Two documents published fifteen days apart frame agentic commerce better than any keynote this year. The first is a federal complaint. The second is a market measurement. Read together, they say something uncomfortable for everyone building shopping agents: discovery is migrating to AI faster than anyone predicted, completion is staying put at the incumbent, and the shelf both agents and humans read from is now formally alleged to have been rigged for most of a decade.

On August 31, the FTC and 22 state attorneys general sued Amazon in the US District Court for the Western District of Washington over what the agency calls a “secret ad surcharge scheme.” On September 15, PYMNTS Intelligence published “Will the 2026 Shopping Season Go Agentic?,” a study of 2,061 US consumers and 60 retail merchants fielded between September 2025 and August 2026. One documents where the money goes. The other documents what the money was looking at. Neither can be understood without the other.

The 59 Percent Funnel

The PYMNTS headline numbers deserve to be read slowly, because they settle a running argument about how far the agent revolution has actually traveled. Nearly 132 million Americans have now bought something with help from AI. That is not a projection. That is the measured base. And yet 59 percent of AI-assisted purchases still end at Amazon.

That split is the real story of 2026. Agents have taken over the top of the funnel. Humans, or habit, or the gravitational field of Prime, still own the bottom. The shopper delegates the research, the comparison and the shortlist, then carries the winner back to the same checkout they have used for a decade. AI finds. Amazon sells.

The price data sharpens the point. Forty-six percent of AI-assisted shoppers used the technology to find the best price or deal, and 53 percent said a lower price could persuade them to leave a large marketplace entirely and buy from a brand’s own website. Consumer loyalty to the marketplace turnstile is thin, and it is priced. What is not priced, and what no arithmetic can settle, is whether the product the agent found is actually good. Price is verifiable by comparison. Quality is verifiable only by interrogating the review corpus, and that interrogation is exactly the step nobody in the current stack performs.

What the FTC Says Was Under the Sponsored Placements

Now look at what the complaint says the shelf was made of, because agents and shoppers were both reading it.

Amazon sells visibility through keyword auctions: Sponsored Products, Sponsored Brands and Display Ads placed alongside the results a consumer sees after a search. For years, per the complaint, Amazon told advertisers, on its website, in training videos and through hundreds of salespeople, that these were “second price” auctions where the winner “would only pay one cent more than the next highest bidder.” That representation matters because auction design changes bidding behavior. In a true generalized second-price auction, bidders bid up to their true value, knowing they will only pay what competition requires. In a first-price auction, they overpay their own bid whenever they win.

The FTC alleges that beginning in 2019, Amazon inserted an undisclosed surcharge it internally called a “soft reserve price,” so that what advertisers actually paid was, in the words of one internal document, “a surcharge hidden in it.” The executive in charge of Amazon Ads explained internally that the price paid “isn’t set by an actual bidder” but by a “proxy 2nd price that we calculate.” Another document acknowledged Amazon uses an “invented auction participant” to raise prices, which the complaint bluntly describes as essentially shill bids. An Amazon employee put the business logic in one line: the surcharges let Amazon obtain prices “beyond what [can] be achieved through advertiser competition.”

The trajectory the complaint charts is the part every marketplace economist will study. For Sponsored Products ads, the share of the time advertisers paid their own full bid rose from between 30 and 40 percent in 2021, to 70 percent in 2022, to approximately 80 percent in 2024. Over seven years, the agency estimates the scheme extracted likely tens of billions of dollars from more than one million brands and sellers, including over 500,000 small and medium-sized businesses. And the timing was tuned: Amazon “carefully ramps up surcharges” ahead of high-volume days like Prime Day and Black Friday to disguise the inflation, while senior executives, including the head of Amazon Ads and the company’s Chief Digital Economist, acknowledged internally that the “clever non-transparent way to charge first price” had been an “incredibly effective way to drive revenue.” Concealment was the strategy, per the complaint, because disclosure would cause “irrevocable damage to advertiser trust” and a “downward spiral” of lower bids.

These are allegations, and Amazon disputes them. Its statement argues the FTC reviewed “1.5 million pages [of documents] spanning six years” and “leans on a handful of simplified communications to allege a companywide effort to deceive,” per AdExchanger’s coverage. The Commission voted 2-0 to file, which is worth noting, and 22 states signed on. The case will take years. But for the purpose of understanding what shopping agents are reading today, the litigation outcome is almost beside the point, which is the next section’s subject.

The Reader Changed. The Corpus Didn’t.

Here is the structural problem for agentic commerce, and it does not depend on any court’s verdict.

A sponsored placement is a claim about money, not about merit. A human shopper at least has years of learned skepticism, banner blindness and a vague sense that the top result is an ad. An LLM reading the same page sees ordered results and, unless told otherwise, treats order as relevance. The model cannot see the auction. It cannot see the soft reserve price, the proxy second price or the invented participant. It sees only the output: this product, this position, this rating, this review count. If the auction that set the position was covertly converted to first-price, then placement flowed to whoever paid most to be visible, and the agent faithfully launders paid visibility into an apparent quality signal, in a text format that strips out the one cue, the “Sponsored” label, that even a distracted human might catch.

