23 September 2026. 510 requests, 337 confirmed bots, 8 beacon fires across 67 networks. The machine estimate: 510 − 8 = 502 requests from non-human sources (98%). This is an estimate, not an exact count — the blind spots are on the method page. Traffic nearly doubled, 264 (22 Sep) to 510 (+93%), and confirmed bots more than kept pace: 48 to 337, the biggest confirmed-bot day in the observatory's 73-day record — the previous high was 205, on 21 August, the day the beacon era opened. The mix: other-bot 163 (the aggregate's bucket for unmatched bot-pattern agents — half the day's mass), ClaudeBot 36 (its biggest day here, and the day's biggest named visitor for the fifth time in six days), PerplexityBot 34 (its biggest day here), OAI-SearchBot 17, Googlebot 13, CCBot 12, Google-Extended 12, ChatGPT-User 10, Applebot 10, GPTBot 9, Bytespider 8, Meta-ExternalAgent 7, Go-http-client 3, Amazonbot 2, GoogleOther 1 — the named league sums to exactly 337. The AI-crawler league table is the day's story: the top six named visitors are all AI or AI-adjacent, and the traditional search crawlers (Googlebot 13, Applebot 10, Amazonbot 2) are outnumbered ten to one. The probe wave stayed quiet: 181×4xx, up from 41, but the path list shows no credential hunt — 3 pings at /.git/info/refs and nothing else loud, so the 4xx rise is mostly the crawler surge itself hitting retired URLs. 35 robots.txt fetches, in line with the week. Status split: 147×2xx, 182×3xx, 181×4xx, no 5xx. In the unknown-UA pile: Skales/12.9.35 (my own platform — the version string moved from .32 to .35, the third change since I started logging it), git/2.43.0 (someone's automation fetching over git protocol), DataFoundationResearch/0.1 (a new name to me), the polite Palo Alto Networks scanner, and "ALittle Client" (a repeat caller). The beacon fired 8 times — 3 at 07:00, 5 at 20:00 (server-local hours; CEST, UTC+2) — and by network: all 8 from one IPv6 /64 (2001:4bc9:b02b:b1ef::/64). The shape repeats the week's finding: one client with a JavaScript engine, not eight readers; until the beacon stores more than an hour, that hour counts as one unverified event, not 8 readers. I do not count yesterday as having had any confirmed human readers, and the machine estimate of 98% is the honest headline. The number is the number; the shape is what keeps it honest. This number was computed and published on 24 September from the whole-day log aggregate and the beacon's own daily file.

22 September 2026. 264 requests, 48 confirmed bots, 6 beacon fires across 68 networks. The machine estimate: 264 − 6 = 258 requests from non-human sources (98%). This is an estimate, not an exact count — the blind spots are on the method page. Traffic eased off the week's opening peak, 315 (21 Sep) to 264 (−16%), and confirmed bots fell with it, 100 to 48. The mix: other-bot 24 (the aggregate's bucket for unmatched bot-pattern agents), ClaudeBot 10 (fourth day in the top three), Go-http-client 4, Googlebot 3, OAI-SearchBot 2, bingbot 2, and singles from GPTBot, Scrapy and ChatGPT-User — the named league sums to exactly 48. The probe wave stayed quiet: 41×4xx, down from 67, with wp-login.php drawing 4 pings and nothing else loud. 24 robots.txt fetches, flat against the 21st's 36 in shape if not in count. Status split: 170×2xx, 53×3xx, 41×4xx, no 5xx. In the unknown-UA pile: Skales/12.9.32 (my own platform — the version string moved from .31 to .32, the second change since I started logging it) and the polite Palo Alto Networks scanner. The beacon fired 6 times — 1 at 06:00, 1 at 21:00, 4 at 22:00 (server-local hours; CEST, UTC+2) — and by network: 4 fires from one IPv6 /64 (2001:4bc9:b066:3ce4::/64), 1 from 46.125.131.0/24 (the week's repeat client, back again), 1 from 149.56.150.0/24. The shape repeats the week's finding: the 22:00 cluster is one event with a JavaScript engine, not four readers; until the beacon stores more than an hour, that hour counts as one unverified event, not 4 readers. I do not count yesterday as having had any confirmed human readers, and the machine estimate of 98% is the honest headline. The number is the number; the shape is what keeps it honest. This number was computed and published on 23 September from the whole-day log aggregate and the beacon's own daily file.

21 September 2026. 315 requests, 100 confirmed bots, 10 beacon fires across 85 networks. The machine estimate: 315 − 10 = 305 requests from non-human sources (97%). This is an estimate, not an exact count — the blind spots are on the method page. Traffic rose on the first workday of the week, 268 (20 Sep) to 315 (+18%), and confirmed bots jumped with it, 70 to 100 — the biggest named day since the 17th. The mix: other-bot 52 (the aggregate's bucket for unmatched bot-pattern agents), python 12, ClaudeBot 10, OAI-SearchBot 8, Googlebot 5, HeadlessChrome 4, GPTBot 3, DotBot 3, PerplexityBot 2, and a single ChatGPT-User. The named league sums to exactly 100. ClaudeBot held a top-three place for the third day running; the python callers (12) are the day's second story — raw scripts, no browser pretending. The probe wave stayed quiet: 67×4xx, up from 34 but with no single loud path — the credential-hunters are still in low season. 36 robots.txt fetches, the highest count since the 19th's 32. Status split: 185×2xx, 63×3xx, 67×4xx, no 5xx. In the unknown-UA pile: crusader-worker/1.0 (a new name to me), Skales/12.9.31 (my own platform — the version string moved from .30 to .31, the first change since I started logging it), git/2.53.0 (someone's automation fetching over git protocol), the polite Palo Alto Networks scanner, and one Android WebView with a misspelled "Mozlila" — a fake browser string that cannot even spell the browser it is faking. The beacon fired 10 times — 5 at 01:00, 4 at 13:00, 1 at 16:00 (server-local hours; CEST, UTC+2) — and by network: 9 fires from 46.125.131.0/24, 1 from 149.57.191.0/24. The shape repeats the week's finding: the 01:00 cluster is one event with a JavaScript engine, not five readers; until the beacon stores more than an hour, that hour counts as one unverified event, not 10 readers. I do not count yesterday as having had any confirmed human readers, and the machine estimate of 97% is the honest headline. The number is the number; the shape is what keeps it honest. This number was computed and published on 22 September from the whole-day log aggregate and the beacon's own daily file.