An AI assistant may understand a shopping request and still be unable to complete it. TechCrunch’s reporting on website access shows why: retailers, airlines and other services do not share a single policy for agents acting on behalf of customers. Some deliberately restrict them, while ordinary bot defenses can interrupt others without an explicit decision to exclude AI.
That distinction is becoming important as personal assistants promise to handle purchases and bookings. A failed transaction does not by itself reveal whether a website rejects the agent, requires a commercial agreement or simply cannot distinguish it from unwanted automated traffic.
A Blocked Purchase Does Not Always Mean a Ban
Amazon has recently blocked Meta’s Muse assistant from browsing its retail catalog and buying products, according to TechCrunch. Complaints about Muse encountering problems at Walmart, however, led to a different explanation. Walmart told the publication that those failures were not intentional and that it wants to work with partners, including Meta.
A human-verification prompt can interrupt an agent’s attempt to finish a transaction. If the assistant cannot complete that check, it may be removed from the session. That technical barrier can look like a policy decision to the customer, even when the retailer describes its position as welcoming rather than prohibitive.
Airlines illustrate another uncertainty. Delta spokesperson Heena Chavda told TechCrunch that the company protects its digital channels against unauthorized automation and currently has no partnerships or integrations allowing third-party agents to shop or book there. Delta is evaluating the security and customer-experience implications.
United pointed to terms restricting unauthorized automated access without written permission, but did not answer a follow-up about which agents it blocks. That response establishes a contractual restriction, not a complete account of how every assistant behaves on United’s website.
Authorization Depends on the Website and the Activity
Other companies draw boundaries around agreements and particular uses rather than declaring every agent unwelcome. Yelp told TechCrunch it allows nonhuman traffic through formal, paid data-licensing arrangements. An assistant without such a relationship may therefore encounter a barrier even if a person would freely browse the same information.
EBay said it does not prohibit all third-party shopping agents. Its restrictions concern unauthorized activities, including scraping and model training. This is a materially different position from treating any automated purchase attempt as forbidden.
The reporting also describes user complaints involving other websites, but those accounts do not establish the companies’ policies. Adidas, Zillow and Pizza Hut either did not respond or declined to comment. Their silence should not be read as confirmation that they intentionally blocked the assistants mentioned by users.
Cloudflare’s protections add another possible layer. Changes affecting AI traffic could influence access on sites using its services, but the company had no specific data to share about the reported failures. Its bot statistics cannot separate beneficial agents from harmful ones. The article therefore does not demonstrate that Cloudflare caused any particular unsuccessful transaction.
A Shared Standard Is Still Being Developed
Meta is working with Walmart, Stripe, Sierra, Genesys, Rocket, NiCE and Decagon on an open standard for agent-to-agent communication in online commerce. The effort aims to help distinguish user-authorized activity from malicious bots, according to Meta’s account cited by TechCrunch. Work remains at an early stage; the report does not describe a finished standard adopted across the web.
The proposed distinction addresses a problem that the company responses make concrete. A customer can authorize an assistant to act, yet the destination website may still require its own permission, commercial relationship or supported integration. The customer’s request and the service’s access rules are separate parts of the transaction.
Meta argues that turning away an agent effectively turns away the customer behind it. Websites’ statements show that they also consider unauthorized automation, security and permitted data use. Those positions leave room for cooperation, but do not imply agreement on what access should be granted automatically.
For now, the limiting factor is not simply whether an assistant can navigate a page. It is whether that page recognizes the assistant’s authority and supports the requested action. TechCrunch’s findings reveal a mixed landscape of deliberate restrictions, conditional permissions and technical interruptions—not a universal ban or a universal right of entry.
