We improved how brand mentions are recognized in Korean-language AI answers in two ways.
First, a brand name that happens to appear inside an unrelated compound word is no longer counted as a mention. Korean text has no clear word boundaries, so a short brand name overlapping part of a longer word could previously be counted as a mention even when the brand was never actually discussed. This false-positive class is now blocked, while legitimate forms — names followed by Korean particles or brand-family service names — continue to be recognized as before.
Second, when a business category is registered with multiple fields joined by the Korean middle dot (for example, "간편송금·핀테크"), topic relevance is now judged against each component field. Previously such entries were treated as one long token, so clearly relevant answers could be counted as off-topic.
Both improvements were validated against a human-reviewed gold set of real AI answers, and we added per-language regression checks so future changes cannot silently degrade Korean measurement quality below the current baseline.