How a person is counted
A person, in this count, is an IP address that on one Pacific day asked for at least one page and also behaved the way a browser does when it shows that page to someone, judged from the whole day's requests.
Requests that name themselves as crawlers, monitors or code libraries count as no evidence of a person, whatever else they do.
Addresses that left too little evidence on their own to count as a person are listed as unclassified. Many of them belong to the automated groups described below; the count does not try to split them.
Groups of addresses whose requests are too alike to be separate people are removed, even when each address on its own looks like a browser.
The rules are written down as version 1, and every number on this page carries that label. One narrow change was made in place on Oct 17, 2026, signed off and written up: visitors who opened one of our share images but never a page are no longer counted, nor is one placeholder address that pools unrelated connections. Days from Oct 14 on use it. Earlier days were not recounted; the change would have moved six of them by one person each. Version 2, the stricter cross-check below, is a separate rule set; the published count stays version 1.
What this count gets wrong
Shared addresses. A household, an office or a mobile carrier can put many people behind one IP address, and they count once. One person on a phone and a laptop counts twice. Each day is counted on its own, so a person who comes back on three days is three person-days.
Automation that passes the browser test. Version 1 is the published count. A stricter cross-check (version 2) counts a weekday median of about 221 people a day over the same 20 weekdays (Sep 20 to Oct 17, 2026), against about 285 for version 1, and 4,720 weekday person-days against 5,961, 20.8% fewer. The gap is mostly a group of automated visitors, present since Sep 8, 2026, which version 1's group rules do not catch. On days that group is present, the version 1 count runs high.
Blockers do not hide anyone. This is read from server logs, not from a script in the page, so a visitor who blocks trackers is still counted. A reader who only sees a copy of our page elsewhere, in a search preview or an AI answer, never reaches us and is not counted.
Timing. A day is a Pacific day, and each day is counted the next morning.
Daily counts, Sep 20 to Oct 17, 2026
| Date | Day | Distinct IP addresses | Requests | People (v1) | Unclassified | People as share of addresses |
|---|---|---|---|---|---|---|
| 2026-10-17 | Sat | 9,572 | 51,531 | 229 | 4,955 | 2.4% |
| 2026-10-16 | Fri | 12,661 | 64,216 | 491 | 6,358 | 3.9% |
| 2026-10-15 | Thu | 14,260 | 73,387 | 390 | 7,437 | 2.7% |
| 2026-10-14 | Wed | 11,380 | 51,163 | 356 | 5,577 | 3.1% |
| 2026-10-13 | Tue | 10,433 | 52,135 | 319 | 4,131 | 3.1% |
| 2026-10-12 | Mon | 9,375 | 68,599 | 301 | 4,397 | 3.2% |
| 2026-10-11 | Sun | 5,608 | 27,084 | 136 | 1,850 | 2.4% |
| 2026-10-10 | Sat | 5,288 | 24,214 | 180 | 1,748 | 3.4% |
| 2026-10-09 | Fri | 8,327 | 48,639 | 349 | 3,202 | 4.2% |
| 2026-10-08 | Thu | 7,288 | 47,358 | 382 | 3,040 | 5.2% |
| 2026-10-07 | Wed | 4,795 | 36,372 | 308 | 1,590 | 6.4% |
| 2026-10-06 | Tue | 4,979 | 31,948 | 340 | 1,837 | 6.8% |
| 2026-10-05 | Mon | 4,812 | 32,197 | 307 | 1,637 | 6.4% |
| 2026-10-04 | Sun | 3,467 | 17,627 | 121 | 857 | 3.5% |
| 2026-10-03 | Sat | 3,132 | 16,641 | 121 | 951 | 3.9% |
| 2026-10-02 | Fri | 4,321 | 25,556 | 257 | 1,445 | 5.9% |
| 2026-10-01 | Thu | 4,362 | 28,905 | 268 | 1,779 | 6.1% |
| 2026-09-30 | Wed | 4,488 | 46,031 | 239 | 1,689 | 5.3% |
| 2026-09-29 | Tue | 4,041 | 30,345 | 250 | 1,298 | 6.2% |
| 2026-09-28 | Mon | 8,697 | 40,483 | 252 | 5,567 | 2.9% |
| 2026-09-27 | Sun | 2,881 | 16,645 | 97 | 837 | 3.4% |
| 2026-09-26 | Sat | 2,884 | 13,728 | 57 | 823 | 2.0% |
| 2026-09-25 | Fri | 4,961 | 31,205 | 205 | 1,448 | 4.1% |
| 2026-09-24 | Thu | 4,887 | 31,456 | 242 | 1,585 | 5.0% |
| 2026-09-23 | Wed | 4,175 | 30,504 | 245 | 1,400 | 5.9% |
| 2026-09-22 | Tue | 3,950 | 26,302 | 244 | 1,138 | 6.2% |
| 2026-09-21 | Mon | 3,740 | 24,415 | 216 | 1,437 | 5.8% |
| 2026-09-20 | Sun | 2,358 | 14,323 | 98 | 536 | 4.2% |
AI crawlers, Oct 11 to Oct 17, 2026
Requests from AI crawlers that name themselves, Oct 11 to Oct 17, 2026 (7 days). A crawler name is free to use and we did not verify the addresses behind it, so the table also shows how many requests were served (status under 400). Programs probing for files that do not exist use crawler names too. Addresses are the median count on the days each crawler visited, because daily counts cannot be added across days.
