The four engines are not one channel. They are four independent distribution networks. Break the roughly one thousand citations we collected down by source type and it becomes obvious that each engine reads a different web. The table below is very nearly the whole guide.
| Source type | GPT | Claude | Gemini | Grok |
|---|---|---|---|---|
| Official site / own domain | 56% | 82% | 66% | 56% |
| Naver blogs | 18% | 0% | 0% | 18% |
| Brunch / Tistory | 1% | 2% | 20% | 8% |
| Restaurant platforms (Diningcode etc.) | 12% | 13% | 3% | 7% |
| Social (YouTube, Instagram) | 0% | 0% | 7% | 8% |
* Share of each engine’s own citations. Naver is South Korea’s dominant search portal and its blog platform is where most consumer reviews live; Tistory and Brunch are the two large Korean long-form blogging platforms. The underlying data is browsable domain by domain on source analysis.
GPT — the engine that searches again every time
GPT runs a search before answering almost every time (88% of measured runs). The consequential property is that asking the same question three days later replaces 97% of its citations — visibility here is a function of whatever the search index says at that moment. So the GPT play is simply holding a search ranking consistently. One oddity worth knowing: GPT was the only engine in our data to cite official KakaoTalk channels (pf.kakao.com). Korea runs on KakaoTalk, and keeping hours, pricing and services listed on that channel is a real route into GPT’s answers.
Claude — the official-website engine
82% of Claude’s citations are the business’s own site. It barely reads blog reviews or social posts at all. If you want to appear in Claude there is one answer: build a page on your own domain that maps one-to-one onto the question, and write the specifics down in a structured way. The flip side is that Claude is also the most reluctant engine to name anyone at all when it isn’t confident — we see it decline repeatedly on legal and fortune-telling questions — which makes it the hardest engine to break into.
Gemini — half the time it answers from memory
Share of runs where Gemini actually executed a search before answering (measured) — the rest came from trained memory alone
In the half where it doesn’t search, only brands that were already famous at training time appear. No amount of marketing today gets you into that half. In the half where it does search, Gemini cites Tistory and Brunch for a full 20% of its sources — it is the only engine for which those Korean blogging platforms are a viable route.
Grok — the only engine reading the Naver ecosystem
Share of all Naver blog citations across the whole dataset that came from Grok alone (measured)
Naver blogs, Instagram, Danggeun (Korea’s neighbourhood marketplace) and YouTube are read, in practice, only by Grok. Put differently: Korea’s enormous sponsored-blogger industry is visible to one AI out of four. Worth checking which engine your budget actually reaches before you spend it.
The short version
- GPT: hold your search ranking, and keep your KakaoTalk channel current
- Claude: one dedicated page on your own site per question
- Gemini: informational posts on Tistory and Brunch
- Grok: Naver blog and social reviews — the only engine sponsored posts reach
What to actually publish is covered, with measured examples, in the four patterns behind getting recommended.