DataDome today published The AI Traffic Report Q3 2026, which counts nearly 25 billion AI agent requests across its network between July and September, up 41.5% from the second quarter, and finds that Meta's Muse agent accounted for 33% of all agentic browser requests in September despite being measured for only nine days of the month.

In Short

A bot-security company counted the AI programs visiting the websites it protects and found that ones which click around sites on a person's behalf grew almost fivefold in a single month. Meta's new Muse assistant was a third of that activity, and almost all of it came from big cloud servers rather than people's home internet connections, which means website checks that look for home connections will not spot it. The numbers come from the security company itself, which sells tools to manage this traffic, and it does not say how many sites its count covers.

A quarter that reset the baseline

The numbers in the report are DataDome's own, drawn from traffic crossing its bot-management network, and they have not been independently audited. According to DataDome, its network processed nearly 25 billion AI agent requests in the third quarter of 2026, up from 17.7 billion in the second quarter. That brings the company's 2026 total to "nearly 55 billion," which means close to half of all the AI agent traffic it has recorded this year arrived in the last three months.

The growth did not come as a steady climb. July dipped 5.6% below June, to 6.23 billion requests. August then jumped 47% to 9.18 billion, and September added a further 4.4% to reach 9.58 billion, the highest month in DataDome's dataset and 45% above June. The three monthly figures add up to 24.99 billion, consistent with the quarterly total.

Those numbers also line up with DataDome's earlier publications. The Q2 figure of 17.7 billion, a 45% quarter-on-quarter rise, had already been cited in IAB Australia guidance on automated traffic in August. And when DataDome released its State of Bot & Agent Security Report on September 22, it put June AI requests at 6.6 billion, which is exactly the June baseline implied by the 5.6% July dip in the new report.

According to the report, "Traffic reset to a new, higher baseline within weeks, and it held." The report is written by Jerome Segura, DataDome's VP of Threat Research, and the accompanying press release went out at 9:00 a.m. Eastern Daylight Time today.

What the report does not provide is the size of the network behind these counts. It refers throughout to "DataDome's network" without stating how many sites or customers contributed traffic in the quarter. The September security report described customer data covering more than 75,000 sites, but the Q3 report does not say whether the same base applies. Any change in the customer mix between quarters would affect quarter-on-quarter growth rates, and the report offers no way to check for that.

Meta still leads, but its share slipped

Meta's two crawlers, Meta-ExternalAgent and Meta-WebIndexer, remain the two largest agents on DataDome's network. Together they generated 12.4 billion requests in the third quarter, up 36.5% on the second quarter, or 49.7% of all AI agent traffic. That is below Meta's 51.5% share in Q2, because non-Meta agents grew faster, at 47%.

Within the quarter, however, Meta's share rose each month: 43.9% in July, 49.7% in August and 53.4% in September. The September figure is driven by Meta-WebIndexer, the crawler that feeds the search index behind Meta AI. It passed Meta-ExternalAgent for the first time in June, fell back behind in July and August, and led again in September with 2.63 billion requests against 2.48 billion. Its climb from about 1 billion requests in July to 2.6 billion in September, while ExternalAgent dipped, "suggests Meta is continuing to invest in the 'answering' side of AI, not only the 'learning' side," according to the report.

Volume is not value, though. The report sets Meta's 12.4 billion Q3 requests against 51,000 Meta AI referrals - visits sent back to sites by Meta's assistant. That is roughly one referral for every 243,000 requests. "Meta's volume is still an infrastructure cost, not a business signal," according to DataDome.

ClaudeBot overtakes GPTBot

ClaudeBot, Anthropic's training crawler, grew 159% quarter on quarter, from 892 million requests to 2.3 billion, and moved into fourth place on DataDome's ranking. It passed OpenAI's GPTBot, which fell 12%. Two new entrants appeared just outside the top ten: Claude-SearchBot with 358 million requests and Amzn-SearchBot with 103 million.

