- DataDome's Q3 2026 report says AI agent traffic on its network grew 41.5% to nearly 25 billion requests, with Meta's bots at 49.7%.
- Meta's Muse agent browses from cloud machines. DataDome says rules that trust home-network traffic will miss it, and name-based bot controls do not cover browser-driven agents.
- Leaders should review bot rules, decide which agents may do what, and check which forms and checkout flows automated visitors can reach.
The question is no longer person or machine
For years, website owners could ask one question about a visitor: is this a person or a bot? DataDome's AI Traffic Report for Q3 2026 points to a more useful question. Whose errand is this, and is it allowed?
The report covers traffic on DataDome's own network, not the whole internet. Help Net Security summarised it. Within that data, DataDome says AI agent traffic grew 41.5% from the previous quarter to nearly 25 billion requests.
What DataDome measured
Q3 alone holds close to half of all the AI agent requests DataDome has logged so far in 2026. DataDome says this growth left capacity plans and bot rules based on the prior quarter's averages out of date.
Meta-ExternalAgent and Meta-WebIndexer together sent 12.4 billion requests. That is 49.7% of the AI agent traffic DataDome recorded. In September, Meta-WebIndexer passed Meta-ExternalAgent. The report's reading is that this shift signals continued investment in AI answer generation. ClaudeBot placed fourth for the quarter and moved ahead of GPTBot.
How today's controls work, and where they stop
OpenAI, Anthropic and Amazon each run separate bots for three jobs: collecting training data, indexing pages for AI search, and fetching pages when a user asks a question. Each bot has its own name. A website owner can allow or block each one through robots.txt, the file that tells automated visitors what they may access.
That split lets a publisher refuse model training and still appear in AI search and answer services. The price is more settings to manage. Vendors do not always say clearly, ahead of time, how each crawler responds to robots.txt rules.
There is a bigger blind spot. An agent that works through an ordinary browser does not present any of those bot names. A rule written for a name has nothing to match, so the allow-or-block choices never reach that agent.
Muse: why name-based rules are unlikely to cover it
Meta's Muse is a personal AI agent. It runs in its own cloud virtual machine, which holds a browser. DataDome files Muse under agentic browsers because its traffic is browser-driven.
DataDome says Muse sent 29.1 million requests in its first 12 days of full availability, from September 22 to October 3. In the same dataset, agentic browsers sent 12.5 million requests in August and 58.2 million in September. Take Muse out, and September still reached 38.7 million, more than triple August. Muse alone was 33% of September's total.
Other agents moved differently. In September, ChatGPT's cloud browser was the largest source in that category. Comet's traffic fell 65% from July to September.
Muse's daily traffic pattern was consistent with user activity rather than a scheduled crawl. DataDome examined a sample of its requests. Almost all of them, 99.3%, came from cloud and CDN networks. Of the sampled requests, 8.9% hit sensitive pages: login, payment, account creation, cart and forms. Forms drew 81% of that sensitive traffic, about eight times what login pages received. The report summary does not give the sample size.
Why a sign on the door is not enough
The lesson here is that robots.txt works like a sign on the door. It helps with visitors who read signs and give their name. Muse is browser-driven and works for a person from its own cloud machine. Name-based controls are therefore unlikely to cover it.
That is our inference. It rests on DataDome's description of the category and on Segura's comments below. The coverage does not say what name, if any, Muse presents to websites.
It matters most where Muse was seen going: forms and checkout. Those pages collect data and start business processes. An unexpected visitor there is a different problem from one reading an article.
Jerome Segura, DataDome's VP of Threat Research, calls Muse "the first mass-market example of where agentic traffic is heading." He notes that because Muse works from its own cloud machine, nearly all its requests look like data-center traffic, not a user's device. Any rule that gives home or ISP connections a pass would not apply to it.
His advice has three parts. Check who an agent is. Read what it is trying to do. Then set permissions agent by agent.
Questions to put to your team
First, were this year's capacity plans and bot rules built on Q2 averages? DataDome says that baseline is already out of date on its network.
Second, which bots do we allow by name, who made that decision, and is it written down?
Third, do any rules treat residential or ISP traffic as trusted by default? Segura says rules like that would not catch agents such as Muse.
Fourth, which forms and checkout flows can an automated visitor reach, and what happens when one submits them?
Fifth, can we verify who an agent is, and have we decided what each one may do?
Asking visitors to announce themselves is a courtesy. Deciding what they may do is a policy.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Help Net Security.
