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 September 6, 2026. The first report to cover a delta from issue #000 (observed July 2026).
This round of observation falls immediately after ChatGPT's model changed from GPT-5.5 (issue #000) to GPT-5.6 Luna. The differences reported below therefore mix whatever changed on the web with whatever changed in the model, inseparably (confounded). This issue makes no claim about which caused what.
With that stated, the change this month reduces to a single phenomenon: citations concentrated on fewer domains. Three angles follow — the citations themselves, business appearance rates, and the differences by category.
On citations: in 11 of the 12 cells, the set of sources returned across 20 runs narrowed to fewer domains. Overlap between runs (Jaccard) rose in all 11, and the number of distinct domains fell in all 12.
On businesses: the count did not thin out the same way. The number of shops and temples in the top group (appearance rate 20% or higher) is roughly unchanged across the 12 cells. What happened was a change of lineup, and its direction differs by language. "Who gets recommended" and "whose information they are recommended from" are moving separately.
On categories: the destination of the narrowing differs. Tea ceremony and kimono concentrated further on referral-type sources (comparison media and OTAs), while zen meditation stayed centered on non-referral sources — public-sector sites, directories, and temples' own sites.
Dominant citation source across the 12 cells (share, %)
| Japanese | English | Simplified Ch. | Traditional Ch. | |
|---|---|---|---|---|
| Tea ceremony | Public 61.2 | Media 59.2 | Top group is OTA and Public (91% combined); rank not determinable | Public 59.1 |
| Kimono | Media 70.5 | OTA 92.5 | Media 54.7 | Rank not determinable (Public 33.3 / Media 31.4 / OTA 25.5 / Owned 9.8) |
| Zen meditation | Top group is Directory and Public (88% combined); rank not determinable | Owned 50.8 (n=19) | Public 52.8 | Public 58.6 |
This issue writes only down to the granularity that can be determined. Where the gap between first and second place is confirmed by bootstrap, the leader is named; where it is not but the top two types are stable, only "the top group is A and B" is written; where neither holds, no ranking is given. Details in the Japanese full text, §5.
Only one movement in this issue can be called a change of leader: in kimono / English, media (45.5% in issue #000) receded and OTAs took first place at 92.5%. Tea ceremony / Simplified Chinese and zen / Japanese look like reversals in the table, but the gap to second place within this issue is 3.6 and 2.6 points — indistinguishable from the error inherent in 20 observations. "Too close to call" is the accurate description.
The structure found in issue #000 has not broken. Splitting sources into referral-type (comparison media and OTAs, which earn referral commissions) and non-referral-type (public-sector sites, businesses' own official sites, directories, UGC): once the error of 20 observations is carried as an interval, kimono's four cells have a lower bound of at least 44.4% even at their lowest, and zen's four cells an upper bound of at most 36.1% even at their highest — the intervals never overlap. The gap widened from 2.6 points in issue #000 to 8.3 points.
Method
Runs used ChatGPT (GPT-5.6 Luna, web version) in a temporary chat while logged in, on 2026-08-18 to 20, in Kita-ku, Kyoto, over a home fibre connection. Queries followed a keyword pattern of "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, logging the business names and cited domains along with date/time, location, network, and model name. Sources were classified into six types — Public, Owned, OTA, Media, Directory, UGC. Exact query strings and classification definitions are in the Japanese full text, §2.
One exception: zen meditation / English (QZ-EN) is aggregated at n=19, because one of the 20 records was found corrupted by a paste error. The run was voided rather than re-taken — selecting a replacement run after a corruption is not a neutral operation, and it was confirmed that the choice of replacement changes the rank determination for that cell.
Scope & limitations
Every difference reported here is fully confounded with the model change (GPT-5.5 → GPT-5.6 Luna). It cannot be written that the web's information ecology itself changed; what is observable is only how ChatGPT's answers differed before and after the model changed. The noise-floor thresholds measure variation within a single model, so across a model change they are treated as a lower bound only. 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=19–20 runs · 676 rows at cited-domain level)
- Business info card pathway placecards.csv (same conditions · 171 rows)
- Noise floor: measurement method and thresholds
- 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.