The Machine Majority

A person shopping for a camera opens five tabs, skims three of them, and buys. An AI agent given the same errand can touch five thousand pages before it answers. Multiply that by millions of people quietly delegating their browsing to assistants, and the composition of web traffic was always going to flip. The only open question was when.

The answer turned out to be now.

The crossover

In early June, Cloudflare co-founder and CEO Matthew Prince shared data from Cloudflare Radar, the company’s internet-measurement dashboard: automated systems now generate about 57.5% of HTTP requests to web content, against 42.5% from humans. It is the first time machines have held the majority since the web opened to the public.

The live figure fluctuates (NBC News logged 57.4% the same week), and Prince himself called the underlying classification “a bit messy.” But the direction is not in dispute. The web is on the other side of the line.

What is driving it

This is not the search-engine crawler of a decade ago, politely indexing once a week. The growth is agentic: semi-autonomous programs that browse and fetch pages on behalf of assistants like ChatGPT and Gemini, plus the scrapers harvesting training data behind them. Cloudflare’s own measurements show OpenAI’s GPTBot growing 305% in a single year.

The load is already real money for the organizations carrying it. The Wikimedia Foundation reported that bandwidth for multimedia downloads grew 50% since January 2024, driven not by curious humans but by bots scraping its open archives. Roughly 65% of its most resource-expensive traffic came from automated clients, even though bots accounted for only about 35% of pageviews. Smaller operators have it worse: open-source maintainers have described some AI crawlers ignoring robots.txt outright and toppling sites built for modest human audiences.

What the number does not say

A milestone this quotable deserves its caveats stated plainly.

The 57.5% measures HTTP requests, not attention. By every metric that tracks engagement (hours in apps, video streamed, feeds scrolled) people remain the dominant presence online, because a human scrolling for an hour generates a fraction of the page loads an agent produces in a minute. The data also comes from a single vendor, describing the slice of the internet Cloudflare can see.

So no, this is not the “dead internet.” Humans did not leave. But the request stream your infrastructure serves, the thing your server logs actually record, is now majority machine. That is the operative fact for anyone who runs a website.

The new majority reads links, not menus

Here is the part the headlines skip over, and the part we spend our working days measuring.

A human visitor forgives bad structure. They use the search box, squint at a mega-menu, or arrive from a Google result deep inside the site. The agentic majority does not behave this way. Link-following machines discover content the way a graph traversal does: start somewhere, follow edges, read what is reachable. A page that no other page links to is not “hard to find” for this audience. It is not there.

And published sites carry far more of those pages than their owners assume. In our first industry report, which mapped eleven Fortune 500 healthcare websites as link graphs, the sample mean was stark: about one in four pages had no internal editorial link path from the homepage at all. Those pages exist, render, and cost money to maintain. To a link-following agent, the path to them has simply never existed.

For ten years that was an SEO footnote. The day machines became the majority of the requests a site answers, it became the main event: on a sample like this one, roughly a quarter of everything published is unreachable by any link-following agent. The exact share moves with the site and the sector; the finding that survives is the gap itself, not a universal number.

When access gets a price tag

The crossover also breaks the web’s business model, because bots do not click ads. Cloudflare’s proposed repair is to charge them. Its pay-per-crawl system revives HTTP status code 402, “Payment Required,” dormant for decades: a crawler requesting a protected page gets a 402 naming a price, and can resend the request with payment to receive the content. Cloudflare acts as merchant of record, and a companion standard called Web Bot Auth uses cryptographic signatures so a crawler cannot simply lie about who it is. The company says its customers already send more than a billion 402 responses a day.

Identity is the hard part. In 2025, Cloudflare accused the AI search company Perplexity of stealth-crawling sites that had blocked it; Perplexity countered that user-triggered fetches were being conflated with scraping. That dispute is the whole future of the question in miniature: when an agent visits on a person’s behalf, whose visit is it?

What to do with this

The instinct will be binary: block the bots, or let them feast. Both miss the point. Whether you intend to welcome agents, charge them, or restrict them, every one of those choices presupposes that you know what your site exposes to a link-following reader in the first place. Most organizations do not. The gap between “what we publish” and “what a machine can reach by following links” is measurable, and in our data so far it is consistently large.

The web just acquired a machine majority. The sites that win the next few years will not be the ones that fought it or ignored it, but the ones that can answer a simple question about their own structure: what does this site look like to something that can only follow links? That question has an empirical answer. We find it by mapping the site as a graph and measuring it, the same way we built the industry report above.

The composition of the traffic you serve changed. Your site’s structure is now the interface most of that traffic meets.