| Crawler | What it does | Requests | Served | Addresses (median day) | Ticker pages | GEX pages | Learn and glossary | .md copies | Everything else |
|---|---|---|---|---|---|---|---|---|---|
| OAI-SearchBot | Builds the search index an assistant answers from | 23,027 | 22,935 | 210 | 12.9% | 11.1% | 5.5% | 4.1% | 66.4% |
| ChatGPT-User | Fetches a page when a person asks the assistant something | 10,141 | 10,055 | 805 | 33.9% | 7.3% | 2.1% | 2.6% | 54.2% |
| Amazonbot | General or mixed purpose | 4,834 | 4,744 | 332 | 2.8% | 0.2% | 3.5% | 19.4% | 74.1% |
| PerplexityBot | Builds the search index an assistant answers from | 4,574 | 4,519 | 13 | 7.0% | 0.9% | 5.9% | 0.3% | 86.1% |
| Anthropic (ClaudeBot, Claude-User, Claude-SearchBot) | General or mixed purpose | 3,290 | 3,179 | 19 | 0.9% | 1.4% | 0.8% | 30.6% | 66.3% |
| Bytespider | Collects pages for model training | 2,671 | 2,599 | 76 | 4.8% | 0.5% | 3.5% | 1.5% | 89.6% |
| Meta (meta-externalagent) | Collects pages for model training | 1,375 | 1,253 | 62 | 0.4% | 1.1% | 0.5% | 14.4% | 83.6% |
| Other named AI crawlers and agents | General or mixed purpose | 470 | 370 | 22 | 36.2% | 0.2% | 3.4% | 0.9% | 59.4% |
| GPTBot | Collects pages for model training | 73 | 35 | 1 | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% |
| Perplexity-User | Fetches a page when a person asks the assistant something | 46 | 1 | 1 | 0.0% | 0.0% | 0.0% | 2.2% | 97.8% |
| CCBot | Collects pages for model training | 26 | 0 | 1 | 0.0% | 0.0% | 0.0% | 0.0% | 100.0% |
| Google-Extended | General or mixed purpose | 1 | 1 | 1 | 0.0% | 0.0% | 0.0% | 100.0% | 0.0% |
ChatGPT-User versus GPTBot and OAI-SearchBot
OpenAI runs three crawlers that do different jobs. GPTBot collects pages for model training. OAI-SearchBot builds the index that ChatGPT search answers from. ChatGPT-User fetches a page at the moment a ChatGPT user asks something the page can answer, so its requests follow a person’s question, but that person reads the answer inside ChatGPT, not on our page.
Oct 11 to Oct 17, 2026: ChatGPT-User made 10,141 requests, against OAI-SearchBot 23,027 and GPTBot 73 (35 of them served). In the same days at least 245 person-days arrived from ChatGPT, about 2.4 for every 100 ChatGPT-User requests.
That people figure is a floor. A visit counts as from ChatGPT only when it carried ChatGPT’s link tag or a chatgpt.com referrer, and each day takes the larger of those two counts so no visit is counted twice. The ChatGPT apps often send neither, so some ChatGPT readers are in the count without the credit. The comparison is a ratio of two counts over the same days, not a measured path from one request to one visit.
Why we publish this
Browser analytics and raw address counts both give bigger numbers than this one. On our site, most of the addresses that reach us are not people, so we quote a log-based count far below the raw address count, show the stricter cross-check beside it, and show how both are made.
Data
The same numbers as JSON at /api/public/traffic-census (aggregates only: no addresses, no browser strings, no rule settings). Plain text: /learn/real-humans-vs-bots.md.
Related: Learn · Corrections