One third. That is the share of agentic-browser traffic that Meta's shopping agent Muse captured in the nine September days DataDome could measure it, according to the bot-detection vendor's Q3 2026 AI Traffic Report, published October 15. Per the report, Muse did this from cloud servers instead of home internet connections, without declaring itself as an AI agent, and while touching logins, carts and payment pages on the sites it visited. For merchants and payments teams who have spent months being told to build agent policies around registered, self-identifying bots, that is the inconvenient part.

DataDome logged nearly 25 billion AI agent requests across its network in Q3, up 41.5% from 17.7 billion in Q2, with a 2026 total of nearly 55 billion. Inside that, the category it calls agentic browsers (tools that open sites and act on a user's behalf rather than crawl for model training) jumped from about 12.5 million requests in August to 58.2 million in September, a 4.7x increase. Without Muse, the figure is 38.7 million, still roughly triple. In DataDome's September split, OpenAI's ChatGPT cloud browser led with 27.4 million requests (47%), Muse had 19.5 million (33%), Anthropic's Claude extension 10.5 million, and Comet 0.8 million. PPC Land's write-up is the most complete public account of the numbers.

The figures come from one vendor, are unaudited, and cover a short window. Treat the 33% as a flag, not a verdict.

What DataDome measured

Meta released Muse in the US on September 8 on iOS, Android, web and WhatsApp. Amazon cut it off from shopping on Amazon.com on September 20; FinTech Weekly quotes Amazon saying third-party purchasing apps "should operate openly and respect service provider decisions" about participating. DataDome's window begins September 22, which it calls "full availability" without defining the term, so the first two weeks are excluded.

From September 22 to October 3, DataDome counted 29.1 million Muse requests. Daily volume grew from 0.8 million on September 22 to 3.7 million on October 1, with a peak of 484,000 requests in a two-hour window on October 3. Traffic rose from 08:00 UTC, peaked at 16:00 UTC and then fell, a pattern the report ties to people using the product rather than a scheduled crawler. Separately, Sensor Tower estimated more than 5 million installs in the US and Canada within 22 days. Because DataDome had only nine September days of Muse data, it treats its September figures as floors.

Two method points matter. First, DataDome says Muse "doesn't declare itself"; its press release says it does not identify as an AI agent. DataDome does not disclose how it recognized Muse traffic. PPC Land infers fingerprinting (identifying a browser session by its technical traits rather than the name it gives) or behavioral analysis, but that is an inference. Second, DataDome gave no sample size for the sensitive-flow figures below; PPC Land's own calculation suggests about 2.8 million sampled requests.

Cloud, not couches

Because Muse runs in a dedicated cloud virtual machine, DataDome says 99.3% of sampled requests came from cloud or content delivery network (CDN) infrastructure: 81.2% through Cloudflare and 18.1% through Fastly. Only 0.6%, about 17,500 requests, came from consumer ISPs such as Comcast, Verizon and Rogers. That contradicts an October 11 claim from Profound that Muse uses residential connections; Profound published no traffic data for it.

Much bot defense still assumes real shoppers arrive from home IP addresses and bots from data centers. The report's author, Jerome Segura, said: "Policies that trust residential or ISP-based traffic won't catch agents like this." The reverse risk is just as real: a filter built on that rule can mislabel a real customer's agent as data-center noise and block it.

robots.txt (the file that tells automated crawlers which parts of a site they may visit) gives precise control over OpenAI, Anthropic and Amazon bots, which use separate names for training, search and user-requested fetches. But that control only works on agents that declare themselves. Browser-based agents have no such name to write a rule against.

What the traffic touched

About 8.9% of sampled Muse requests reached what DataDome calls sensitive flows: roughly 202,600 forms, 24,800 logins, 9,600 payments, 6,800 account creations and 5,200 carts. These figures describe contact with those pages. The report does not say how many actions completed, failed or were blocked, so this is a map of where the agent went, not of what it bought.

DataDome also separates volume from value. Meta's 12.4 billion requests produced 51,000 Meta AI referrals, about one per 243,000 requests. Its framing: "an infrastructure cost, not a business signal." The merchant pays to serve the traffic whether or not a sale follows.

The industry is building for agents that raise their hand

American Express released its Business Playbook for Agentic Commerce on October 6, including Agent Purchase Protection for eligible charges caused by errors from registered AI agents. Yahoo Finance called it "an announced intent, not a live, functional product"; terms, including what counts as a registered agent, are undefined. TechCrunch covered the framework on October 13. Mastercard is testing in the US a score that estimates whether a transaction originated from an AI agent, announced in September.

Both efforts assume you can tell an agent from a person, ideally because it says so. Muse, as DataDome measured it, did not.

A separate survey shows the same gap from inside risk teams. Media Copilot reports a Fingerprint survey of 300 US financial services leaders, June to August: 79% had detected malicious bots impersonating legitimate AI assistants or search crawlers, and 95% were confident they could tell real assistants from impostors, yet only 45% said they detected and mitigated spoofed traffic in real time. 61% still rely on client-side JavaScript tags or native SDKs, and only 26% had full visibility into traffic that bypasses JavaScript-based detection. 40% mainly feared losing customers by blocking a legitimate assistant. Fingerprint CEO Dan Pinto: "Many AI agents circumvent this model, so the challenge now is knowing whether automated traffic is actually what it claims to be." The two studies use different methods and populations, and neither establishes how often impersonation hits publishers or retail checkout.

On the merchant side, the checkout layer is already live. Muse launched with Stripe's Link checkout at more than 1 million businesses, per FinTech Weekly. A Checkout.com merchant sample in the UK and US, with no disclosed sample size or method, reports 3% of transactions involve agents, 89% of merchants are preparing and 42% are testing. Fed Governor Christopher Waller said the biggest barrier to the agent-delegated model is "building sufficient trust among buyers and sellers."

What is unproven

This is one vendor's telemetry from its own customer base. DataDome does not state its network size, so a change in customer mix could move growth rates. The 33% share rests on nine days of data after a window DataDome starts on its own definition of "full availability." The method for identifying Muse and the sample behind the sensitive-flow numbers are unpublished. And DataDome sells bot management to the merchants it is warning; it joined Experian's Agent Trust programme on October 1. None of that makes the findings wrong. It means 33% is not a stable, verified measure of Muse's share of the agentic web, only the clearest public signal that it is large and undeclared.

What merchants can do now

DataDome's recommendations, minus the sales pitch, are reasonable: classify agents by function; set explicit rules for forms, login, cart and checkout; make a deliberate decision about Muse instead of letting default settings decide; and stop treating the user agent (the self-reported browser name any client can set to anything) as proof of identity. The report points to Web Bot Auth, an emerging standard for cryptographically verifying a bot's identity, plus IP range checks, as more durable than residential-versus-data-center heuristics.

On Monday, pull 30 days of server logs, filter for requests from Cloudflare and Fastly IP ranges that touched your login, cart or payment endpoints, and count how many got through under rules written for human traffic.

Sources