Here again, crawl activity and returned traffic moved in opposite directions. According to the report, Claude's referrals fell while ClaudeBot crawling rose. OpenAI's pattern ran the other way: ChatGPT-User, the fetcher that retrieves pages when a ChatGPT user asks a question, fell 37% in a single month, while ChatGPT referrals reached a record. The report does not give absolute referral figures for either company, so the scale of those movements cannot be checked. "Crawl volume is a poor proxy for value," it concludes.

Three bots per lab

The report's description of the crawler landscape matters to anyone managing robots.txt files. OpenAI, Anthropic and Amazon now each run three kinds of bot, according to DataDome: one that collects content to train models, one that indexes pages for AI search, and one that fetches a page in real time when a user asks a question. Each carries its own name, so a site can allow or block them separately.

That split, according to the report, "gives publishers what they asked for: a way to refuse model training without disappearing from AI search and answers." It also lets the AI companies build their own search indexes and "depend less on Bing, Google or Brave." The cost falls on site operators: every new bot name is another allow-or-block decision, and "the companies don't always say up front how each bot treats robots.txt."

There is a larger gap. Agents that load pages inside a regular browser do not carry any of these bot names, according to DataDome, and are not covered by these controls at all. That is the category where the quarter's sharpest growth happened.

Agentic browsers grew 4.7x in a month

Agentic browsers are tools that load and act on pages on a user's behalf rather than simply fetching content. Many are built on a computer use agent that reads the screen and clicks like a person. According to DataDome, total agentic browser requests on its network rose from about 12.5 million in August to 58.2 million in September, a 4.7x increase. The press release headline rounds this to "nearly 5x."

Muse explains part of the jump, but not all of it. Excluding Muse, agentic browser volume still tripled, from 12.5 million to 38.7 million. The September breakdown, according to the report, looks like this:

  • ChatGPT's cloud browser: 27.4 million requests, up 4.5x from 6.1 million, and 47% of the category
  • Meta's Muse: 19.5 million requests, 33% of the category
  • Claude's browser extension: 10.5 million requests, up from 3.6 million, nearly triple
  • Perplexity's Comet: 0.8 million, down from 2.4 million in July

Those four add up to 58.2 million, matching the total. The report describes Comet's drop as 65%; on the rounded figures given, the fall is closer to 67%.

The order is very different from the picture HUMAN Security painted earlier in the year. In its May data, Comet held 47% of agentic traffic and OpenAI's Atlas 20.3%, with the Claude Chrome extension at 18.6%. Since then, OpenAI began sunsetting Atlas on July 9 and folded browsing into its ChatGPT desktop app. The two vendors measure different networks with different methods, so the shares cannot be compared directly, but the direction of travel for Comet is the same in both.

DataDome adds a caveat about Muse's place in the list. "Muse is not a standalone browser," according to the report. "It is a personal AI agent that operates inside its own dedicated cloud virtual machine, which contains a full browser, to complete multi-step tasks on a user's behalf." DataDome counts it with browsers because the traffic it generates is browser-driven.

Muse, seen from the server side

Muse generated 29.1 million requests on DataDome's network between September 22 and October 3, which the company calls its "first 12 days of full availability." Of those, 19.5 million fell in September, about 1.9x the Claude browser extension's monthly total. Because detection only covers the last nine days of September, DataDome describes those figures as floors.

That start date needs a note. Meta released Muse in the United States on September 8, on iOS, Android, the web and WhatsApp, and Sensor Tower estimates put installs above 5 million across the US and Canada within 22 days. DataDome does not explain what "full availability" on September 22 refers to, or whether Muse traffic before that date went undetected or was simply absent. The 29.1 million figure therefore excludes the agent's first two weeks in the market.

A daily rhythm

According to DataDome, daily Muse volume rose 4.4x in nine days, from 0.8 million requests on September 22 to 3.7 million on October 1, roughly 18% growth per day. The peak was 484,000 requests in a single two-hour window on October 3.

