Ask an AI assistant "best Korean dentist in Koreatown LA" and ask it "LA 한인타운 치과 추천해줘," and you are not asking the same question. In our measurement, the two languages produced answers built on almost entirely separate sets of sources — and therefore surfaced different businesses.
For a Korean-American business, that is not a curiosity. It means the visibility you have in one language tells you very little about the visibility you have in the other, and most owners are only ever checking one of them.
What We Measured
On August 24, 2026 we ran three local-business questions through three AI engines from a US location and collected every source each answer cited.
| Item | Detail |
|---|---|
| Date | August 24, 2026 |
| Questions | "LA 한인타운 치과 추천해줘" · "best Korean dentist in Koreatown LA" · "뉴저지 포트리 한인 회계사" |
| Engines | ChatGPT, Gemini, Perplexity |
| Location | United States |
| Citations collected | 108 |
The first finding is the boring but necessary one: all nine answers named real businesses, with addresses and phone numbers. The common worry that AI simply does not know small ethnic neighborhoods did not hold up. It knows. The question is only whether it knows you.
The Two Languages Cite Different Worlds
Here is where every citation came from, split by the language of the question.
| Source type | English questions | Korean questions |
|---|---|---|
| Korean-American media and business directories | 0% | 61% |
| The business's own website | 65% | 32% |
| Mainstream US directories (Yelp, Zocdoc, Reddit) | 35% | 7% |
Read the first row again. In English answers, Korean-American media and directories were cited zero times. In Korean answers they carried the majority of the sourcing.
This is the practical consequence: a Korean-American business that measures its AI visibility with an English-only tool — which is what every mainstream visibility tool is — is measuring the half of the market that is often not its core customer base. And a business listed thoroughly across Korean-American directories may still be absent from every English answer.
The 61% Deserves an Asterisk
We would rather publish the caveat than let you over-read our own number.
That 61% is an average across three engines, and Perplexity pulls it upward because Perplexity cites far more sources per answer than the others do. Split by engine, Korean-language answers look quite different:
| Engine | Share of citations that were the business's own site | Share that were Korean-American media |
|---|---|---|
| ChatGPT | 70% | 20% |
| Gemini | 33% | 67% |
| Perplexity | 23% | 68% |
So the honest version of the finding is: in Korean, Gemini and Perplexity lean on Korean-American media, while ChatGPT leans on the business's own website — even in Korean. Since ChatGPT is the assistant consumers actually use most for this kind of question, that last column matters more than its share of our sample suggests.
The Sites That Publish in Korean Get Cited in Korean
One more pattern in the data. Of the 17 business websites cited anywhere in the Korean-language answers, four actually had Korean-language content on them. All four were accounting firms.
Those four appeared only in Korean answers — zero appearances in the English answers. The eight sites that were English-only had no Korean content at all and did not surface in Korean answers.
That is a small sample and we are describing correlation, not proving cause. But the shape of it is consistent with everything else here: engines answer a Korean question with pages written in Korean, and an English question with pages written in English. If your website exists in one language, you are eligible for one of the two conversations.
What We Are Not Claiming
- Three questions, one snapshot. This is a single day of measurement, not a longitudinal study. AI answers vary between runs; we have watched rankings move substantially in a single day in other datasets we maintain.
- Correlation, not causation. We did not prove that publishing Korean content causes Korean citations. We observed that the sites with Korean content were the ones cited in Korean.
- Directories still matter. Zero citations in English does not make Korean-American directory listings worthless — they carried most of the Korean sourcing for two of three engines. They are complementary, not central.
What This Suggests You Should Do
- Check both languages. Ask each engine your customers' actual question in English and in Korean. Compare who gets named. If you have never done this, the gap is usually the surprise.
- Treat your own domain as the asset. It was the most-cited source type in English and the leading one for ChatGPT in Korean. Directory profiles are shared ground that every competitor also stands on; your own site is the only surface you control.
- Publish in both languages if you serve both. Not translated boilerplate — pages that answer the questions customers actually type.
- Measure repeatedly. A single check tells you almost nothing, because the answers move.
FAQ
Does this apply outside of dentistry and accounting?
We measured those two categories in LA Koreatown and Fort Lee, New Jersey. We would expect the language split to generalize to other Korean-American service categories, but we have not measured them yet, and we will publish the numbers when we do rather than assume.
If AI already names businesses in my neighborhood, why isn't it naming mine?
An AI answer names a handful of businesses, not a page of results. Being findable is not the same as being selected. In our data, the businesses that got named were the ones whose own sites or whose Korean-American directory entries gave the engine something to cite.
How often do these answers change?
Often enough that one measurement is not a conclusion. In a separate 99-business study we ran in Korea, one business moved from 25th to 8th between two measurements taken a day apart with no intervention.
RECO AI Lab measures how ChatGPT, Gemini and Perplexity answer questions about local businesses — in English and Korean — and publishes bilingual content on the client's own domain. Our methodology and pricing are public at recoailab.com.