Meta's shopping agent lands as direct traffic from Linux Chrome

Meta’s Muse shopping agent has no published user agent, IP range or robots.txt token, so the visits it makes to retail sites arrive in analytics as ordinary desktop Chrome traffic. In Botify’s writeup, Jordan Long, who runs the North America services team at Botify, which sells an enterprise SEO platform, reports that live Muse sessions captured by the bot detection vendor DataDome came from desktop Chrome on Linux, on Cloudflare and Fastly addresses rather than Meta’s own network.

Muse launched on September 8 and gives each user a dedicated cloud machine running full Chromium, so page scripts execute exactly as they would for a person and Google Analytics, Adobe Analytics and pixels all fire. GA4’s built-in bot filter mostly catches bots that announce themselves, and Muse announces nothing, so the sessions pass into reports untouched. By HUMAN Security’s count, Muse is now about 72% of the AI-agent requests the fraud detection firm sees, with 84% of them landing on product and search pages. The agent reads the accessibility tree rather than the rendered page and cannot run scripts of its own, but as far as measurement goes it behaves like a browser with a person in it.

Botify’s post splits Muse activity into three groups by how it reaches the site:

• Direct traffic from agent browsing, which is most of it. Muse opens a product URL from a catalog or a search lookup, carries no referrer, and gets logged as direct. An agent comparing several stores buys from one at most, so direct volume rises while the direct conversion rate falls.
• Referral traffic, which only appears when a person taps a link on the muse.ai website. Clicks from the mobile app usually arrive with no referrer and land as direct. GA4’s AI Assistant channel, added in May, catches only the visits that carry a referrer.
• Orders with no session, which happen when Muse buys through a connector or a payment service such as Shop Pay or Link by Stripe. The agent skips the checkout pages, no tags fire, and the order lands in the commerce platform with nothing behind it. Shopify merchants get Meta sales-channel attribution for those orders; everyone else sees revenue with no source.

Botify lists the symptoms without ranking them. Ranked by what they cost a business, the gap between the groups is wide. Inflated direct traffic and a sagging direct conversion rate are a reporting annoyance a team can carry for a quarter. Orders with no session are revenue that cannot be attributed to anything at all, and HUMAN Security found Muse 25% more likely than other agents to reach a checkout page, so the group hardest to see is the one carrying the money. Analytics teams have had a thinner version of this problem since ChatGPT clicks started arriving as direct traffic, but those sessions still had a buyer at the end of them.

What to do

Start with the baseline, because the remaining steps all need a number to work against. In GA4 or its BigQuery export, isolate desktop sessions from Chrome on Linux inside direct traffic and compare the weeks either side of September 8. A sudden jump in volume, or one browser version taking over that segment within days, is the pattern Botify says to look for: Meta updates its whole fleet at once, and no population of real shoppers changes version together.

Server or CDN logs cover the part analytics cannot. Muse can also fetch a page’s text without running JavaScript, and those fetches reach the server and never reach GA4, so log data is the only place they exist.

Reconcile checkouts on the commerce platform rather than in the tag. Record the order source there and push agent purchases into analytics through the Measurement Protocol or the data warehouse. Without that step, the connector orders never reach analytics.

Watch out for

Filtering on Meta’s published IP ranges will not find Muse, since those ranges cover Meta’s declared crawlers and fetchers and the Muse sessions come from Cloudflare and Fastly. Blocking those ranges instead would hit real people: Botify notes the same Cloudflare addresses carry consumer VPN and privacy traffic, so a datacenter IP is a clue and not proof. Neither will robots.txt help, because there is no Muse token to write a rule against. Detection that works today is pattern-based, and several bot management vendors now flag Muse by name. When one of them does, label the session in the tag manager, set its source and medium, and send that medium to a dedicated agent channel so agent sessions and agent revenue report on their own lines.