The traffic followed a clear daily cycle. Volume rose from 08:00 UTC, peaked at 16:00 UTC and fell back until 08:00 UTC the next day. That pattern is "consistent with traffic following user activity rather than a scheduled crawl," according to the report. A 16:00 UTC peak corresponds to noon on the US East Coast and 9:00 on the West Coast in October, which fits an agent then available only in the US and Canada.

99.3% from the cloud

The more consequential finding concerns where Muse's requests come from. In a sample of Muse traffic, 99.3% originated from cloud and CDN infrastructure: Cloudflare accounted for 81.2% and Fastly for 18.1%. Only 0.6%, or 17,500 requests, came from consumer internet service providers, spread across eight networks: Comcast, Bell, Verizon, Charter, Rogers, AT&T, T-Mobile and TELUS.

The explanation, according to DataDome, lies in how Muse is built. Each user's agent works in its own cloud virtual machine, so requests leave from data-centre infrastructure rather than from the user's phone or home router. The result, in the report's words, is that "device and home-IP signals will not separate Muse from other cloud-hosted traffic, even though a real user's task sits behind the requests."

"Muse is the first mass-market example of where agentic traffic is heading. It works from its own cloud virtual machine, so nearly all of its requests look like data-center traffic rather than a user's device, and a meaningful share lands on forms and checkout flows," Segura said in the press release. "Policies that trust residential or ISP-based traffic won't catch agents like this."

This finding sits awkwardly against earlier reporting on the same agent. Profound, in research covered by PPC Land on October 11, stated that Muse and the Instinct agent route traffic through residential internet connections and present as ordinary human visitors. DataDome's sample shows the opposite for Muse: almost none of it arrives via residential connections. Profound's statement covered two agents and did not publish traffic data to support it, while DataDome publishes a breakdown by network. Neither vendor's data has been independently verified.

What Muse touches

The same sample shows where on a site Muse goes. According to DataDome, 8.9% of Muse requests hit what the company calls sensitive flows: roughly 202,600 form requests, 24,800 login requests, 9,600 payment requests, 6,800 account-creation requests and 5,200 cart requests.

Forms make up 81% of that sensitive traffic and run about eight times higher than logins. DataDome reads this as Muse "reaching past browsing and reading into forms and checkout, while mostly acting logged out rather than signing in as the user."

The report does not state the sample size. It can be inferred, however. If 17,500 requests represent 0.6% of the sample, and the five sensitive categories (about 249,000 requests in total) represent 8.9%, both figures point to a sample of roughly 2.8 million requests, or a little under a tenth of Muse's 29.1 million total. That is a PPC Land calculation, not a figure DataDome published, and the report does not say how the sample was drawn or which dates it covered.

HUMAN Security's earlier work on Muse pointed in a similar direction. Its data, published September 28, found Muse averaging 72% of daily AI agent requests on HUMAN's network, with 84% targeting product and search pages and a higher likelihood than other agents of reaching checkout. The 72% and DataDome's 33% are not in conflict: HUMAN measured Muse against all AI agents it tracks on a daily basis, while DataDome measured it against agentic browsers only, over a month in which Muse was present for nine days.

How DataDome identified Muse

One question the report leaves open is how DataDome recognised Muse traffic in the first place. Muse "doesn't declare itself," according to the report, and the press release states that "Muse does not identify itself as an AI agent." Yet DataDome attributes 29.1 million requests to it with enough precision to separate them by two-hour window.

The report does not describe its detection method. DataDome's product is built on behavioural and signal analysis, so the attribution likely rests on fingerprinting rather than a declared identity, but that is an inference. Amazon faced a similar puzzle: when it cut Muse off from shopping on Amazon.com on September 20, one of its stated grievances was that Muse does not identify itself while browsing, and it did not say how it recognised the traffic either.

