The Web's Fastest-Growing Users Aren't Human: AI Agent Traffic Jumped 41.5% Last Quarter

Something strange is happening to the web's traffic statistics. The fastest-growing class of visitors is not human, and it is not the familiar crawlers that index pages for search engines either. According to new data from bot-management company DataDome, AI agent traffic grew 41.5% quarter over quarter in the third quarter of 2026, with nearly half of all agent traffic recorded this year arriving in just the last three months. The agentic web has stopped being a prediction. It is now measurable, and it is accelerating.

DataDome's network processed nearly 25 billion AI agent requests in Q3 2026, up from 17.7 billion in Q2. That brings the 2026 total to nearly 55 billion. The growth did not arrive as a steady climb. July dipped 5.6% below June, to 6.23 billion requests, then August jumped 47% to 9.18 billion, and September edged up another 4.4% to 9.58 billion, the highest month in the dataset and 45% above June. Traffic reset to a new, higher baseline within weeks, and it held.

The browser that browses for you

Most of that traffic still comes from crawlers: software that reads pages. But the fastest-growing slice is something new: agentic browsers, AI systems that don't just read the web but act on it, filling in forms, clicking through checkout, navigating login flows. In September, agentic browser requests jumped 4.7 times in a single month, from 12.5 million to 58.2 million. Even excluding Meta's new Muse browser, volume tripled to 38.7 million. ChatGPT's cloud browser and Claude's browser extension both grew rapidly, while Perplexity's Comet browser fell 65% from July to September.

The distinction matters because a crawler reads a page and leaves, while an agentic browser does things: it logs in, it adds items to carts, it submits payment details, it creates accounts. Every session raises the stakes, because a browser that can transact on a user's behalf can also transact on an attacker's. DataDome's conclusion is blunt: capacity and bot-management policies built around last quarter's averages are already out of date.

The agentic web has stopped being a prediction. It is now measurable, and the fastest-growing visitors on the internet are software acting on people's behalf.

Muse: the browser nobody sees

The standout newcomer is Meta's Muse, which spins up a remote browser in the cloud for each user rather than running on the user's device. It was fully available only from September 22, yet it made up 33% of September's agentic browser volume in just nine days, generating 29.1 million requests between September 22 and October 3. Here is the detail that should worry every site owner: Muse does not identify itself as an AI agent. In DataDome's traffic sample, 99.3% of its requests came from cloud and CDN infrastructure, indistinguishable from ordinary data-center traffic.

And it reaches deep into the customer journey. Of the sampled Muse requests, 8.9% touched forms, login, payment, account creation, or cart endpoints, the parts of a site that carry the most fraud risk. Jerome Segura, DataDome's VP of threat research, says Muse is the first mass-market example of where agentic traffic is heading, and warns that policies which trust residential or ISP-based traffic will not catch agents like this. Organizations, he argues, need to validate agent identity and understand intent, then decide deliberately which agents to allow and what they are authorized to do.

The identity crisis

That question, who is this agent acting for, and what is it allowed to do, is now the central problem of the agentic web. On October 15, the same day DataDome published its report, a Boston company called Proof launched x401, an open HTTP protocol that lets any website ask an AI agent to prove who it is acting for and what that person approved. Proof works with Meta's Muse, ChatGPT, Grok, Claude, and other assistants that support the Model Context Protocol (MCP).

The mechanism is a digital identity wallet. A consumer holds their verified legal identity in Proof's agent identity wallet and approves, on their own device, exactly what their AI agent may do. The agent carries the request; the person's wallet signs it. The receiving business can independently confirm the verified identity and check the cryptographic signature on the transaction. If the agent changes the amount, recipient, or other signed terms, the signature no longer validates. With the launch, a consumer can authorize a wire transfer, approve an itemized purchase, or sign a payment mandate through an AI agent.

The timing is not accidental. According to Proof's announcement, the launch comes weeks after Amazon blocked Muse from its marketplace, a dispute that exposed the core problem: there was no reliable way to know who an AI agent was acting on behalf of, or whether that person had authorized what it was doing. And it follows September guidance from FinCEN and federal banking regulators recognizing digital certificates and verifiable credentials, under certain conditions, as a permissible way for banks to verify customer identity.

What the numbers imply

Step back and the pattern is clear. Meta still generates the most agentic traffic, 12.4 billion requests, but its 49.7% share slipped from 51.5% in Q2 as non-Meta agents grew faster (47% versus 36.5%). ClaudeBot surged 159% quarter over quarter to 2.3 billion requests, passing GPTBot, which fell 12%, to become the fourth-ranked agent. Meta-WebIndexer retook the top spot in September at 2.63 billion requests, just ahead of Meta-ExternalAgent at 2.48 billion. The agent ecosystem is diversifying fast, and no single vendor's behavior defines it anymore.

For website owners, the message is twofold. First, traffic from software will keep growing faster than traffic from people, and every form, login, and checkout flow is now a potential API endpoint for agents that may or may not be legitimate. Second, identity is becoming infrastructure: protocols like x401 point toward a web where agents carry credentials the way humans carry passports. The sites that figure out which agents to welcome, and which to block, will have a decisive advantage over the ones that treat all software visitors the same.