Bot Traffic Passed Human Traffic in May, a Year Ahead of Forecast

“I have called it wrong at every point along the way.”

That’s a chief financial officer, on an earnings call, describing his own forecasting record. Thomas Seifert said it on Cloudflare’s second-quarter call while telling analysts that machine traffic had overtaken human traffic on the internet in May. Here’s the full sentence:

“To give you a sense of how this trend is playing out, and with the big caveat that I have called it wrong at every point along the way, if the current trends continue, we think in five years, non-human traffic will be as much as 1,000 times as much as human traffic.”

He then gave The Register the line that got quoted everywhere:

“In other words, humans will be a rounding error on the internet, not because human traffic goes down, but that’s just how fast we’re seeing non-human traffic grow.”

Nearly every outlet that ran the 1,000x figure cut the clause in the middle of it. The number travelled; the doubt attached to it did not.

That’s worth pausing on, because the doubt is earned rather than performative. Cloudflare put the crossover in the second half of 2027 back in November 2025, then moved it up to the first half of 2027 in March. It arrived in May. The curve keeps outrunning the people with the best seat to measure it, and Seifert is the rare forecaster who says so into a microphone with analysts on the line.

What a marketer is actually looking at

Every number in your weekly report was designed for the half that’s now smaller. Sessions, bounce rate, time on page, pages per session, engagement rate: each one describes a person deciding whether to keep reading. None of them describes a machine fetching a page to answer somebody’s question somewhere else.

That gap has been widening under the reporting for a while. We wrote in July that publishers were rebuilding their sites so agents could read them, and Time now sells sponsored blocks inside the bot-readable version. What Seifert added is a date. The audience split flipped in May. Very few analytics setups were reconfigured in June.

The agent as a customer

Julien Bek, a partner at Sequoia Capital, put the consequence on 20VC on August 24 in a way that isn’t about infrastructure at all:

“We’re 3 years into AI, and we’re already at parity in terms of agent traffic to human traffic.”

His argument runs past the measurement problem to a demand-side one:

“on the demand side, you have a new customer that we’re not treating as good as human customers as the agent.”

Which is a strange sentence until you sit with it. Marketing spends heavily on how a page feels to a person: the hero image, the scroll, the proof above the fold. An agent gets none of that. It reads the text, takes what parses, and leaves. The budget line that wins a human is invisible to the reader that now makes up most of the requests.

Someone is already building the index they read

The day after Bek’s clip went up, TechCrunch reported that Keenable had raised $26 million led by Accel to run a web index of more than 100 billion documents, built for AI systems rather than people. Andrey Styskin, who previously ran search, AI and cloud at Yandex, co-founded it. He described why the shape of the thing differs from a consumer search engine:

“This actually creates a new flywheel that is different from what Google learned from human behavior.”

Fifteen engineers, doubling by year end, several AI labs already using it in production.

The difference carries a budget consequence. Google’s ranking learned from what people clicked. An index built for machines learns from what answers a query well enough for a model to use it. Those are different signals, and there’s no submission form for the second one. If a meaningful share of your buyers’ research runs through an assistant calling an index like this, your visibility depends on a system you can’t audit, can’t appeal to, and can’t buy placement in.

Who profits from the number being large

All three sources sell something the number justifies. Cloudflare sells bot management, and traffic growing 1,000x is the best case its sales team will ever have. Keenable sells the index. Sequoia funds the companies serving the agents. A forecast from someone who profits from the forecast is a sales document until a disinterested party confirms it, and on the 1,000x figure nobody has.

The May crossover is different. It’s a measurement of something that already happened, taken by the company routing the traffic, and it landed ahead of that company’s own published expectation rather than conveniently on top of it. Treat that one as real and the five-year projection as marketing.

The verdict

The number to distrust is 1,000x. The number to act on is that the split already flipped, in May, roughly a year before the best-informed public estimate expected it. Go and find out what share of the requests to your own site are not people, then ask which of the metrics on your weekly report still means what you think it means. Most teams have never separated the two, which means the trend line they’ve been reporting all year is measuring an audience mix that changed underneath it.