Claude (an AI model) reads each post and labels its stance toward the brand it was collected for. The overall mood of the text doesn't count. Positive means praise or a recommendation, negative a complaint, neutral no stance (most news headlines about store openings or earnings land here), and mixed both praise and a complaint. Posts that turn out not to be about the brand, such as name clashes or spam, are dropped before counting.
A brand's sentiment score is (positive − negative) ÷ all labeled posts in the selected window. It runs from −1 (all negative) to +1 (all positive). Weeks with a lot of news sit closer to 0, since factual articles count as neutral. Each post also gets one to three topics (taste, price, service and so on) for the Topics chart.
Scores are kept separate by source because each source starts from a different baseline. The All / News / Posts / Comments & X toggle and the four tiles at the top follow this split.
News is press coverage. It is mostly neutral and turns negative when there is a controversy. Posts are blog, Instagram and Threads posts people wrote themselves. They lean positive, because people who bother to write a café review usually liked it. Comments & X are replies under posts plus X posts, which read like comments. This is the closest to unprompted customer reaction. The brand's own Instagram posts are left out of all three. Compare a score only with the same source for other brands.
Sponsored posts. Posts that disclose a paid review (체험단·협찬·원고료) are tagged 협찬/체험단 in the article list. Korean law requires the disclosure, but many bloggers skip it, so the real number of sponsored posts is higher than what is tagged here.