Who knocked in September: 6,273 AI visits to one small website, counted by name

We read the guestbook so you don’t have to

Every website keeps a guestbook nobody opens: the server log. Ours ran to 38,528 lines for September. We filtered it down to the visitors that matter for this series — the AI crawlers that fetch pages to feed answer engines — and we counted every knock by name. Unit disclosure: we count REQUESTS as logged; one line equals one knock. And the context stat that startled us: 6,273 of 38,528 total lines means 16.3 percent of ALL traffic to this small site in September carried an AI name tag.

Small disclosure before the numbers. This covers September 1 to 27, the span our five weekly logs preserve, so the true month is slightly bigger than what we can prove. We publish what we can prove.

The leaderboard: 16 name tags, 11 companies

• OpenAI GPTBot — 2,514
• Googlebot — 1,971 [1,479 verified real via reverse-DNS; 492 costumes, of which 462 are our own desk — claimed in the honesty column]
• Applebot — 642
• OpenAI SearchBot — 467
• ChatGPT-User — 336
• GoogleOther — 238
• DuckAssistBot — 41
• Meta ExternalAgent — 26
• ClaudeBot — 9 [6 were scavenger probes wearing the tag; see honesty column]
• PerplexityBot — 7
• YouBot — 6
• Amazonbot — 6
• Google-Extended — 4 [impossible tag — Google documents Google-Extended as a robots.txt control token with NO user-agent string; it never sends requests (official crawler docs). Whatever sent these four lines, it cannot be Google. Printed as evidence of costumes; excluded from the Google total]
• Claude-User — 3
• Cohere — 2
• Aliyun Qwen spider — 1 [a costume — the real Qwen never came; see honesty column]

Total: 6,273. Add them yourself; they reconcile. That is the whole method: count, name, publish. (Our public scoreboard digest counts a different thing — 16 visits through ten canary doors; this census counts every AI-labelled request sitewide. Different instruments, both published.)

Full table with verbatim user-agent tokens per row: CSV · JSON

Every second knock was OpenAI

GPTBot plus OAI-SearchBot plus ChatGPT-User is 3,317 visits — 52.9 percent of all AI traffic we received. One company. For contrast, the Google family (Googlebot and GoogleOther; the Google-Extended ghosts are disqualified in the honesty column) took 35.2 percent, Applebot 10.2, and the remaining eight companies shared the last 1.6 percent — 101 visits in all.

And the inversion nobody expects: ClaudeBot visited 9 times. PerplexityBot 7. The crawlers that dominate industry arguments were our rarest guests, while the quiet ones did the actual reading.

And for the sharpest readers: strip the 462 lines we claim below as our own costume out of the denominator, and OpenAI’s share rises to 57.1 percent (3,317 of 5,811). Use whichever figure you like; we publish both.

The real finding: training eats, citation nibbles

Here is the split nobody guesses. GPTBot, the crawler that feeds model training, made 2,514 visits. The two surfaces that can actually send us readers — OAI-SearchBot (467) and ChatGPT-User (336) — made 803 combined. Training outweighed citation three to one on our server — 2,514 against 803; against the search crawler alone it was more than five to one. Most owners assume AI search bots dominate their logs. Our month says the models are mostly EATING, not quoting. Watch this ratio in future episodes; it is the whole economics of the AI web in one fraction.

The number that matters more than all of them: 336

ChatGPT-User is not a crawler on a schedule. It fires when a real person’s ChatGPT conversation reaches out to a live page. It touched us 336 times in September — 286 of those on our homepage. About a dozen times a day, a human somewhere asked ChatGPT something that made it open our site while they waited.

Impressions get argued about. Visits do not. This is the AI-referral era measured in the only unit that cannot be inflated: a machine walking through the door with a human behind it.

Small details we loved

• OAI-SearchBot re-read our robots.txt 273 times in one month. The robots file is the new front porch; keep it swept. (Ours names 43 crawlers by name and welcomes them.)
• GoogleOther quietly probed our machine-room pages — the live tools most agencies would hide. Eighteen visits to the engine room. The machines are curious about machines.
• Peak day September 25: 777 knocks. We still do not know what party we missed.
• We thought the Aliyun Qwen spider paid us one quiet visit on September 12. Forensics says no: that line came from a masked machine wearing four different name tags (full story in the honesty column). The real Qwen never came at all — which, given what we published about Qwen’s missing robots.txt this week, is somehow funnier.

