This page is a summary edition. To keep maintenance effort focused on the observations themselves and on the accuracy of the Japanese original, non-Japanese editions are published as summaries. The full report — all sections, the business-level analysis, and the determinability rules — is published in Japanese: full text (日本語). The observation data is language-neutral and can be downloaded below.
Summary
First edition, published October 4, 2026. The third round of observation, following issue #000 (July 2026) and issue #001 (August 2026).
Issue #001 reported a sharp rise in the share of referral-type sources — comparison media and OTAs, which earn referral commissions — across most cells. This round, that rise did not hold evenly.
Zen meditation stayed flat and in the single digits across all three rounds (4-language average 6.8% → 9.5% → 6.1%). Tea ceremony (24.2% → 59.0% → 32.6%) and kimono (49.3% → 81.3% → 55.5%) spiked in issue #001 and partially receded this round, but both remain above their issue #000 baseline.
Cell by cell, in 6 of the 12 cells the net change from issue #000 to issue #002 still exceeds the noise-floor threshold: 2 cells (tea ceremony/Japanese, kimono/English) moved below their starting point, and 4 cells (tea ceremony/Traditional and Simplified Chinese, kimono/Simplified Chinese, zen/English) remain above it. Kimono/Simplified Chinese, despite receding from its issue #001 peak, is still the single highest value across all 12 cells at 66.7%. It is too early to call the issue #001 spike entirely temporary.
Individual business appearance rates also shifted substantially between issues #001 and #002. With only 20 runs per cell, this metric is inherently noisy, so this issue treats it as a secondary record rather than a main finding.
The size of the pullback differs by category and language: Japanese tea ceremony and English kimono pulled back the most, moving below their issue #000 starting point, while Traditional and Simplified Chinese tea ceremony and Simplified Chinese kimono remain well above it.
Referral-type share across three rounds (category average, %)
| Issue #000 | Issue #001 | Issue #002 | |
|---|---|---|---|
| Zen meditation | 6.8 | 9.5 | 6.1 |
| Tea ceremony | 24.2 | 59.0 | 32.6 |
| Kimono | 49.3 | 81.3 | 55.5 |
Referral-type = comparison media and OTAs (earn referral commissions); non-referral-type = public-sector sites, businesses' own official sites, directories, and UGC. Per-cell detail and the noise-floor thresholds (9–23 points, cell-dependent) are in the Japanese full text, §3.
A further refinement: within the "media" category, a meaningful share of domains are themselves affiliate sites that earn referral commissions from OTAs such as GetYourGuide or Klook, rather than independent editorial media. Classified as media under the project's domain-level rule, they are closer to referral-type in revenue structure. Reading the OTA-label decline alone as a retreat of referral-type sources overall would be inaccurate; what appears closer to the mechanism is that OTA-labelled platforms receded and affiliate-style media absorbed part of that position. Kyoto's official tourism site, ja.kyoto.travel, stayed in the top two cited domains in 10 of 12 cells throughout, unmoved by this swing. Full discussion in the Japanese full text, §5.
Method
Runs used ChatGPT (GPT-5.6 Sol, web version) in a temporary chat while logged in, on 2026-09-21 to 30, in Kita-ku, Kyoto, over a home fibre connection. Queries followed the same keyword pattern ("Kyoto + category term + recommendation term") in Japanese, English, Simplified Chinese, and Traditional Chinese, for tea ceremony, kimono rental, and zen meditation. Each query was run 20 times (n=20) consecutively. Sources were classified into the same six types as previous issues — Public, Owned, OTA, Media, Directory, UGC.
Two cells — kimono/Simplified Chinese (an account-plan error) and zen/Traditional and Simplified Chinese (an incomplete citation-logging format) — were re-run under proper conditions within the same month; all 12 cells are reported at n=20. Exact query strings and classification definitions are in the Japanese full text, §2.
Scope & limitations
Every difference reported here is confounded with two model changes across the series (GPT-5.5 → GPT-5.6 Luna → GPT-5.6 Sol); no causal attribution is made. While aggregating this issue, a citation-parsing bug was found and fixed — two legacy citation formats were not being recognized, which had affected citation counts in 10 of the 12 cells — and 11 previously unclassified domains were newly classified as part of the fix. The noise-floor thresholds were calibrated for single-period variation within one model; applying them across two periods (issue #000 to #002), as this issue does for the 6 threshold-exceeding cells, is treated as a lower-bound estimate only and has not been separately verified for multi-period comparisons. The 72 individual-brand threshold-exceeding changes noted between issues #001 and #002 are kept as a secondary record rather than a main finding, given the high sampling variance inherent to n=20 binary observations. This study is limited to ChatGPT, a single observer, and a single execution location, and covers only recommendation-type queries. Full discussion in the Japanese full text, §7.
Data
- Raw data data.csv (CC BY 4.0 · 12 cells × n=20 runs · 648 rows at cited-domain level)
- Business info card pathway placecards.csv (header only this issue — no matching citation format detected; see data README)
- Noise floor: measurement method and thresholds
- Issue #001 (the prior round): Monthly Report #001
- Issue #000 (t=0, the baseline for comparison): Monthly Report #000
This is a personal research and information-sharing project, unaffiliated with any organization, university, or company. Data is published free of charge under CC BY 4.0 and is not for sale.