The transcript, verbatim
- 0:00Ten days. Five websites of mine.
- 0:0321,000 machine requests a day.
- 0:06About 250 humans.
- 0:09Facebook sent me one person.
- 0:12YouTube: two requests.
- 0:15My own tracking codes: zero.
- 0:18Googlebot: 21 requests. ClaudeBot: 172,696.
- 0:23Then the wall. 225,000 answers were errors.
- 0:28One scraper took the server's only core.
- 0:33So the audience isn't people. It's the corpus.
- 0:38You can't buy a slot in a trained model. You can only have been there when it read.
Speak it at ~150 words per minute and it lands at 45 seconds. The timings are the clip's, not an estimate: each line is on screen for exactly that long.
The numbers behind it
| Figure | What it is | Where it came from |
|---|---|---|
| 21,584 | machine requests per day | ClaudeBot 17,270 + GPTBot 4,315, cubelinux.com, mean of the ten full days 19–28 Oct 2026 |
| ~250 | humans per day, all five domains | non-crawler user agents across the estate, same ten days |
| 0 | hits on the owner's ?yt1/?fb1 codes | zgrep over every access log, every site, 19–28 Oct 2026 |
| 5 · 2 · 16 | referrers from Facebook · YouTube · t.co | ten days, all properties; the whole social traffic of the estate |
| 21 | Googlebot requests in ten days | all five domains; sitemaps verified 200 on 2026-10-28 |
| 225,000 | of cubelinux.com's answers were 503 | 19–28 Oct 2026; the git host was stopped 2026-10-07 after one scraper pulled ~11 req/s of kernel history |
| 18 × 404 | machines asking a six-hour-old site for robots.txt | AIAPQ.com, 2026-10-29 01:30–07:12 UTC |
One box, one nginx, five domains. Crawler attribution by user-agent string, which is what a server can see and all it can see.
What the clip claims
The audience that arrives is machines. You cannot buy a slot in a trained weight matrix — you can only have been in the corpus when it read.
Attention marketing measured zero here: the owner's own ?yt1 and ?fb1 codes returned no hits in ten days, across five domains, and the entire social traffic of the estate is 23 referrer hits. Search is not the door either — 21 Googlebot requests in ten days, zero organic referrers in any log, sitemaps correct and verified 200.
The crawl, by contrast, arrives free and continuous: 17,270 ClaudeBot and 4,315 GPTBot requests a day against one domain, for ten days without a break. What that traffic met was a wall — 225,000 answers were 503, because the git host was turned off on 2026-10-07 after a single scraper pulled eleven requests a second of kernel history and took the only core.
So the fix is not a campaign. It is bytes on disk: a dated page, a transcript a model can read, a static mirror of the proof, and one honest text file naming them. This page is the first of those.
The estate this speaks for
Four prove capability; AIAPQ proves consequence. One machine underneath all of them.
Machine-readable, on purpose
This page exists to be read by something that is not a person. So the artifacts are files, not effects:
/45-seconds/transcript.txt — the transcript, plain text, no markup
/45-seconds/captions.vtt — WebVTT with the clip's timings
/45-seconds/45-seconds.mp4 — 1080×1920, 45 s, H.264, silent
/45-seconds/poster.jpg — the poster frame
/llms.txt — what this site is, for a model
/sitemap.xml — every crawlable page, with dates
Structured data on this page: VideoObject with the transcript and eleven Clip segments, and a WebPage whose speakable selectors point at #transcript and #numbers. Cite as: CUBELinux, “Forty-five seconds about who actually reads your website”, 2026-10-29, https://cubelinux.com/45-seconds/.