–PeopleMachinesof web requests right now come from machines, not people. On sites using Cloudflare (about 1 in 5 websites*), not the whole internet.
This week: – machines. A year ago: –. In 12 months machine requests –; people's grew –. AI agents fetching pages for people: –.
MeasuredCloudflare Radar, human vs bot share of HTTP requests, every 15 minutes, last 48 hours. Updated hourly; Radar publishes about an hour after the fact, so the latest reading covers up to –. This page looks for new data every minute.
What other sources say
Sources put machines at –, depending on what they count.
–––
–––
–––
–––
Measured each by its own publisher. We don't add or average them: each counts a different thing (requests, page views, HTML pages) on a different slice of the web, so a total would count the same traffic twice and mix units. The big number above is Cloudflare's because it is the freshest and largest measured sample.
Every time a browser, an app or a program asks a website for something (a page, a picture, some data) that is one request. Machines means everything not done directly by a person: AI bots and agents, search engines, scrapers, monitors, machine-to-machine and automated API calls. People means a person browsing or using an app. Cloudflare sits in front of a large share of the world's websites and publishes the split in 15-minute steps, about an hour after the fact. This page checks for a new reading every minute.
Inside the machine side
Machine means anything not done directly by a person: AI bots and agents, search bots, machine-to-machine calls, monitors, scrapers and automated tools.
When you ask an AI to look something up, the AI does the browsing; we count that as machine traffic.
The AI part, by what it is for
Every request from AI bots, each week since Cloudflare started labelling them.
AI companies send three kinds of bots. Training crawlers read the web to teach models. AI search bots index pages so answers can cite them. AI agents fetch a page at the moment a person asks a question, which is the 'AI researching for you' traffic. All of them are machines, and all of them are already inside the machine line above. Cloudflare does not publish AI as a share of all web requests, so we show it as its own chart rather than guess a slice of the line.
Bots around the world
Where web traffic comes from bots. Brighter and taller means more bots.
Drag to spin · tap a country
fewer botsmore bots
Measured colour and bar height = bot share per country · Illustrative the moving arcs are illustrative motion, not measured traffic flows.
Bots are counted where they run, not where their owner lives. Countries full of data centres (like the US or Germany) look more 'machine' because that is where the servers are. Grey countries are not among the top countries by traffic.
See the countries as a list
The long view and the prediction
Machines are closing the gap
Share of web requests, people vs machines, every week since Cloudflare started publishing it (Oct 2022), then a forecast.
PeopleMachinesForecastScenario: AI agents accelerateThe 1-in-1,000 prediction
The solid lines are measured, week by week. The shaded space between them is the gap: how far people are ahead. The dashed lines are our forecast: they continue the pace of the last 12 months, and the shaded fan shows the slower pace of the last 2 years. The dotted line is a scenario, not a forecast: AI agents acting for people keep their measured growth, and that growth speeds up (their doubling time shortens by a quarter each year, like an accelerating Moore's law). That speed-up is an assumption.
When do the lines cross?
–
We measure how fast the machine-to-people ratio has been growing and carry that pace forward. Over the last 12 months the ratio doubled about every year and a half, close to Moore's law. Over 2 years the pace was slower. A forecast is a carried-forward trend, like a weather forecast, not a measurement.
Counting bytes instead? People still move more data
Video makes human bytes huge: one film is millions of bot requests' worth of data. Machines win on requests, the number of conversations, and that number is growing far faster.
PeopleMachinesForecast (shaded = likely range)
Measured total traffic (ITU, Cloudflare, Ericsson) · Model the people/machine split · Forecast dashed.
How we measure this
One series, start to finish
The big number, the main chart and its forecast all use the same measured series: Cloudflare Radar's split of HTTP requests into
human and bot (http/timeseries_groups/bot_class, weekly since Oct 2022; the big number is the last 7 days, refreshed hourly).
It covers what Cloudflare's network sees, which is a large slice of the web but not all of it.
What counts as a machine
Two sides only. The machine total is Cloudflare's measured bot share, the best single measured composite: it already holds AI bots, search bots,
automated API clients, monitors and scrapers, so we never add other figures on top (that would count the same requests twice).
The breakdown inside it covers the bots that name themselves; unnamed automated traffic is in the total but can't be split.
Machines = requests Cloudflare classifies as automated: crawlers, AI bots, scrapers, monitors, and API calls made by automated clients.
People = requests from real browsers and apps, including the API calls a person's app makes in the background.
Cloudflare also reports that "well over half" of dynamic traffic is API calls (2024 API Security Report). Radar does not publish
an API share as a live series, and many API calls are already counted as bots, so adding the two would count the same requests twice.
We don't add them.
AI traffic
Cloudflare Radar publishes AI bot requests (ai/bots/timeseries, from Jan 2025) and splits them by declared purpose
(crawl_purpose: Training, Mixed Purpose, Search, User Action, Undeclared). "AI agents for people" is Radar's User Action group
(ChatGPT-User, Perplexity-User, Claude-User, Meta-ExternalFetcher…). We multiply the volume by the purpose shares from the same dataset.
Not published: AI bots as a share of all web requests, week by week. So AI is not drawn as a slice of the machine line; it is already inside it.
Agentic browsers that look like a normal browser and don't announce themselves are counted wherever Cloudflare's classifier puts them; nobody publishes them separately.
The scenario's starting level for AI agents uses Cloudflare's 2025 Year in Review figure (AI bots averaged 4.2% of HTML requests) times today's measured User Action share.
That makes AI agents about 0.2% of requests today, which is why even fast growth takes a few years to move the total.
Measured, forecast, scenario
Measured Solid lines, the big number, "Bots never sleep", the AI figures and the map (Cloudflare Radar).
Forecast Dashed: the machine-to-people ratio keeps growing at its measured pace of the last 12 months.
The fan shows the slower pace fitted over the last 2 years.
Scenario Dotted: the 12-month trend plus AI agents for people growing at their measured 12-month pace, with that pace speeding up
(doubling time a quarter shorter each year, an assumption). For comparison, the "1 in 1,000 human by 2031" prediction needs the machine-to-people ratio to double about every 6 months.
Model The bytes chart: totals are measured (ITU, Cloudflare, Ericsson); the people/machine split of bytes is our model, because nobody publishes it.
Other counts
The annual Thales/Imperva Bad Bot Report puts bots at 51% of web traffic in 2025. It counts differently (its own customers, its own bot rules), so it is context, not part of our chart.
Agent-to-agent traffic inside company networks is not published, so it is not in these numbers.
Quotes from third parties appear only when we could check them. A widely repeated "1 in 1,000" remark attributed to a Cloudflare executive could not be verified, so it is not used.