There is no reason to ban Perplexity’s AI agents from Amazon’s marketplace. This was the conclusion reached, a few weeks ago, by the United States’ largest federal appellate court.
Set against the backdrop of a landmark dispute in the world of agent-driven commerce. The focal point: the Comet browser, to which Perplexity began granting access in July 2025.
In the run-up, Amazon had warned: in broad terms, the integrated assistant must not access its marketplace. The threat sharpened in November with a cease-and-desist letter. Amazon urged Perplexity not to “sneak into” its e-commerce perimeter by “masquerading” its agents as Chrome instances. It demanded transparent identification, invoking the rights of service providers to supervise agent activity, in the name of preventing behaviors that degrade the shopping experience and customer trust and that could create risks for users’ data.
Perplexity called this a “threat to all Internet users.” In its view, a company would have no right to discriminate among users based on the AI they choose to represent them. Amazon, it argued, should welcome a simpler shopping experience, which translates into more transactions and happier customers. If it did not adopt that stance, the argument went, its interest lay in serving you advertising, sponsored results, and steering your purchasing decisions with “confusing” offers …
Amazon ultimately filed suit, alleging a violation of the Computer Fraud and Abuse Act (which punishes certain intentional intrusions into computer systems). In March 2026, it obtained a victory. The court granted the requested injunctive relief. Perplexity appealed promptly, which had the effect of suspending the enforcement of the judgment.
Perplexity did not access Amazon’s servers
The Court of Appeal overturned the trial court’s decision. The reasoning boiled down to this: it is not the AI assistant that accesses Amazon, but the user.
Perplexity framed its defense around this point. According to them, their systems never directly access Amazon’s servers: they access data that is sent to the user’s computer. A distinction that the jurisprudence takes into account, referencing the Meta v. BrandTotal case. In that instance, a company exploited a browser extension that captured data while users browsed Facebook. In that sense, the court noted, it did not access Meta’s servers, but it did access data sent to users.
According to Amazon, the parallel with the BrandTotal case is not apt: here we are dealing with active navigation rather than passive data collection. As for the agent’s actions, they should be attributed to Perplexity, since it is their servers that control the agent.
Perplexity does not view it that way. They contend that their assistant has no independent intent; its intent is that of the user.
The EFF, Mozilla, and EleutherAI, among others, filed as amici curiae in support of Perplexity. Their conclusion: Perplexity’s servers never directly access Amazon’s servers.
“Computer” does not imply “federal crime”
The Court of Appeal reasoned more straightforwardly: the language of the Computer Fraud and Abuse Act implies that an assistant cannot “access” a server since it is not a person, but a tool.
It also noted that ruling in Amazon’s favor would amount to endorsing a novel interpretation of the statute, aimed primarily at punishing hacking. Such a reading would also expose Comet users to criminal penalties for facilitating unauthorized access.
The Court acknowledged more broadly that precedents call for caution when it comes to reclassifying harmless conduct as federal crimes simply because a computer is involved.
It likewise rejected the lower court’s finding about irreparable harm. The argument about a degradation of the shopping experience appeared too abstract, as did the question of how much responsibility users would assign to the assistant.
* The Court also relied on a decision it had issued in a dispute between Facebook and Power Ventures. Power Ventures operated a social-network aggregator. One of its referral campaigns prompted messages to be sent to the Facebook friends of users. Power Ventures was deemed responsible for not terminating those actions, given its evident access to Facebook’s servers.
