Why your Korean brand is invisible in US AI search

Ask ChatGPT for the best Korean sunscreen for oily skin. You get a shortlist. Run the same question for baby carriers, air purifiers or pet food and read what comes back — that is the test this whole post is about, and it takes four minutes. Whatever names appear, the ones that do not appear are not absent because the product is worse.

The brands in that long tail are not there because their product is worse. The causes are structural. Once crawler access is ruled out — always the first check, because it is the cheapest — these four are what we expect to find. Three map onto causes Searchd diagnoses directly; the second is the corroboration cause in its harder form, where the outside source exists and is wrong.

1. Your English is translated, not written

Translation optimises for fidelity. It takes a Korean marketing sentence and produces an English sentence that means the same thing. What it does not do is introduce the specific, checkable claims that make a passage quotable.

The result is copy that reads perfectly well and offers a model nothing to lift. “Our sunscreen provides comfortable daily protection with a light finish” is grammatical, accurate, and useless — every competitor says it, and no assistant will quote it.

What a model can use: “SPF 50+ PA++++, chemical filters, tested at the full 2 mg/cm² application rate, no white cast on Fitzpatrick IV–VI.” (Sample copy for a fictional brand — the shape is the point, not the values.)

The fix is not better translation. It is writing English copy natively against the facts, then having it checked in Korean for accuracy — the opposite of the usual direction.

2. Amazon and your retailers own your product facts

When a model looks for your product’s specification, it finds whichever source is most retrievable. For a cross-border brand that can be a marketplace listing, written by a distributor rather than by you, and out of date by a product revision or more.

Four errors worth checking for on your own listings:

• Weight limits stated in pounds that were mis-converted from kilograms
• Age ranges that reflect an older product revision
• Ingredient lists missing a reformulation
• Dimensions transposed between length and width

Two consequences. First, the model may be describing a product you no longer sell. Second, and this is the part brands under-price: for regulated categories, an assistant confidently stating a wrong SPF or a wrong age rating is a compliance exposure with your name on it.

3. The English-language conversation happened without you

AI answers lean heavily on the sources where a category is discussed at length: forum threads, YouTube round-ups, comparison articles, review publications.

For a Korean brand, that conversation may already exist — in Korean. Naver carries long-form review posts, Korean YouTube carries multi-brand comparison videos, and the discussion runs deep in Korean communities. The model answering an American shopper’s question in English is not reading any of it.

The English-language record, meanwhile, might be three Reddit comments and a listicle that misspells the brand.

This is the constraint that takes months rather than weeks. It is also the reason on-site work alone plateaus.

4. Your entity does not resolve

Check how many strings your brand exists as. The usual ones:

• The brand name in English on the website
• A romanised transliteration used in some meta tags
• The legal entity name in the footer and privacy policy
• A seller name on Amazon that matches neither

A model naming a brand is making an assertion about an entity. Faced with four candidate strings it cannot confidently unify, the low-risk answer is to name a competitor whose identity is unambiguous.

We scope the fix at a day of decisions and a week of edits: choose one canonical string, use it byte-identical everywhere, and link your profiles with sameAs in your Organization markup.

The order to fix them in

Sequence by speed and dependency, not by size of prize.

Times below are estimates we scope engagements against, not measured averages. When we have real figures they will replace them.

The first three move the number fastest because they need no one else’s cooperation and no time to age. The last two are where the ceiling is.

Measure before you start

Every item above is arguable until there is a baseline. Run twenty category questions across the engines your market uses, record who gets named and which sources they cite, and keep the file. Otherwise in three months you will have opinions about what worked and no way to settle them.

If you would rather not do it by hand, we run that panel for free.