AI Picks Its Shortlist Before It Searches — 72 Queries Across 12 Korean Categories

Summary: AI search does not run the question you typed. It composes its own search query first — and in some categories that query already contains brand names the user never entered. We asked a category-leader question across 12 Korean categories and captured 72 queries the two models actually executed. Categories separated cleanly into those where brand names appear in the query and those where they never do. Which side your category falls on determines what work matters next (as of 2026-08-17).

---

Three-line summary

• AI writes its own search query before fetching anything. In food delivery, the names of three Korean operators appeared inside that query without the user typing any of them.
• In others — eye clinics, interior contractors, pilates studios — no brand name ever appeared across every run. These are markets with many providers and no dominant brand.
• The work differs by side. If names are already in the query, the task is entering that shortlist. If no name appears yet, the task is establishing position while the shortlist is still unformed.

---

How we observed it

We read the search-query field recorded in the model API response, not the browser screen. OpenAI responses record each executed search call and its query string; Google responses record the list of queries the model used. Neither is visible in the UI.

For 12 Korean market categories we asked, in Korean, which brands lead that category, and repeated it three times per model — 72 observations. Every raw query string was retained.

Labelling rules were fixed before collection. Language was decided automatically by whether the query contains Hangul. Brand presence counted only real proper names; procedure terms such as LASIK, generic nouns, and generic acronyms such as ATS were not counted.

Perplexity was excluded because its API does not expose the query. That is itself part of the result — for the same question, one of the three major engines gives you no way to inspect the query at all.

All 72 collected queries are published (JSON dataset, CC BY 4.0). Every table and reading below can be checked against that file directly.

Company and service names the models wrote into their queries, along with any domains, are masked in the published data. The analysis rests on how many names appeared, not on which companies they were, so masking does not affect verification. It makes clear that this is not an article evaluating named companies.

---

Finding 1 — Markets with a shortlist already fixed, and markets without one

Results for the 10 domestic categories, 60 observations.

Category	Brand name in query	No search executed
Tax bookkeeping services	4 / 6	1
Food delivery	2 / 6	4
Meal-kit subscription	1 / 6	0
Electronic approval / groupware	1 / 6	2
Marketing automation SaaS	1 / 6	0
Eye clinics	0	1
Apartment interior contracting	0	4
Pilates studios	0	0
Recruiting management software	0	1
AI search visibility diagnostics	0	0
Total	9 / 60	13 / 60

What went into those queries: food delivery drew in three Korean delivery apps by name, tax bookkeeping four tax platforms, and marketing automation three overseas marketing SaaS vendors. Everything around the names was generic wording — the Korean equivalents of "Korea", "leading brands", "official", "features" — with the company names slotted in between.

The user entered no brand names. The model added them before executing any search.

The separating factor was market structure. All five categories where brands appeared are markets divided among a small number of providers. The categories where no brand ever appeared fall into two types: fragmented local markets with many providers spread across regions (eye clinics, interior contracting, pilates), and emerging markets where no leading provider is established yet (recruiting software, AI search visibility diagnostics).

Food delivery needs both columns read together. Two of six runs put brand names in the query, and four of six answered without searching at all. Whether it searched or not, the brands it would name were already determined. For a new entrant in that market, ranking well in search results is not sufficient on its own.

---

Finding 2 — The model targets a specific brand's official site

Four of the collected queries were not open searches but domain-scoped lookups using the site: operator, each pointing at one company's official domain — a tax platform, a food conglomerate, and a workplace collaboration suite. The rest of each query named the feature being verified: bookkeeping service, subscription delivery, electronic approval.

Once a brand is confirmed as a candidate, the model goes directly to that brand's official site to verify. At this stage the brand's own pages serve as the evidence, not third-party media. But that lookup only happens if the name entered the candidate set in the earlier step.

---

Finding 3 — Some categories produce English queries, and models differ

Across the 10 domestic categories both models searched in Korean. Adding categories whose discourse formed in English changes that.

Category asked	OpenAI model	Google model
Category name containing English acronyms	English 3 / 3	Korean 3 / 3
Product analytics tools (no acronym)	English 3 / 3, brands in every run	Korean 1 / no search 1 / mixed 1

The English queries followed the shape Korea product analytics tool brand service … local Korean analytics tools, with five overseas product analytics vendors named in the middle. The question was about the Korean market, yet both the query and the candidates came from the English-language side.

