The agent didn’t get in: why Amazon blocked Meta’s Muse AI agent

On September 8, Meta launched Muse, a personal AI agent (software that doesn’t just answer questions but carries out tasks itself: booking, paying, negotiating) ¹. Muse opens a browser, fills out forms, and — with a user’s approval — completes purchases. Within a week it topped the free app chart in the US App Store ¹, and according to an estimate by analytics firm Sensor Tower (a third-party estimate, not an official Meta figure), it was downloaded roughly 2.5 million times within two weeks ¹³. On the evening of Sunday, September 20 (US time), Amazon started showing Muse users a popup: “Continued access by an unauthorized AI agent violates Amazon’s Conditions of Use” ¹. Shopping on Amazon through Muse is now blocked.

Amazon shut the door while others hold it open

Amazon cited three reasons for the block: Meta never disclosed that Muse would access the store, the agent doesn’t identify itself as AI while browsing, and, Amazon says, it appears to capture and store user credentials ³. Meta had said the opposite at launch: Muse “has no visibility into people’s passwords or payment methods” — credentials go into “secure storage,” so Muse can use them without seeing them ¹. Muse is free at the basic level, with paid subscriptions costing $20 and $100 a month ⁸; per Digiday, the seller — not the buyer — pays the commission on purchases made through Muse.

This isn’t Amazon’s first attempt to keep an outside agent off its territory. In November 2025 it sued Perplexity over the Comet browser and its built-in shopping agent, won a preliminary injunction in March 2026, and on August 4, 2026 the Ninth Circuit Court of Appeals vacated that injunction and sent the case back for further proceedings, finding that Amazon was “unlikely to succeed” on the merits: it is the user, using Perplexity’s tool, who “accesses” Amazon’s servers — not the company itself — so the federal Computer Fraud and Abuse Act (CFAA) likely doesn’t apply ⁴. The court separately noted that its ruling left untouched a site’s own terms of use, which a company writes itself ⁴. That’s exactly what the Muse popup cites ³. eBay added an outright ban on unauthorized third-party chatbots and buy-for-me agents to its user agreement back in February 2026 ³ — same logic. Walmart, Target, Best Buy, Home Depot, Macy’s and Etsy went the opposite way, integrating directly with ChatGPT Shopping, Perplexity and Microsoft Copilot ³. In our view, retail has split not by company size but by how dependent each business is on advertising sold on its own pages.

Why this is really about advertising

In 2025, advertising brought Amazon more than $68 billion — roughly a tenth of the company’s total revenue [1; 2]. That sum depends directly on shoppers browsing pages themselves, seeing sponsored products, and clicking them. An agent that searches, compares and checks out off-page takes away exactly the ad inventory (the ad placements on its pages) that Amazon sells to advertisers ³. Emarketer analyst Max Willens put it plainly: for years, Amazon was where most US consumers went first to buy something online; increasingly, they now start with an AI assistant instead ⁵. A September 2026 PYMNTS Intelligence study found that nearly 132 million US adults have already bought something with AI help, and 22% now begin product research inside an AI tool [3; 6]. Yet 59% of AI-assisted purchases still end on Amazon ⁶. Even an open agent, in other words, often just leads the shopper to where they’d have ended up anyway — and that’s precisely the inbound flow Amazon wants to keep under its own control.

Meta, for its part, says Muse carries no ads: according to Mark Zuckerberg at Meta Connect on September 23, monetization will come through a small transaction fee, not an ad auction (the bidding system in which advertisers compete for each impression) [7; 8]. But Digiday points out a structural problem: if Muse scales, a banner ad simply won’t work inside it — an agent doesn’t “scroll” a page the way a person does, it reads structured product data (specs, price and availability in a machine-readable format) and moves on ⁸. So any advertising model that eventually emerges will be about something different. Digiday imagines what that could look like: you couldn’t buy the agent’s actual answer, only a “sponsored” placement next to it ⁸. We see the real risk of the coming quarters here: advertisers may end up paying not for a person’s attention but for a slot in a list a machine reads — and no platform has yet said clearly what that will look like or cost.