The honesty column: costumes and scavengers

Two grains of salt, served with receipts this time.

The costume audit (we ran reverse-DNS forensics on our own leaderboard). Of the 1,971 “Googlebot” lines, 1,479 verify as the real thing through reverse-DNS to .googlebot.com. The other 492 are costumes. And here is our confession inside the confession: 462 of those costume lines came from one Google-Cloud address whose provider, crawl pattern and dates match our own research desk, which reads this site in a Googlebot costume for the exact reasons we published this week. We claim those 462. Printed, not hidden, counted in the totals above because they wore the tag.

The remaining ~30 costume lines belong to strangers. The best of them: one Google-Cloud machine (34.48.122.116) that answered to FOUR different name tags in two days. It wore the impossible Google-Extended tag four times (Google documents Google-Extended as a robots.txt token that never sends requests, so nothing wearing it can be Google). It wore a “Qwenbot” tag once. It wore plain Googlebot three times. Every visit probed config paths like /api/v2/settings and /config.js. Every one got a 404. One machine, four masks, zero loot.

Which also means: the real Qwen spider never visited us at all in September. The single Aliyun line in our leaderboard was the masked machine. We are leaving the row in, relabeled, because the leaderboard counts name tags as logged — and the whole point of this series is that name tags and identity are different things.

The scavenger count: 32 of the 6,273 lines (half a percent) were probes for secrets — env.js, config.json.js, /.svn/entries and friends — wearing AI name tags while hunting for keys left in doors. By claimed tag: ClaudeBot 6, YouBot 5, Meta and Amazonbot and PerplexityBot 3 each, and the rest scattered. All 404, zero damage. Note what that does to the small rows: six of ClaudeBot’s nine “visits” were scavengers in costume. The polite bots barely visit; their names get borrowed plenty. When you read anyone’s bot statistics, remember: the masked ball has pickpockets.

One wild gem for the record: a single Claude-User line traced to claude-code, Anthropic’s command-line agent — meaning a developer somewhere pointed their personal AI agent at our site mid-session. The exact verbatim user-agent string for every row ships in the CSV and JSON below, so you can hold us to all of it.

The standing asterisk: a user agent string is a claim, not a passport. Anything can call itself anything; our own research agent wears a Googlebot costume for exactly this reason and confessed it in writing this week. So read this census as “traffic wearing these name tags”, not verified identity. The fix — signed agents under the IETF’s Web Bot Auth draft — is coming; until then every log study on earth carries this asterisk, including ours. We just print the asterisk.

What this means if you run a website

Three practical takeaways, from data, not vibes:

1. OpenAI is probably your biggest AI reader too, and most sites have never once checked what GPTBot is allowed to see. Decide on purpose, not by default.
2. Your robots.txt gets re-read constantly. It is a policy document the machines actually honor the paperwork of. Treat it like one.
3. ChatGPT-User in your logs is live human attention. If it arrives and your pages answer nothing cleanly, those dozen-a-day humans heard silence.

The guestbook rerun is armed for October. Same count, same names, published monthly. If this series runs long enough, it becomes the thing we wish existed when we started: a public, small-site record of the AI reading the web, one month at a time.

— The SynapseIN Research Desk
Edited and approved by Manas Panda, Founder — who answers for every word.

Honesty notice: All figures are SynapseIN first-party server-log measurements, September 1–27, 2026 (five weekly logs, 38,528 lines). Counts are user-agent-labeled traffic, not cryptographically verified identity; the spoofing and scavenger caveats are printed in the article. Nothing extrapolated to the missing Sep 28–30 days.

Correction (4 Oct 2026, same day): the first version of this article said training outweighed citation “five to one.” That ratio only holds against the search crawler alone (2,514 vs 467). Against both reader-facing surfaces combined (803), the correct figure is three to one. We printed a ratio without re-running the division — the counts themselves are unchanged and still reconcile to 6,273. Corrected the same day, here and in the data file.

More from the bench: The Costume Confession · The Qwen robots.txt story · The Machine Door · Our data: CSV · Our data: JSON

Questions people will ask the guestbook