Live data

The web is being read by agents

59 crawlers from 23 operators, counted and named. This is what they are, which of them can be proven, and what they came to do.

Agent requestsin October 2026
Human visitsin October 2026
%
Agent sharein October 2026
agents humans
JanFebMarAprMayJunJulAugSepOct

Since we started counting

agent requests human visits

Every request recorded across every site on TrueStat, counted server-side. No site, page or visitor is identifiable in these figures.

The agent takeover

Human readership holds steady while agent readership compounds

The same pages, the same months. One line of readers barely moves; the other multiplies.

x
Agent growth this year

Jan to Oct, across every site.

x
Human growth, same span

The comparison that makes the first number mean something.

%
Agent share, latest month

Of every request recorded on the platform.

The week

People stop reading at the weekend. Agents do not.

The same sites, the same seven days. One audience keeps office hours; the other runs on a schedule that has none.

agents humanseach scaled to its own peak
Mon
Tue
Wed
Thu
Fri
Sat
Sun

Where from

Humans arrive from where they live. Agents arrive from datacentres.

The busiest countries on the platform, counted separately for each audience. The two rarely rank the same, and that difference is the reason a single visitor map cannot describe both.

agents humans
  • 🇺🇸US8821,050
  • 🇨🇳CN71165
  • 🇩🇪DE48284
  • 🇹🇷TR4110,606
  • 🇸🇬SG34452
  • 🇨🇦CA467
  • 🇫🇷FR1218
  • 🇮🇪IE125
  • 🇮🇳IN01,345
  • 🇵🇰PK0389
  • 🇳🇱NL0179
  • 🇬🇧GB0161
  • 🇮🇹IT0124
  • 🇧🇷BR0111
  • 🇷🇴RO093
  • 🇪🇸ES085
  • 🇮🇱IL071
  • 🇵🇱PL063
  • 🇵🇹PT040
  • 🇨🇭CH036
and 4 more countries0126

Ranked by agent requests, the longest bar being 882. The bar's fill is the split between the two. Countries fewer than three sites have traffic from are not published.

Who is reading

Not one crowd. Companies with names.

Every operator the platform has actually recorded, with the share of agent requests it accounts for. We name 23 in total and show the ones that have turned up. “AI traffic” is a queue of identifiable companies, and each one is a separate decision.

OperatorRequestsShare
  • OpenAIprovable11,45126.6%
  • Metaclaimed11,34426.4%
  • Amazonclaimed5,75213.4%
  • Anthropicprovable5,15212.0%
  • Googleprovable3,7318.7%
  • Microsoft Bingprovable2,9306.8%
  • Perplexityprovable1,0792.5%
  • Appleprovable6021.4%
  • DuckDuckGoprovable4201.0%
  • Baiduprovable3600.8%
  • Yandexprovable1620.4%

What they came for

“AI traffic” is four jobs, and four robots.txt decisions

Answering a question being asked right now is not the same as collecting a training corpus. A single number for all of them tells you nothing you can act on.

%
Answering a live question

requests

%
Refreshing a search index

requests

%
Collecting training data

requests

%
A reader an assistant sent

requests

A category at zero is drawn at zero. It means the panel sorts for it and no site has recorded it yet — not that it was left out.

Who can be proven

56% of agent requests could be proven. We say which.

An operator that publishes address ranges or reverse DNS can be verified by anyone. One that publishes neither cannot be — by us or by anyone claiming otherwise. Both are counted; only one is presented as confirmed.

Provable

13 operators

Publishes address ranges, reverse DNS, or both — so a claim can be checked.

  • Google7
  • OpenAI5
  • Anthropic4
  • Microsoft Bing3
  • Mistral3
  • Moonshot AI3
  • Baidu2
  • DuckDuckGo2
  • Huawei2
  • Perplexity2
  • Apple1
  • Common Crawl1
  • Yandex1

Takes its word

10 operators

Publishes neither. The visit is counted and labelled a claim, never dropped and never presented as confirmed.

  • xAI5
  • Meta4
  • Alibaba3
  • Amazon3
  • ByteDance3
  • Allen AI1
  • Cohere1
  • DeepSeek1
  • You.com1
  • Zhipu AI1

Methodology

Aggregated and anonymised across every site on TrueStat

How these figures are produced, so you can judge them.

Counted server-side

An agent never runs JavaScript, so a browser tag records none of this. Every figure here comes from the 59 crawler patterns the detector ships with, matched on the server.

No site is identifiable

Nothing published here carries a project, a hostname, a path or a visitor. Any bot family fewer than five sites have seen is dropped entirely, because an aggregate over one customer is that customer's data.

Verification is a claim we can check

13 of 23 operators publish address ranges or reverse DNS. Where an operator publishes neither, the visit is counted and labelled a claim — never dropped, never shown as confirmed.

All of it, all the time

The chart shows the months of the current year; the total beneath it is every request since we started counting. Nothing is sampled and nothing is extrapolated.

See every crawler we name, and how each one is verified →

See this for your own site

One script tag, and the agent panel from the first request. 14 days, no card.