AI shopping agents promise to search listings, compare details and help people find products. But their usefulness depends on retailers letting them into the places where those products are sold. A recent change in what Meta’s Muse could retrieve from eBay puts that access question in focus.
Muse stopped returning eBay listings
A reader told The Verge that Muse had been assembling a daily digest of watches listed across sites including eBay. The digest included links, prices, shipping details and notes about condition or style. After several weeks, the agent stopped pulling eBay links and details, saying it was blocked from the site.
The Verge reproduced the problem with other search requests. In some cases, Muse sent users to outside sites that mirrored eBay listings instead. That workaround can point shoppers toward information about products, but it is different from retrieving the listings directly from eBay.
eBay did not confirm whether it had recently introduced new limits on AI agents. Spokesperson Sybille St. Arromand said the company takes a selective approach to how third parties interact with its marketplace. She said eBay has restricted access in some cases, including unauthorized scraping and training on its data.
Browsing and buying are different questions
eBay had already updated its user agreement to prohibit tools such as “buy-for-me agents” and “LLM-driven bots” from placing orders without human review. Until recently, agents like Muse could still browse the site and retrieve listings for people. The reported change suggests that access for browsing may also be constrained in some circumstances, but eBay did not say directly whether its rules had changed.
The company says AI is becoming part of discovery, personalization and selling on its own platform. St. Arromand said eBay is discussing potential partnerships and evaluating how third-party agents might help new customers find its inventory. She also said the company wants to preserve the experience buyers and sellers expect.
That leaves a distinction between an agent that helps someone search and one that acts on their behalf. Retailers may see value in outside tools sending customers to their products, while still wanting control over automated access, data use and purchases. The terms eBay gave point to those as separate concerns.
Retailers are choosing different levels of access
Online marketplaces and service providers have taken varied approaches to third-party agents. Amazon blocks external agents from browsing or shopping. Target has restrictions that let agents browse in some cases but block them in others. Shopify has opened its catalog of a billion listings so shoppers can configure agents to buy items for them.
These different policies show that “AI shopping” is not one uniform experience. An agent’s ability to find a product can depend on the retailer, the way it accesses a site and whether it is only gathering information or attempting to complete a transaction.
There are practical reasons for retailers to be cautious. Many shopping sites were built around people browsing at human speed, not large numbers of automated agents. The Verge previously found that a Muse agent could repeatedly add an item to a cart on The RealReal, effectively keeping a product listing unavailable to other shoppers. The RealReal changed its bot system after that report.
How much agent-led shopping is happening?
Retailers are adapting their defenses as they encounter automated activity, but the scale of agent-driven commerce remains uncertain. The examples in the article suggest that interest in agents is ahead of their use for many real purchases.
Uber said third-party agents completed only four rides in Q3 and around 500 Uber Eats orders during the week of Halloween. DoorDash reported “several hundred” orders placed through third-party agents, compared with 993 million total orders last quarter.
For now, shoppers may find that an AI assistant can help with some parts of a purchase while failing to access particular listings or retailers. The eBay episode shows that the promise of an agent depends not only on what the software can do, but also on whether marketplaces allow it to see and interact with their inventory.