Undeclared agents are not new to DataDome's research. A June roundup of AI crawler user agent strings cited DataDome data finding that 80% of AI agents do not identify themselves properly. In December 2025, Segura's team documented Grok sending 16 requests from 12 IP addresses for a single query, none of which identified as xAI or Grok.

What DataDome proposes

The report ends with recommendations, which reflect the services DataDome sells. It argues for classifying agents by function rather than by vendor, setting explicit rules for what agents may do on forms, login, account creation, cart and checkout, and making "a deliberate Muse decision" to allow, limit or block it. It also calls for validating agent identity beyond the user agent string, naming Web Bot Auth and checks against published IP ranges as "the minimum bar," and for applying behavioural analysis, since "form interaction, session flow, and API sequencing patterns are detectable, but not with signature-based approaches designed for crawlers."

DataDome describes itself as "the leader in bot and agent trust management," and its press release cites a Leader placement in The Forrester Wave for bot and agent trust management software in Q2 2026. The company has a direct commercial interest in the finding that ISP and device signals no longer separate good traffic from bad. On October 1, it joined Experian's Agent Trust programme to provide what the two companies call continuous intent verification for AI shopping agents.

Why this matters for marketers and publishers

For advertisers, the Muse figures raise a measurement question before they raise a security one. If an agent operating on a real customer's instruction reaches a product page or a cart from a Cloudflare or Fastly address, analytics and bot filters built around residential traffic will either count that visit as data-centre noise or block it outright. Both outcomes hide a shopper who exists. The 8.9% of Muse requests reaching forms, carts and payment pages suggests these are not idle visits.

The report also complicates the case for letting crawlers in on the promise of referrals. Meta's 12.4 billion quarterly requests yielded 51,000 Meta AI referrals on DataDome's network, and Claude's referrals fell as ClaudeBot's crawling rose. A Botify and Nectiv analysis in April found the same pattern from another angle, with OpenAI's total crawl roughly tripling after GPT-5 while ChatGPT-User events fell 28%.

For publishers, the separate bot names for training, search and user-triggered fetching offer more granular control, though only over agents that declare themselves. Muse, the fastest-growing agent in the report, does not. And the rankings move quickly: in one quarter, ClaudeBot more than doubled, GPTBot shrank, ChatGPT-User fell 37% in a month, Comet lost two thirds of its volume and Muse went from zero to a third of a category. "Rankings and bot names will keep shifting, so static rules won't hold," according to the report.

Whether 33% of agentic browser traffic in nine days will hold as Muse moves beyond the US and Canada remains open. DataDome's Q4 report, if it follows the same schedule, would cover the first full quarter of Muse traffic.

Timeline

Summary

Who: DataDome, a bot and agent security company, through a report by Jerome Segura, its VP of Threat Research. The data concerns Meta's Muse agent and crawlers, Anthropic's ClaudeBot and Claude browser extension, OpenAI's GPTBot, ChatGPT-User and cloud browser, Perplexity's Comet and Amazon's new search bot.

What: The AI Traffic Report Q3 2026 counts nearly 25 billion AI agent requests on DataDome's network in the third quarter, up 41.5% from 17.7 billion. Agentic browser requests rose 4.7x from August to September, to 58.2 million, with Muse accounting for 33%. In a sample, 99.3% of Muse requests came from Cloudflare and Fastly infrastructure and 8.9% reached forms, login, payment, account creation or cart endpoints. The report does not state its network size, its Muse sample size or how it identified Muse traffic. All figures are vendor-supplied.

When: The report and press release were published today, October 15, 2026. The data covers July through September 2026, with Muse figures from September 22 to October 3.

Where: Traffic across DataDome's customer network, with no geographic breakdown given; Muse was available in the US and Canada during the measurement window, and its consumer ISP traffic came from US and Canadian networks.

Why: Muse runs each user's agent in a cloud virtual machine and does not declare itself, so bot rules that trust residential connections or rely on declared bot names do not separate it from other data-centre traffic. That affects how advertisers, retailers and publishers count, filter and allow agent-driven visits that may represent real customers.