And the sponsored layer is only the first contaminated stratum. Underneath it sits the review substrate, from which Amazon’s own trust reporting says it proactively blocked more than 275 million suspected fake reviews in 2024. Two layers of the product record, the placement layer and the reputation layer, are both publicly documented as contested terrain. The agent reads both as ground truth, because nothing in its stack is built to do otherwise.

This is why the 59 percent figure cuts both ways. Today, the contaminated shelf mostly steers humans, who at least skim. Every forecast says the next reader is a model. Salesforce’s holiday accounting found AI and agents already drove 20 percent of global retail sales, $262 billion, in the 2025 season, that shoppers arriving from AI search channels converted nine times more often than social referrals, and that retailers running their own agents grew sales 59 percent faster than those without. Salesforce’s Caila Schwartz called it “a definitive shift to a new era of ‘agentic’ shopping.” The agent side is industrializing in parallel: on September 2, Anthropic published open-source blueprints for Claude-based shopping and merchant agents that search products, compare options, add items to carts and connect with checkout systems.

Higher-intent traffic meeting an unaudited corpus is not a neutral combination. It means the highest-value shoppers in the funnel are the ones whose judgment is being formed by the least-verified data in the stack.

Merchants Are Building a Second Shelf. It Is Curated Too.

The PYMNTS merchant data reveals a second, quieter shift. Asked how they would expose inventory to AI agents, only 28 percent of merchants said they would offer agents their full product range on the same terms as other channels. One third would offer the full range but with different prices, promotions or delivery choices. Another 28 percent would limit agents to selected products.

Read that again. The majority of merchants are preparing a curated, agent-specific shelf: different terms, different prices, or a hand-picked subset. This is rational retail, and it is also the Amazon pattern rebuilding itself one level up. A shelf that is assembled rather than earned is a shelf whose assembly process is the trust question. The industry is standardizing the plumbing for these feeds, merchant-defined product data flowing to agents through the new rails from Mastercard, Visa and the protocol stacks, but as we walked in Monday’s analysis of the payments trust gap, a merchant-defined feed is legibility, not verification. It is the seller’s own claim, delivered with a network’s logo on the pipe.

The liability numbers complete the picture. Ninety-three percent of merchants say the AI or agent provider should cover the loss when an agent makes the wrong choice, and 80 percent expect that provider to verify the agent’s authority before it acts. In other words: merchants want the agent’s identity checked and the agent’s vendor invoiced, but the question of whether the information the decision was made from was true goes unassigned. Everyone has an opinion about who pays for a bad decision. Nobody has an opinion about what would have prevented it.

What the Shortlist Needs Before Q4

The requirements fall out of the data as cleanly as they did last week. Price sensitivity is the agent’s strongest suit, with 46 percent of AI-assisted shoppers already tasking it with deal-finding, and 53 percent ready to defect from the marketplace over a better number. Those are arithmetic problems, and models are good at arithmetic. Product quality is an evidence problem, and it needs infrastructure:

  • Filter before scoring. Smart Score runs 0 to 100 on review quality and authenticity, computed only after manipulated and low-information reviews are removed. Purchased rating volume stops being an input to what agents and humans trust.
  • Curation with an audit trail. GoBuy shows the top seven verified products per category instead of thousands of ranked listings, which removes both the placement auction and the twenty-five-option shelf from the discovery path. Fewer options, each defensible, beats many options, each purchasable.
  • Persistence as the test. The GoBuy Verified badge requires holding a filtered score of 80 or above for 90 days, which is precisely the property the FTC complaint says the surcharge scheme lacked: behavior that holds steady when nobody is shopping, not prices that spike around Prime Day.
  • Machine-native delivery. Agents consult the filtered trust layer over MCP at gobuy.ai/api/mcp before recommending anything, and the Chrome extension injects the same trust panel directly onto Amazon product pages, so the human inspecting the shortlist sees the same evidence the agent saw.

An agent that checks price everywhere and trust nowhere is a machine for finding the cheapest product on a rigged shelf. The fix is not more caution from the model. It is a verified substrate under the model.

What to Watch

Four signals into year end. First, whether the 59 percent completion share holds through the first genuinely agentic holiday season, or whether in-agent checkout starts pulling it down as the new payment rails mature. Second, the FTC case’s discovery phase, which will surface more internal auction documents and will be read closely by every platform selling visibility to sellers. Third, whether any merchant agent-shelf program ships with independent verification of the product data it serves, or whether all of them remain merchant-defined. Fourth, the first documented case of an agent recommending a product on the strength of manipulated reviews at scale, which will convert this argument from analysis into incident.

Discovery has been delegated. Verification has not. Do both: check the shortlist before you, or anything acting for you, buy at gobuy.ai, with agent integration docs at gobuy.ai/agent-docs.