Two things follow. First, English queries appear even without an English acronym. 프로덕트 애널리틱스 (product analytics, written in Korean) carries no acronym, yet the queries came out in English. The separating factor is not the spelling but which language the concept's discourse formed in. Second, models differ. Under identical conditions the Google model kept Korean.

Where queries come out in English, domestic providers are structurally disadvantaged. The search runs against an English index, so domestic documents are not included in the results, and adding a country name to the query still returns overseas services. In that case the documents competing for candidate entry are English-language documents.

---

Compared with the overseas measurement of citation rates

This study only covers what went into the query. A published analysis measured the next step — citation rate.

"Seven words, no brand names. Before it fetched anything, ChatGPT wrote itself this search." — Suganthan Mohanadasan, Search Engine Journal (2026-08-14)

In that analysis, 21 of 27 conversations had brands in the first query that the user never typed, and citation rates separated as follows.

Group	Sample	Citation rate
Named in the query	119	68.9%
Fetched but never named	515	2.1%

These figures should not be taken at face value. They come from a single account over two days, and the author states the questions skew toward software and AI tools. The author also distributes a GEO analysis tool. As the author puts it, "these are my account's shortlists, not ChatGPT's." Treat the direction of the gap as the takeaway, not the ratio.

The direction holds in the author's separate dataset as well. There, Reddit was fetched 278 times and cited 11 times, while YouTube was fetched 201 times and cited zero times (author's write-up). Being fetched and being cited are different outcomes.

Structurally this connects to query fan-out. Google has confirmed that AI Mode issues 12 searches by default for one search (Google official blog). What this study looked at is what fills those queries.

---

What to do about it

First, determine which type your category is. Ask the category-leader question five times and record the queries the model executed. A browser extension (FanoutFox) works, and so does reading the search-query field in the API response.

If specific brands already appear in the query, the task is entering that set. What the model consults when assembling that set is third-party comparison and recommendation material. Adding more owned blog posts does not change the set, and neither structured data nor page speed addresses this class of problem.

If no brand appears in the query, the shortlist is not fixed yet. The model reads fresh search results each time and assembles an answer from them. Owned content contributes more under these conditions. Conditions change once the market consolidates, so re-check on a schedule.

If the queries come out in English, this judgment precedes both of the above. Growing Korean-language content will not place you in the results of an English query. Confirm which language the target documents are in before allocating resources.

---

Frequently asked questions

If our name is not in the query, is good content pointless?

No. Half the categories in this study never had a brand name in the query, and in those markets the model reads fresh search results to build its answer — conditions where content contributes more, not less. Where names already appear, content alone will not change the set. The answer depends on which type your market is.

Owned blog content or third-party mentions first?

If specific brands already appear in the query, third-party comes first. The material a model consults when assembling a candidate set is comparison and recommendation content, and owned documents are rarely used as evidence for that set. That said, as this study showed, once a candidate is confirmed the model looks up the official site directly. Owned pages are used at that stage.

With a small budget, where do we start?

Start by inspecting the query. It costs nothing and takes about ten minutes. If the queries come out in English and the budget goes into Korean-language content, improvements will not appear in those results and the measurement will not move.

Can we apply these numbers to our own company?

We do not recommend it. Three runs per category is not enough to fix a rate, and the same category produced different queries across runs. What this study shows is the contrast between categories, not an exact rate for any one of them. Measuring in your own category is more accurate.

If models disagree, which one should we trust?

Both. In this study one category produced English queries from the OpenAI model and Korean queries from the Google model. Looking at one model shows half the market. Some engines, such as Perplexity, do not expose the query at all, so what is observable differs by engine.

Once our name is in the query, does it stay?

Even across three runs of the same category, the queries differed. Runs with brand names and runs without them both occurred. This is a metric to re-check periodically, not to confirm once.

---

Related reading

• One Query Becomes 12 Searches — Sub-Query Coverage for AI Search Visibility
• Best Lists Rule AI Citations — But Self-Promotion Backfires Unless the Fit Is Right
• 89% of AI Search Demand Has No Clear Brand Owner — What 1,094 Categories Reveal
• When AI Recommends Competitors but Omits Your Brand — Closing the Visibility Gap