Someone asks Perplexity's Comet browser to reorder the same detergent at the best available price. The agent visits three retailer pages, extracts price and availability from each, and completes the purchase. None of the three pages' promotional banners, hero images, or limited-time framing ever entered the decision, because the agent was never looking at the page the way a person would.
The advertising industry built its modern business model around a simple premise: a human looks at a page, notices something, and clicks. An agent works from an explicit goal, cheapest option meeting these constraints, and evaluates against price, spec, and availability rather than against the emotional or visual cues advertising has spent a century learning to deploy.
Publishers and advertisers whose revenue depends on impressions from human attention are the most exposed, because agentic browsing does not generate impressions the way a human scrolling a page does. It generates structured data extraction, often without rendering anything an ad server would register as a valid view.
Attribution breaks first. The systems that track whether an ad influenced a purchase were built around human browsing behavior, and they have no clean way to credit or even detect an agent's decision process. A merchant can win an agent's comparison on price and pay zero advertising spend in doing so, while a merchant that spent heavily on brand advertising can lose the sale simply because an agent never rendered the page carrying that message.
The Interactive Advertising Bureau has begun publicly grappling with how to even classify agent-driven traffic, since existing bot-detection frameworks were built to filter out fraud, not to price a legitimate new category of buyer. Amazon's own Rufus assistant, built directly into the platform its competitors' products live on, gives Amazon a structural advantage in deciding what an agent even sees as an option in the first place.
What is opening up is a scramble for new surfaces to influence: bidding not for a human's attention but for placement within structured feeds and product APIs designed to be legible to automated buyers, and early experiments in paying for preferred treatment inside an agent's comparison logic itself, which raises its own disclosure questions.
The businesses most at risk are those whose entire value proposition to customers was mediated by successful persuasion rather than a genuine price or quality advantage, the layer of the market advertising has always subsidized. An agent that compares on stated criteria rewards the underlying product and has no mechanism to reward the marketing that used to compensate for its absence.
Advertising reinvented itself for broadcast, then search, then social, each time finding a new bottleneck to sit inside. Comet and Atlas are not a new medium to buy space in. They are a browser built specifically not to notice the space at all, and nobody selling ads has found the equivalent move yet.