Automated requests now outnumber human ones on the web, but much of that is ordinary infrastructure and people are still here; what deserves worry is manufactured consensus, untrustworthy engagement and fraud.
- It counts requests, not people. The automated share measures requests to web servers, and a large part of it is crawlers, monitors and other plumbing.
- Agents are the new bots. Traffic from AI agents is growing fast from a small base, and it fits neither side of the human-or-bot split.
- About half of new articles are machine-written, and that share has levelled off. Those articles reach top search results far less often than their volume suggests.
- A famous forecast failed. The widely repeated prediction of a mostly synthetic web by 2026 did not come true, and the record keeps it for that reason.
- Worry about deception, not volume. Treat engagement counts as cheap, trace repeated claims back to their source, and watch for fraud.
Update — August 2026: the bots are agents now
A 2026 benchmark report found automated traffic growing about seven and a half times faster than human traffic year over year (23.51% against 3.10%), and within that, agentic traffic up 7,851%.HUMAN Security, 2026 State of AI Traffic & Cyberthreat Benchmark Report, 26 Mar 2026, read at source 9 Sep 2026: “traffic from AI agents and agentic browsers grew 7,851% year over year”, and “automated traffic across the internet grew 23.51% year over year, while human traffic increased 3.10%”.
That percentage is enormous because the baseline was near zero, and it should be read that way. The direction is the finding, not the multiple. What was scraping and spam is increasingly software acting on someone's behalf — booking, buying, filling forms, reading pages in order to do something rather than to index them.
The traffic counts on this page split the web into human and bot. An agent acting on a person's instruction fits neither cleanly. It is automated traffic with a human intent behind it.
Expect the headline percentages to keep rising while meaning less, until the measurement catches up with the category.
Imperva / Thales, 2026 Bad Bot Report, published April 2026, read at source 23 Sep 2026: “automated traffic continues to outpace human activity online, accounting for more than 53% of all web traffic in 2025, up from 51% the year before”.
That is the honest headline number. The rest of this page is the part that gets left out of it — because "half the internet is bots" is true, widely repeated, and routinely misunderstood.
Read agent-traffic growth for its direction, not its size: it starts from a small base, and agents blur the human-or-bot split these counts rely on.
What the 53% actually counts
The figure measures requests, not people and not content. Every time something asks a web server for a page, that request is classified as human or automated. In 2025, more than half were automated — up from 51% in 2024, with human activity falling to roughly 47%. How much of the automated share is malicious is given in the full report, which sits behind a contact form and was not read at source here.Source: Imperva (a Thales company), 2026 Bad Bot Report, read at source 9 Sep 2026: “automated traffic continues to outpace human activity online, accounting for more than 53% of all web traffic in 2025, up from 51% the year before” — thirteenth annual edition, published Apr 2026, based on full-year 2025 telemetry across the Imperva global network. Vendor telemetry, not a universal census of every website. Re-read 17 Sep 2026.
Earlier editions of the same report put the automated share lower. For 2022: “nearly half (47.4%) of all internet traffic came from bots”. For 2019 the report split it into bad bots at 24.1% and good bots at 13.1% — about 37% together. The method changes between editions, so read the direction, not the decimals.Imperva, 2023 Bad Bot Report: key insights and Bad Bot Report 2020, both read at source 17 Sep 2026: “In 2019, bad bot traffic comprised 24.1% of all website traffic” and “Good bot traffic consisted of 13.1% of traffic”. The 37% is this page’s sum of the two.
"Bot" is not a synonym for "fake." A large share of that automated traffic is ordinary infrastructure: search engine crawlers, uptime monitors, accessibility tools, feed readers, security scanners, and the AI crawlers that index the web — including, quite possibly, the one that brought you this page. The internet has been substantially automated since long before generative AI, because automation is how the web is indexed, monitored, and kept running.
The line worth drawing is not human versus machine. It is automation that serves people — the crawler that lets you find this page — versus automation that deceives or attacks them: credential stuffing, scraping for resale, fake engagement, fraud. The second category is the problem, and it is only part of that 53%.Reasoning, September 2026 — the 53% counts all automated requests, useful and hostile alike; the malicious share is in the full Imperva report, which sits behind a contact form and was not read here.
