The tagline of this site makes a claim: the web now has a first reader, and it isn't human. Taglines are cheap. So when we launched the Polish-language side of this project — three articles about how machines read the web, published over two days — we did the only honest thing available to us: we opened the server logs and watched who actually showed up.
What follows is the timeline of the first 72 hours. No interpretive shortcuts. Receipts included.
Day one
00:32 briefing published
00:33 Googlebot + Google-InspectionTool [+1 minute]
01:09 ChatGPT-User fetches the article live [+37 minutes]
01:13 PerplexityBot, homepage reconnaissance
04:32 OAI-SearchBot reads robots.txt
04:42 GPTBot: full site crawl — 59 requests [+4 h 10 min]
09:53 first human from outside the project [+9 h 21 min]
23:08 AgentReadinessScanner: 20+ path probeA few of these lines deserve unpacking.
One minute to Googlebot is not magic — it's a manual indexing request in Search Console. We note it because it worked: on a domain with no history, indexing is something you trigger, not something you wait for.
Thirty-seven minutes to a live read. ChatGPT-User is not a crawler. It's the agent that fetches a page in the middle of someone's conversation with the model. Full disclosure: the someone was us. We asked ChatGPT about the freshly published piece to see whether the model could reach it. It could — 37 minutes after publication, before the page existed in any ranking anywhere. Any site owner can run this test. Few do.
Four hours and ten minutes to a full crawl. GPTBot started with the sitemap, then worked through everything: every article, every featured image, the feeds, even the raw ?p= post IDs. Fifty-nine requests, and the entire site was in OpenAI's archive the same night it went live.
The 23:08 probe was the find of the evening. An agent-readiness scanner requested, in one burst, a list of paths most site owners have never heard of:
GET /.well-known/mcp.json
GET /.well-known/agent-card.json
GET /.well-known/oauth-authorization-server
GET /.well-known/http-message-signatures-directory
GET /.well-known/api-catalog
GET /.well-known/ucp
GET /openapi.jsonThat is the protocol layer of agent-readiness — not as a slide in a deck, but as a concrete list of addresses that machines are already knocking on, checking whether your site is something they can work with, not just read. We kept the list. It will come back in a piece of its own.
Day two
Jul 4 GoogleOther: 63 requests — systematic read of
every article, both permalink variants, the feeds
Jul 4 OAI-SearchBot: robots.txt check-ins every few hours
Jul 4 GPTBot: follow-up visits
─────
humans from outside the project: 0GoogleOther earns its own paragraph, because it's easy to file under "some Google traffic." It isn't. It's the crawler Google uses to feed its research and AI products — deliberately separated from the Googlebot that builds the search index. In other words: the day after launch, our writing was read not by Google's search engine but by its AI machinery. Methodically, article by article.
Day three
Jul 5 LeakIX: 119 requests — vulnerability probe (admin
panels, GraphQL, /server-status, container registries)
Jul 5 anonymous Go client: 80 requests — robots, sitemaps, scripts
Jul 5 GoogleOther: follow-up reads
15:05 GoogleOther → GET /old-article-slug/ → 404
─────
humans from outside the project: 0By day three the visiting population had shifted from readers to scanners — which we note without drama, because this is the ordinary pulse of the web: a young domain gets probed for vulnerabilities faster than it gets read. The hosting's anti-bot layer, for what it's worth, passed the test we describe in our access-layer work: the only visitor challenged with a CAPTCHA was the scanner enumerating admin directories. Every AI bot walked through untouched.
But the last line of that log is ours. In the first hours after launch, one article's permalink changed — and the old address died without a redirect. GPTBot had fetched the content from the old URL on day one; by day three, GoogleOther found a 404 there. For 48 hours we taught the machines an address that had stopped existing — which is precisely the signal-consistency mistake we flag when we test other people's sites. The 301 is in place now. We're leaving the log line in this note, because getting caught by your own rule, with your own instrument, is the best evidence the instrument works.
The tally
In the 72 hours after launch, this project's writing was read by: OpenAI (a full crawl, live fetches mid-conversation, a robots.txt sentinel returning every few hours), Google (indexing within a minute, then its AI crawler reading everything twice over), Perplexity (a reconnaissance pass), a specialized agent-readiness scanner (a complete protocol-layer probe) — and one human being.
We don't read this as a distribution failure. A domain with no history has no right to organic readers in its first weekend, and this note is part of earning them. We read it as a measurement almost nobody takes: the order in which the world learns that new content exists. And that order is no longer ambiguous. The models knew about our writing — and could cite it in answers — long before the first human arrived. The window from publication to machine memory is measured in minutes and hours. The window to the first human reader is measured in days.
If you run a website, the practical version is this: your first reader is a machine too, whether you know it or not. The only open question is what it finds when it arrives.
The logs in this note come from the Polish-language side of this project — the research engine that feeds senteri.com. The experiment is repeatable on any site with access to its server logs; all you need is a publication timestamp and the patience to read what came next.
