The Agent Doesn't Care Who Paid for the Ad

Based on the research ofRothschild, Immorlica, Lucier, Mobius, and Slivkins, "The Evolution of Digital Search: From Blue Links to Delegated Decision-Making," arXiv preprint, 2026

The Auction Never Priced the Product. It Priced the Click.

Google's ad auction, Amazon's sponsored shelf, and twenty years of affiliate-link publishing all monetize the same resource: a human being's attention as it moves down a list. The product being sold to advertisers was never really the product; it was position, and position only has value because a person is standing there, scanning top to bottom, more likely to click the first plausible thing than to scroll to the fourth. Strip out the human and you strip out the reason anyone would pay for the first slot.

That dependency is already visible in how fragile click behavior turns out to be. Pew Research found that searches which triggered an AI Overview saw only 8% of users click through to any linked source, compared with 15% in ordinary results without a summary, according to reporting aggregated by eMarketer in August 2025. Separately, publisher trade group Digital Content Next measured a median year-over-year drop in referral traffic from Google of about 10% across premium publishers over an eight-week stretch in 2025, with declines running as high as 25% at the extremes, non-news publishers losing roughly 14% on average. None of this required an agent that transacts; a summary box that merely satisfies the query before a human clicks anything already cuts the audience the auction was built to sell. The agentic layer the paper describes is a much larger version of the same mechanism, because an agent does not just skip the click, it replaces the human's judgment entirely.

An Agent Loyal to You Has No Reason to Click the Sponsored Link

This is the pivot the paper is built around: once an AI agent interprets your goal and compares, negotiates, or transacts on your behalf, it has no structural incentive to respect a ranking someone paid for. A human can be swayed by a bolded "Sponsored" tag sitting in the top slot because a human is pattern-matching on trust and convenience under time pressure. A shopping agent optimizing for price, fit, and your stated preferences has no equivalent soft spot, unless its maker deliberately builds one in.

The industry is already running this experiment in public. Perplexity shipped Comet, an AI-native browser whose built-in Comet Assistant can read a page, navigate sites, and complete multi-step tasks on a user's behalf, launching in July 2025 according to TechCrunch's hands-on review. OpenAI followed in October 2025 with Atlas, a browser with an agent mode that, per MIT Technology Review's testing, could be asked to fill a shopping cart unprompted and act on inferred intent. Amazon's Rufus, meanwhile, is no longer just an answer box: Modern Retail reported in April 2026 that monthly active users were up 115% year over year and engagement up 400%, with Amazon CEO Andy Jassy telling analysts Rufus had driven nearly $12 billion in incremental annualized sales and that users of it were 60% more likely to complete a purchase. Each of these agents is, in some form, the thing the paper describes: a system that interprets intent and executes a decision rather than handing back a ranked list for a human to parse.

The tell is in how OpenAI chose to describe its own commerce feature. When it launched Instant Checkout in ChatGPT in September 2025, built on an Agentic Commerce Protocol developed with Stripe, OpenAI stated plainly that merchants pay a small fee on completed purchases but that this "doesn't influence ChatGPT's product results," adding that "Instant Checkout items are not preferred in product results," per Thurrott's coverage of the launch. That is a direct, public disclaimer of the sponsored-rank logic that funds the rest of the search economy, offered because OpenAI understands that an agent caught favoring a paying merchant over the best match for the user would break the one thing that makes a shopping agent worth using at all: the user's belief that it is on their side.

Governance Is Already the Fight, Not a Future One

The paper's central claim is that the determining variable is not better ranking algorithms or smoother natural-language interfaces, it is whether the agent layer itself ends up open, transparent, and competitive, or captured by a small number of gatekeepers. That fight is not hypothetical; it is the actual content of product announcements happening right now. OpenAI open-sourced its Agentic Commerce Protocol so any merchant or developer could integrate with it. Four months later, at NRF 2026, Google announced a competing Universal Commerce Protocol, built with Shopify, Etsy, Walmart, and Target and endorsed by more than twenty companies including American Express and Visa, covering discovery through checkout within AI Mode and Gemini, according to eMarketer's reporting from January 2026. Two different standards, from two different gatekeepers, both racing to become the rails that merchants plug into before the other one locks in the market.

An agent that is actually loyal to the user has no reason left to respect a ranking someone paid for.

Amazon is pointedly not joining either camp. Jassy told analysts in that same reporting period that "the experience just hasn't gotten great with these third-party horizontal agents yet," arguing they lack personalization data and shopping history, and predicted that shoppers "going to a particular retailer they like to do business with" would stick with retailer-built tools like Rufus rather than a neutral intermediary. That is a platform explicitly trying to prevent an open, portable agent layer from forming at all, because an agent a user trusts to compare across every retailer is exactly the thing that erodes Amazon's advantage as a closed destination. eMarketer also noted that Amazon's own "Buy for Me" agent frustrated both shoppers and third-party sellers over the 2025 holiday season when AI-generated listings didn't match real inventory, a reminder that even the incumbents are not close to having solved this cleanly, let alone agreed on who should govern it.

The Numbers Are Still Small, Which Is the Window That Matters

None of this has fully arrived yet, and that is precisely why the paper's governance argument is worth taking seriously now rather than after the fact. Salesforce data cited in eMarketer's January 2026 reporting found that AI and agents influenced about 20% of global retail sales over the most recent holiday season, but mostly in a customer-service capacity rather than completing purchases outright, and eMarketer's own forecast puts the number of US consumers expected to use AI platforms and assistants to shop this year at 63.3 million. That is a meaningful and fast-growing slice of commerce, not yet a majority of it. The protocols, disclosure norms, and ranking rules being written today, while adoption is still partial, are the ones every later agent will inherit, in the same way that Google's original ten-blue-links auction became load-bearing infrastructure for an entire advertising industry that nobody deliberately designed to be structured that way.

The twenty-year-old SEO and SEM industry was built to win a fight for a human's attention as it scanned a page. An agent does not scan; it computes a decision from a mandate you gave it. If that mandate is genuinely to serve you, the entire architecture that monetized a persuadable human has no one left to persuade, and the industry's real fight is no longer about who ranks first. It is about who gets to build the thing standing between you and the ranking at all.