Half of Web Traffic Is No Longer Human, and the Advertising Model Cannot Survive That Fact
Thirty years of digital economy built on an assumption that no longer holds: that on the other side of the screen there is a person. In 2024, for the first time in a decade of systematic measurement, bots surpassed humans as a source of internet traffic. According to the Imperva report, automated traffic reached 51% of the global total. Malicious bots alone accounted for 37%. Human traffic, by definition, is already the minority.
This is not a poorly contained cybersecurity problem, nor a passing anomaly. It is a structural change in the architecture of web usage that invalidates the assumptions upon which the dominant monetisation models were designed. Digital advertising, in its current form, was built to capture the attention of someone who looks, feels, hesitates, and buys. None of those conditions apply to an artificial intelligence agent that receives an instruction, consults twelve sites in parallel, extracts the relevant data, and executes the transaction without having processed a single banner ad.
The problem is not that there are too many bots. The problem is that the bots most dangerous to the advertising model are the most sophisticated: AI agents with the capacity for action. According to Human Security, that segment of automated traffic grew eight times faster than human traffic in 2025. These are systems that book flights, refill medical prescriptions, or compare prices on televisions without opening a visible browser tab. They are not clumsy scrapers: they are autonomous buyers with a delegated credit card and zero willingness to be interrupted by advertising.
The Business That Financed the Free Web Operated on a Condition Nobody Wrote Into the Contract
The first advertising banner on the internet was published in October 1994. AT&T paid for space on HotWired and obtained a 44% click-through rate. It was a world in which the novelty of the format did the work on its own. Over the following three decades, the industry built an increasingly complex monetisation architecture — real-time bidding, behavioural targeting, retargeting, viewability measurement — but never abandoned the foundational assumption: there is a human being on the other side who can be persuaded.
That assumption worked while humans were the majority. Now they are not. And the inflection point was not gradual: it was accelerated by the mass adoption of large language models that turned the automation of web browsing tasks into something accessible to millions of users with no technical knowledge.
The consequence for publishers is immediate and mathematical. If traffic grows but the human proportion falls, the advertising inventory inflates while the real value of each impression erodes. Advertisers pay for the illusion of an audience that in part does not exist as a persuadable subject. The session and page view metrics that underpin commercial negotiations with agencies have lost their precision as a proxy for human attention. The Harvard Business Review described it in April of this year as a direct threat to the revenue streams of platforms such as Google and Meta, and to the open web as a whole.
For independent publishers, the impact is more immediate because they have less capacity to absorb it. Without human clicks, the generation of first-party data dries up. Without first-party data, targeting capability degrades. Without precise targeting, the cost per thousand impressions falls. The chain is short and has no weak link: the whole thing fails.
The Toll as a Model, Not as a Metaphor
The replacement of the billboard by the toll gate is not a poetic image. It is an operational description of what is already happening in the infrastructure of the web.
Cloudflare, which processes traffic for approximately one fifth of all sites on the internet, activated HTTP status code 402 — "Payment Required", reserved in the protocol specification since the 1990s and practically never used — to charge artificial intelligence crawlers for access to content. The logic is precise: if an agent extracts value from a page without generating any advertising revenue, access must have a direct price.
TollBit operates with the same logic but as a layer over thousands of publisher sites, including the publishing arm of the Washington Post and the Philadelphia Inquirer. It installs a per-page charging mechanism for pages read by bots. According to available figures, close to one fifth of those sites are already generating revenues of tens of thousands of dollars per month through this channel. It is not yet advertising-scale revenue, but it is a proof of concept that works.
The problem that immediately emerges is one of financial plumbing. An AI agent executing a search for "65-inch television at the best price" can generate dozens of requests to different sites in fractions of a second. The advertising model charged an advertiser once for an impression. The toll model has to charge thousands of machines, constantly, in fractions of a cent. The traditional payments infrastructure, designed for consumer transactions with minimum viable values above the operational cost of processing, cannot handle that level of granularity.
The answer to that plumbing problem arrived this very month. On 10 June, Mastercard launched Agent Pay for Machines, a payments layer designed specifically for the high-frequency, low-value transactions that AI agents execute. Visa advanced in parallel with its own agentic checkout protocol, with integrations that allow agents to make purchases on behalf of a user with delegated authentication. The two card networks that for decades charged a fraction of every consumer transaction have just positioned themselves to charge that same fraction on every machine-to-machine micropayment.
The architecture that emerges is not difficult to read. The publisher builds the toll gate, but does not have the capacity to process it alone. It needs Cloudflare at the door to manage traffic and identify the bot, and Visa or Mastercard at the till to settle the micropayment. The visible margin belongs to the publisher. The structural margin — the one that does not depend on whether the content is good or the site is popular — belongs to the infrastructure networks.
Where Value Settles When Attention Ceases to Be the Product
The shift that this moment describes is not only technological. It is a redistribution of power along the digital value chain that has been concentrating for decades in platforms that controlled human attention and that now face a depreciating asset: the eyeball economy.
Google and Meta built empires on the capacity to direct advertising messages to people at their moment of greatest receptivity. That model works with human behavioural signals — searches, likes, dwell time — that AI agents do not generate, because they have no emotional states, no purchase impulses, and cannot be segmented by affinity. An agent searching for the cheapest flight between Madrid and Mexico City has no airline brand preference unless the user has explicitly programmed one. There is no space for branding advertising in that flow.
The strategic consequence for advertising platforms is that the growth of agentic traffic degrades the denominator on which they calculate their value. More total traffic with proportionally less human attention means lower effectiveness per impression, greater pressure from advertisers to verify audience quality, and higher audit costs to demonstrate that users are real.
The companies best positioned for the new model share a common characteristic: they monetise by transaction or by infrastructure usage, not by attention. Cloudflare charges for processed requests, for applied security rules, for protected bandwidth. Mastercard and Visa charge for settlement, regardless of whether the person initiating the transaction is a consumer in front of a screen or an agent executing an instruction at three in the morning. Their revenue model has no column called "human attention."
For publishers who built their businesses on organic traffic and programmatic advertising, the path of adaptation is narrower. They can diversify toward subscriptions, toward commerce with commission per verified transaction, or toward content licensing with the very AI models that today consume their content without paying. Some are already attempting this. But any of those options requires relinquishing traffic volume as the central metric of value, which means accepting that the majority of the visits they receive today are, in economic terms, noise.
The fracture between what the advertising market measures and what web traffic actually represents is already structural. Patching it with better bot detection tools buys time, but does not change the direction of movement. AI agents will continue to multiply because they create value for the user who delegates to them. The advertising attention model will continue to lose its share of persuadable audience. And the money will continue moving toward where it always ends up: in the layers of infrastructure that every transaction must cross, with or without a human being involved.
