Korea AI Pulse
Weekly Insights

Every round,
the data tells a new story.

Ask the same questions a week later and the answers change. This page records, round by round, what was newly observed that week. Figures are preserved as they were at observation time — for current totals, see Stats.

  1. 2026-08-01

    Round 2 — visibility is a flow, not an asset

    GPT replaced 97% of its citations within days

    Across 19 same-question, same-model pairs asked ~3 days apart, GPT fully swapped its cited domains in 18 (overlap 0.03). Claude (0.26) and Grok (0.21) tend to return to the same sources. GPT visibility is not an asset you keep — it re-derives from search rankings every time.

    See the underlying data →

    The engines started citing this site

    Gemini cited koreaaipulse.com in 16 answers this round, and Grok directly cited two of our question pages. The archive that records AI answers has itself become a source for AI answers — a loop we will keep documenting transparently.

    See the underlying data →

    The same question flipped its answer in a week

    Last round GPT answered “best fintechs to move to as a developer” with Korean startups that had open roles; this round both GPT and Claude listed only foreign companies. A concrete case of how violently answers swing week to week.

    See the underlying data →

    Gemini is citing far more often now

    Gemini answers with citations jumped from 40 to 69 out of 81. The switch to a search-grounded model plus a retry-once policy keeps paying off.

    See the underlying data →
  2. 2026-07-18 ~ 07-27

    First observations — four different worlds

    In 86% of questions, zero brands were shared by all engines

    In 86% of questions answered by two or more engines, no brand was recommended by all of them. 60–73% of each engine's brands were exclusive to it — the four engines are effectively four independent distribution channels.

    See the underlying data →

    Only one engine really reads Naver blogs

    Grok alone produced 68% of all Naver-blog citations. Claude drew 82% of its citations from official websites, while Gemini pulled 20% from Tistory/Brunch blogs — each engine reads a different web.

    See the underlying data →

    Self-published claims become the recommendation

    A clinic's own line — “over 10,000 orthopedic surgeries” — was quoted verbatim as the reason to recommend it, and law firms with keyword domains were recommended for exactly that. 64% of all citations point to official or self-owned sites.

    See the underlying data →

    Refusals are data too — empty seats exist

    GPT and Claude flatly refused to name fortune-telling services, and even high-volume questions like indemnity insurance had zero named brands on some engines. Empty seats go to whoever gets cited first.

    See the underlying data →

A new entry lands here after every collection round. To check whether your brand appears in any answer, search the collection log.