 1.03 million buyer personas as 78% informational, 17% commercial, and 4% transactional. We examined the next step: what searches those questions become inside AI.
What did we measure, and how?
Since 2026-08-16, all Intrix AI query experiments have also recorded the search queries AI actually used while generating its responses. From those records, we classified 387 user questions and 12,809 rewritten search queries into four intents: informational (methods and concepts), commercial (finding and comparing candidates), transactional (immediately before an action, such as applying or checking prices), and navigational (finding a specific site, institution, or document). The paper used three categories, but we added a fourth because rewritten search queries frequently specified particular sites and institutions.
What were the numbers?
| Category | Count | Commercial | Navigational | Informational | Transactional |
|---|---|---|---|---|---|
| User questions | 387 | 78.3% | 0.3% | 20.9% | 0.5% |
| All rewritten search queries | 12,809 | 46.8% | 42.8% | 8.8% | 1.5% |
| Rewrites - OpenAI | 11,066 | 41.7% | 49.1% | 7.6% | 1.7% |
| Rewrites - Gemini | 1,813 | 77.8% | 3.4% | 18.1% | 0.7% |
The shift is clearer when broken down by experiment. In the market citation study, 96% of user questions were commercial, but rewritten search queries split into 49% commercial and 49% navigational. In the individual authority experiment, 93% of questions were informational, but 43% of rewrites became navigational. In the preliminary tire rental study, transactional intent rose to 12.5% because AI independently generated queries immediately preceding an action, such as searches for monthly rental fees or contract terms.
What does 43% navigational mean?
When AI receives "recommend a dermatology clinic in Seoul," it searches by specifying institutions and domains, using queries such as "Association of Korean Dermatologists specialist finder" and "site:snubh.org dermatology." The user asked for candidates, but instead of choosing candidates, AI looks for places that maintain candidate lists. This fits the same pattern as citations concentrating on associations, institutions, and directories in industries such as law and medicine. Even if an individual company appears in search results, it cannot become a candidate if AI does not run that search.
Why do OpenAI and Gemini differ?
Half of OpenAI's rewrites are navigational, compared with 3% for Gemini. Gemini keeps search queries close to the user's wording and often skips searching, while OpenAI issues multiple queries, many of which contain site: operators or institution names. This means the search queries to optimize for differ depending on which model answers, even for the same question. The Gemini sample is small, at 1,813 queries, because it includes only responses in which a search actually took place.
What do we do differently as a result?
The optimization target is the search query AI uses, rather than the user's question. If half of those queries point toward institutions and directories, accurate listings there take priority over keywords on your own pages. The remaining half of commercial queries include locations and conditions, requiring pages that answer those combinations. Intrix's research reports also show rewritten search queries by industry.
Limitations
Intrix designed the user questions for experiments, so they skew commercial and do not represent the distribution of questions from general users. This is not a direct comparison with the paper's distribution; it examines the next step along the same dimension. Intent classification is automated, so the boundary between commercial and transactional intent may vary. We record only the rewritten search queries AI exposes externally and cannot know whether the model conducted additional searches internally. These results are not evidence that any particular citation or ranking is guaranteed.
Frequently asked questions (FAQ)
Q. How can I see rewritten search queries for my industry?
They are not visible simply by asking questions directly in the AI interface. You can review rewritten search queries by industry in Intrix's query research reports.
Q. If transactional intent is 1.5%, does that mean searches immediately before purchase are rare?
The share was low for general questions, but it rose to 12.5% for products with complex conditions, such as rentals. Product structure generates transactional searches.
Q. How should we respond to navigational search queries?
Check whether you are listed with the institutions and directories AI specifies, and make the identity of your official site clear so that your domain becomes a target of site: searches.