Separate automation that serves people from automation that deceives or attacks them. Only the second is the problem, and it is much smaller than the headline share.
How much of what you read is machine-written
A different question, and a different measurement. Research firm Graphite sampled English-language articles from Common Crawl and classified them with detection tools. Their latest study finds primarily AI-generated articles reaching about half of new articles by early 2025 and hovering around half ever since — 50.9% in Q4 2025, 49.9% in Q1 2026. The plateau is the robust finding, not any single decimal; Graphite's own earlier study, with a different detector, ran 3.3 points higher.Source: Graphite, “AI Now Writes as Many Online Articles as Humans”, May 2026, read at source 17 Sep 2026: “However, since Q1 2025 the percentage of primarily AI-generated articles has plateaued at roughly 50%.” and “In Q4 2025, primarily AI-generated articles surpassed human-written at 50.9%, before returning to 49.9% in Q1 2026.” It supersedes the earlier Graphite study, whose estimates it puts “3.3 percentage points higher”.
Two things about that curve matter more than the headline. First: it flattened. The share climbed steeply, hit about half in early 2025, and has hovered there since — it did not continue to a machine-only web. Second: publishing is not reading. The same research found AI-generated articles are produced in enormous volume but appear far less often among top-ranking search results, which still skew human-written or heavily human-edited. An enormous amount of machine text exists that essentially nobody reads.
THE PREDICTION THAT FAILED
You have probably encountered the claim that 90% of online content would be synthetic by 2026. It was repeated endlessly in the name of a 2022 Europol report, Europol has since removed the statement behind it, and it did not happen.Europol Innovation Lab, Facing reality?, 2022, publication page read 17 Sep 2026: “In the updated version, a statement from an inaccurate source on the expected future share of synthetically generated content was removed.” It is on this page because a record that only keeps the frightening numbers is not a record. When a forecast fails, that failure is data too.
Judge the machine-written web by what people actually read: the volume is large, but it has levelled off and reaches the top of search results far less often.
So is the dead internet theory true?
Partly — and not in the shape it is usually told. The strong version says the web is now essentially machines talking to machines, with humans a vanishing minority and the whole thing orchestrated. The measurements support a much narrower claim: automated requests outnumber human ones, roughly half of newly published articles are machine-written, and both trends arrived fast. They do not support the idea that the people are gone. They are still here, still reading, still overwhelmingly what the top of search results is made of.
The honest summary is that the internet did not die. It got much louder, and the cost of producing plausible noise fell to nearly zero. That is a real change with real consequences — for search, for trust, and for fraud — without requiring anyone to believe the strong version. The full accounting is in the explainer, and the current figures live in the quarterly record.
Keep the narrow claim and drop the strong one: the web got louder and cheaper to fill, but the people are still here, reading.
What this changes for you
- Engagement counts mean less than they did. Views, likes, and replies are the cheapest things on the internet to manufacture. Treat them as decoration, not evidence.
- Volume is no longer a signal of effort. A site with ten thousand articles may represent an afternoon. Depth, sourcing, and specificity are the scarce things now.
- Consensus can be manufactured. If a claim appears in fifty places with the same phrasing, that is one source and forty-nine copies. Trace it back before believing it.
- The fraud is the concrete harm. Not the volume — the deception. Nearly $900 million was lost in US fraud complaints that mentioned AI in 2025 alone.FBI, 2025 IC3 Annual Report, read at source 17 Sep 2026: “In 2025, IC3 received more than 22,000 complaints reporting AI-related information. Adjusted losses of these complaints exceed $893 million.” See how AI fraud works and the defences.
Half the traffic being automated is not in itself alarming — much of it is the plumbing that makes the web usable. What deserves attention is narrower and more specific: manufactured consensus, engagement that cannot be trusted, and fraud that now speaks in a familiar voice. Worry about those. The figures on this page are dated because they will change, and if next year's numbers move the other way, that will be recorded here too.
Spend your worry on manufactured consensus, engagement you cannot trust and fraud, not on the share of traffic that is automated.