Your Site Can't Tell Which AI Agents Are Shopping
Meta's Muse made up 72% of the AI agent requests HUMAN measured, and it runs in a regular browser, so blocking agents by name misses it.
On September 28, HUMAN Security reported that Meta's Muse averaged 72% of the daily AI agent requests it observed. That is 5.2x the volume of ChatGPT Agent, the next largest. Of Muse's requests, 84% went to product and search pages.
Those numbers are impressive, to say the least, and with Muse passing 10 million set ups in the US and Canada in just 6 weeks, there’s reason to believe we’ll continue to see growth there. Whether it’s Muse, ChatGPT, or some other agent platform, it’s clear that AI agents are already on commerce sites in volume. The tricky bit for businesses is that many agents don't say who they are.
A caveat before we build plans around these numbers. HUMAN publishes no request counts, sample size or observation window, and it sells bot management, so treat the 72% as directional. It is also a share of the agent traffic HUMAN can see, not of all traffic to your site, and none of the reports we found measure agents as a share of total site traffic. The direction is consistent across vendors, though the denominators differ: DataDome’s Q3 report found Muse was 33% of September’s agentic browser traffic, and its first 12 days of full availability produced 29.1 million requests. Either way, these are big numbers, which is why it matters whether you can identify them.
Muse doesn't look like a bot
According to DataDome, Muse runs in a cloud virtual machine with a browser, so its traffic looks like browsing rather than crawling. The report names no bot identifier for it. In DataDome's sample, 99.3% of Muse requests came from cloud and CDN infrastructure, which means rules that trust residential or ISP traffic won't catch it. Rules that block a named bot won't catch it either, because there is no name to block.
The same sample showed 8.9% of Muse requests reaching sensitive endpoints: login, payment, account creation, cart and forms. Bot-management vendor Cequence reported that it found Muse-matching traffic at more than half the customers it studied, and says Muse presents as a standard Chrome browser and doesn’t sign its requests.
What Meta says comes next and why you need a doorman, not a bouncer
Retailgentic's account of Meta's Q3 call says CFO Susan Li expects the market to move away from agents that screen-scrape like a human toward authenticated APIs and approved agent partnerships. Meta also said it plans to charge businesses a fee on transactions Muse facilitates. If that is the direction, the open question for a merchant is which agents get access, and to what.
A bouncer decides at the door, usually by name. A doorman knows who is coming in and what they are allowed to do once inside. HUMAN's own advice to merchants points the same way: set rules by activity rather than by agent alone, allowing useful automation while blocking risky actions such as adding to cart or logging in.
That approach only works if your site has separate surfaces to apply rules to.
Where architecture comes in
In a composable stack, catalog, pricing, cart and checkout are separate services behind APIs. A merchant can offer read access to catalog data to any verified agent, require authentication for cart, and require stronger authorization for checkout. A storefront that serves everything through one page flow can only allow or block the page. A storefront built around one page flow can still block or allow pages, but it has fewer clean points to apply different rules to browsing, cart and checkout.
What to do now
- Measure it. Separate agent traffic from human traffic using behavior and infrastructure signals, not only the user agent string.
- Set policy by action. Decide what unverified agents can do on read, cart and checkout, and write it down.
- Offer an authenticated path. Give approved agents API access so they have a sanctioned alternative to scraping your pages.
- Watch agent identity standards. HUMAN backs Web Bot Auth, a signed-identity approach, and notes that most AI providers don't use it yet. We don't know that merchants can reliably verify Muse today.
- Read vendor data with care. The companies measuring agent traffic also sell the tools to manage it.
The harder question is commercial. Once agents are a real share of shoppers, deciding who gets in and what they can do becomes a merchandising decision as well as a security one, and your architecture determines how many options you have.
Leigh Bryant
Editorial Director, Composable.com
Leigh Bryant is a seasoned content and brand strategist with over a decade of experience in digital storytelling. Starting in retail before shifting to the technology space, she has spent the past ten years crafting compelling narratives as a writer, editor, and strategist.