AI Agents Do Not See Ads: What Agentic Shopping Means for Brand Marketing Budgets

A US appeals court ruling this month settled a narrow legal question about web scraping. Buried in the arguments was a much larger commercial one, and most businesses have not registered it yet: the advertising model that funds online retail assumes a human is looking at the screen.

Increasingly, nobody is.

What the court decided

Amazon sued Perplexity in November 2025, seeking to stop the startup's Comet browser agent from shopping on its site. Comet can be pointed at a product and will then navigate Amazon and complete a purchase on the user's behalf.

In March 2026, a district judge granted Amazon an injunction, finding that Comet had accessed Amazon's site at the user's direction but without the retailer's authorisation. Amazon had argued the agent violated the Computer Fraud and Abuse Act, a law written decades ago for a very different kind of intrusion.

In August the Ninth Circuit overturned that injunction. The reasoning matters more than the outcome. Because Comet acts only when a user tells it to, the appeals court concluded that it is the user, not Perplexity, who is accessing Amazon's servers. A tool doing a person's bidding is legally that person.

That removes one of the sharpest instruments platforms had for keeping third-party agents out. As eMarketer noted, Amazon has been fighting to keep agents off its site because it wants to control the shopping experience, including the advertising shoppers see.

The argument nobody expected to be about advertising

Amazon's public case rested on security. Its filings said Perplexity's agents could act inside protected systems, including private customer accounts behind passwords.

But the complaint also made a second argument, and it is the one businesses should read closely. Amazon told the court that AI-generated traffic created problems for its advertising business, because when AI systems produce ad impressions, those impressions have to be identified and filtered out before advertisers can be billed. Building that detection capability required changes to Amazon's advertising systems, which the company said were necessary to meet its contractual obligations to advertisers.

Perplexity’s response was blunter. In its appellate filing the company argued that "AI agents don’t have eyeballs to see the pervasive advertising" that Amazon shows its users, and suggested protecting ad revenue was the real motivation for the lawsuit.

Set the legal dispute aside and a plain commercial fact remains. Amazon generated roughly $68.6 billion in advertising revenue in 2025. That money comes from brands buying placement in front of shoppers. An agent completing a purchase does not view a sponsored placement, does not compare it against an organic result, and is not influenced by the creative work the placement paid for.

Every business currently funding visibility on a marketplace is funding a mechanism that assumes human attention.

Why this is a strategy problem, not a technology problem

It is tempting to file this under emerging technology and revisit it in two years. The reason not to is that the intermediation is already happening at scale.

Amazon's own assistant, Rufus, drove nearly $12 billion in incremental annualised sales in 2025, according to figures the company disclosed in February earnings materials. Amazon has also blocked dozens of outside agents from its platform, including OpenAI's, while continuing to build its own.

That combination describes the actual competitive situation. Platforms are not resisting AI-mediated shopping. They are resisting AI-mediated shopping they do not own. A marketplace that runs the assistant keeps the interface, the data, and the advertising inventory. A marketplace that lets someone else’s agent in keeps only the transaction.

For a brand selling through those platforms, the distinction is not academic. It determines whether the money spent on visibility reaches a person, an algorithm, or nothing at all.

What actually changes for brands

Three things, in ascending order of difficulty.

Product data starts doing the work that copy used to do. An agent parsing a request for a waterproof jacket under a certain price, delivered by a certain date, is matching structured attributes. It is not reading persuasive product descriptions or responding to lifestyle photography. Businesses that have treated product feeds as a compliance chore rather than a marketing asset will find that the chore has quietly become the marketing.

The measurement layer stops describing reality. Click-through rate measures a human interacting with a page. When intent is parsed into structured constraints and resolved without a page view, the metric survives while the behaviour it describes does not. Marketing teams will keep receiving numbers that look normal for some time after they stop meaning what they used to.

Advertising budgets need a stated position. This is the hardest one, because there is no settled answer yet. Some share of marketplace traffic will continue to be human and will continue to respond to paid placement. Some share will not. Nobody currently knows the split, and the platforms with the best data on it are also the parties with the strongest interest in how it is reported.

The practical response is not to cut spending on a guess. It is to know what the current spend is actually buying, at a level of detail most businesses do not have to hand. That means separating advertising performance by placement and by intent type, and being able to see when the return on a placement moves for reasons that have nothing to do with the creative or the bid. It is unglamorous analytical work, and it is why agencies that manage marketplace accounts for brands are fielding a version of this question from clients well before anyone has a confident answer.

The case for not panicking

There is a reasonable argument that this is all moving slower than the coverage suggests.

Amazon’s chief executive Andy Jassy has said agentic commerce has a chance to be really good for e-commerce, while adding that agents are not yet good enough at personalisation and pricing accuracy. That is a fair assessment of the current state. Anyone who has watched an agent select a plausible but wrong product knows the technology is not finished.

Consumer adoption is also not guaranteed to follow the trajectory the industry expects. People delegate booking and reordering more readily than they delegate considered purchases. A category where shoppers want to see the thing before buying it may be intermediated far more slowly than one where they do not.

And the legal position is unresolved. The Ninth Circuit lifted an injunction; it did not decide the underlying case, and platforms have other instruments beyond the statute Amazon chose.

What to watch

The useful signals over the next year are not about the technology.

Watch whether platforms open formal agent access on their own terms rather than fighting it in court, because a paid, permissioned channel for agents is the outcome that preserves the advertising model. Watch whether any large retailer publishes a breakdown of agent-originated versus human-originated traffic, because the first credible number will reset a lot of budget assumptions. And watch the terms of service rather than the press releases, since that is where the operative rules will change first.

For most businesses the correct action right now is neither a reallocation nor a rebuild. It is making sure the product data is accurate and complete, and knowing precisely what current advertising spend returns, so that when the split between human and machine buyers becomes measurable, the comparison is available rather than reconstructed after the fact.

The court settled who is allowed in the door. It did not settle who pays for the room.