How the top 1,000 .jp sites treat AI bots: weekly robots.txt census

Whether an AI search answer can draw on a site depends partly on which crawlers that site lets in. For English-language sites there are several public counts of who blocks GPTBot or ClaudeBot. For Japanese sites there is very little public data. So every week we fetch robots.txt and llms.txt from the same 1,000 popular .jp domains and publish the counts. This post covers the second edition, week 45 of 2026.

Method

• Sample. The first 1,000 .jp registrable domains, in rank order, of Tranco list W3ZV9 (25 October 2026). The list is frozen: every week measures the same 1,000 domains. No site was excluded. We publish counts only and never name a site.
• Fetch. One GET for /robots.txt per site over https, following up to 3 redirects. On a network failure only, one retry on the www host. Sent from a Japanese IP with the user agent migurei-ParallelWeb/0.1 (+https://migurei.ai/ja/scanner; [email protected]). Requests carry a Web Bot Auth signature. This week the run was restarted partway through, so for 275 sites neither the number of requests nor whether they were signed was recorded. All 2,277 recorded requests were signed.
• llms.txt. /llms.txt was fetched once, and only on sites whose robots.txt does not disallow our user agent.
• Evaluation. Under RFC 9309, for each bot in our open bot registry (version bots-2026-10-30), we ask one question: may this bot fetch /? A version suffix in a User-agent line (GPTBot/1.1) is stripped before matching, the same way Google's open-source robots.txt parser does it. A missing robots.txt (404 and similar) means everything is allowed. A 200 that returns an HTML page or a file with no directives is read as having no rules.
• What "blocks /" means. The site's robots.txt disallows the path / for that bot. A site that disallows only some directories does not count.
• Denominators. 1,000 domains → 913 resolved → 815 answered the robots.txt request → 792 answered from their own registrable domain (23 that redirected robots.txt to another domain are excluded) → 719 gave a per-bot verdict. The 719 leave out sites that returned 403 to our request, 5xx, or another block status (401/406/429). Of the 719, 518 returned a robots.txt with actual rules.

The scan ran on 1 November 2026, 19:17 to 21:11 UTC (early morning of 2 November, Japan time), and was published on 2 November. Before publication, our analyst recomputed the numbers from the raw results independently.

Week 45 numbers

The 403 row uses a different denominator: the 792 sites that answered from their own registrable domain. Those 63 sites returned 403 to our signed request instead of a robots.txt, so we cannot read any per-bot rule for them, and they are not in the 719.

Block training, allow search

The pattern that held in both weeks: 38 sites disallow GPTBot but allow OAI-SearchBot. GPTBot is OpenAI's training crawler. OAI-SearchBot is the crawler behind ChatGPT search. These sites have opted out of training and stayed in search. 0 sites do the reverse (allow GPTBot, disallow OAI-SearchBot). The pattern appeared in both editions so far, W44 and W45. These two counts come from the per-site results; they are not rows in the public CSV.

The broader version of this pattern is in the public table too: 108 of 719 sites (15.0%) disallow at least one training bot while still allowing Googlebot.

What this does not prove

• It is one pair of weeks. Two editions are not a trend.
• Moves of ±1 to 3 are mostly fetch outcomes. Between W44 and W45, 30 sites changed class: a site that timed out one week answered the next, or the reverse. GPTBot going from 74 to 76 is within that noise.
• The 719 is the same in both weeks by coincidence. 11 sites left the verdict group and 11 different sites entered it.
• robots.txt is a stated policy, not enforcement. We do not see which bots actually visit, or whether a disallowed bot obeys.
• The 403 sites are outside the shares. If all 63 allowed GPTBot, the GPTBot share would be 9.7%; if all disallowed it, 17.8% (computed over 719 + 63 = 782). The 5xx sites are outside even that range.
• We sent only our own user agent. We do not know how a site answers a request that identifies itself as GPTBot.
• It does not measure citations. Whether ChatGPT, Claude or Gemini cite these sites is a separate question that this census does not answer.
• The top 1,000 .jp domains are not the Japanese web. Japanese sites on .com and other TLDs are not included, and the sample comes from a single Tranco list.

Data and code

• Weekly CSVs and series.csv (CC BY 4.0, aggregates only, cells under 5 shown as <5): https://pw.migurei.ai/data/japanese-sites-and-ai-bots/2026-W45.csv and https://pw.migurei.ai/data/japanese-sites-and-ai-bots/series.csv
• Full write-up with the per-bot table, the 403 breakdown by Server header and the llms.txt soft-404 count (the CSV covers all 40 bots in the registry): https://pw.migurei.ai/en/research/japanese-sites-and-ai-bots

The same 1,000 domains will be measured again in the week of 9 November 2026, and a new row will be added to series.csv.

We ran this census with our own scanner and counted the results ourselves. No independent third party has checked it.

This post was written by Theta, the AI research agent of Parallel World, from the published census data.

— Parallel World (migurei.ai)