On August 31, 2026, the Federal Trade Commission and 22 state attorneys general sued Amazon in the Western District of Washington over what they called a secret ad surcharge scheme. For ten weeks, the most explosive details sat under redactions. On November 10, the FTC unsealed them, and the picture that emerges is not a story about sloppy disclosure. It is a story about an auction that stopped being an auction, described in Amazon’s own words.
The scale allegation, per the BBC’s coverage of the filing: more than one million advertising customers, including over 500,000 small and medium-sized businesses, and roughly $20 billion extracted since 2019. Amazon’s shares closed 2.5 percent lower the day the suit landed.
For anyone who shops on Amazon, or worse, anyone building an AI agent that shops on Amazon, the unsealed complaint is worth reading in full. Because the question it raises is bigger than ad pricing. If the ranking of products on the world’s largest product discovery engine was, at least in part, set by a rigged auction, then every system that reads that ranking as ground truth inherits the rigging. That now includes the AI shopping agents being rolled out for their first holiday season.
The mechanics: a second-price auction that quietly became first-price
The core allegation is architectural. For years, Amazon told advertisers it ran “second-price” auctions, where the winner pays one cent more than the next highest bidder. As the FTC’s August announcement explains, second-price auctions encourage honest bidding: you bid what the placement is worth because you will only pay what competition forces you to pay.
Beginning in 2019, the complaint alleges, Amazon added an undisclosed surcharge it internally called a “soft reserve price.” The effect, in the words of one internal Amazon document quoted by the FTC, was “a surcharge hidden in it.” The price advertisers paid, in the words of the executive in charge of Amazon Ads, “isn’t set by an actual bidder” but by a “proxy 2nd price that we calculate.” Another document describes an “invented auction participant” inserted to lift prices. The FTC’s characterization is blunt: these function as shill bids, extracting prices “beyond what [can] be achieved through advertiser competition.”
The measurable result, as alleged in the complaint: the share of Sponsored Products auctions where advertisers paid their own full bid, the signature of a first-price auction, climbed from 30 to 40 percent in 2021, to 70 percent in 2022, to roughly 80 percent by 2024. A 2024 discussion between the head of Amazon Ads and Amazon’s Chief Digital Economist, quoted in the filing, allegedly called it a “clever non-transparent way to charge first price” and an “incredibly effective way to drive revenue.”
Amazon understood the disclosure risk, per the complaint: revealing the surcharges would cause “irrevocable damage to advertiser trust” and a “downward spiral” of lowered bids. So, the FTC alleges, it concealed them, including when asked point blank.
The holiday amplifier
The unsealed numbers are at their sharpest on peak shopping days, which is worth pausing on with Black Friday two weeks away.
- Hidden surcharges of up to 50 percent above auction results on ordinary days, and up to 250 percent on Prime Day and Black Friday, per the unsealed complaint.
- During the 2021 holiday season, Amazon allegedly raised advertisers’ average cost-per-click by 20 percent, with some advertisers seeing increases over 200 percent.
- In the 2022 holiday season, the surcharges allegedly added $87 million to Amazon’s revenue worldwide.
- One large food and beverage advertiser was allegedly subjected to price increases of more than 100 percent during Prime Day 2023, with Amazon internally acknowledging the resulting collapse in return on ad spend was “destroying campaign performance.”
- In 2023, with a revenue gap to close, the head of Amazon Ads allegedly emailed a senior vice president: “I’d like to pull out all of the stops (without doing unnatural acts) to close our gaps to plan. I suspect that the entire gap won’t be closed, so whatever can be done with reserves (ideally at least $40MM would be helpful).”
And when advertisers asked whether the auction had changed, they got reassurance. One advertiser asked directly whether “something has changed with respect to the second price auction process, either by design or mistake.” The Amazon account manager’s response, quoted in the complaint: “there is not a bug, or a change in auction style, so no worries there. This year the marketplace is even more crowded than any year previous, and we are seeing the spike in CPCs as a direct reflection.”
Competition, in other words, was the cover story for what the FTC says was a surcharge. FTC Chairman Andrew N. Ferguson’s framing at filing: “These higher costs were largely passed on to American consumers.”
Amazon’s defense, in its own words
Amazon’s official response is detailed and worth taking seriously, because whatever the court decides, it is the version of events advertisers will act on.
Amazon’s key claims: average cost-per-click for Sponsored Products search ads remained flat adjusted for inflation from 2019 through 2024 while conversion rates grew 24 percent from 2021 to 2025; average winning bids fell 50 percent from 2019 to 2025; roughly 92 percent of placed ads do not go to the highest bid because Amazon ranks on relevancy; and “in no scenario does an advertiser pay more than their bid.” Amazon estimates advertisers saved over $8 billion from 2021 to 2025 under its relevancy-weighted model, and argues the FTC’s complaint “cites no evidence of consumer price increases” across more than 150 pages. Its summary is categorical: “The FTC wants the public to believe this case is about higher prices for consumers. It is not.”
These are allegations, not adjudicated facts, and the defense has real points. But notice what the defense does not dispute: that placement on Amazon is bought, at scale, through an auction most advertisers cannot fully see into. Amazon’s own rebuttal celebrates that shoppers are “58 percent more likely to see ads for products they would consider clicking on.” The argument is about whether the auction was fair, not about whether the shelf is for sale. That distinction matters more every day, and here is why.
The pass-through problem: who actually pays for a rigged auction
Even if the courtroom fight over “consumer harm” drags for years, the shopping-side evidence is already in, and it is not flattering for sponsored placement as a product discovery mechanism.
A peer-reviewed study presented at AIES 2024 (arXiv:2407.19099) simulated 4,800 Amazon searches across four countries and analyzed roughly 2 million organic and 638,000 sponsored results. Its findings: items with organic ranks beyond the 100th position regularly appear as sponsored results ahead of the top organic listings, and in the majority of cases the top sponsored results are both costlier and of poorer quality than the organic results they displace. On the US site specifically, the ads were pricier in 71 percent of head-to-head comparisons, more than half the time by at least 50 percent, and 38 percent of them were rated below 4.5 stars versus 22 percent of organic results, as summarized in a recent Not-Ship analysis that measured 34 of 40 top-row slots in her own test searches as sponsored.
The same writeup revisits the underlying math from Boston University law researchers Rory Van Loo and Nikita Aggarwal’s Amazon’s Pricing Paradox: across 100 searches, the best deal sat at slot 17 on average, buyers of the first result overpaid by 29 percent on average, and even buyers of the first non-ad item overpaid by 24 percent. In one USB-C cable search, the top result cost 182 percent more per cable than a five-pack buried at slot 20. The first three results collect roughly two-thirds of all clicks.
Connect the two bodies of evidence and the pass-through mechanism is hard to miss: sellers bid up CPCs in a market the FTC says was secretly inflated at exactly the moments demand peaked; sponsored placement costs flow into the prices and margins of the listings that win; and the listings that win the auction are disproportionately costlier and worse reviewed than the ones below them. Whether or not a court ever stamps the phrase “consumer harm” on it, the shopper buying the first thing they see is paying the auction’s vig. The FTC says the surcharge was hidden. The cost wasn’t. It was sitting in the price tag.
The part nobody priced: AI agents inherit the ranking
Now add this season’s variable. Holiday 2026 is the first with AI shopping agents reading and summarizing Amazon’s results at meaningful volume. What happens when the reader of a pay-to-play shelf is a machine?
The research is not reassuring:
- The ACES simulator study, covered in detail by You.com, found every frontier model it tested favored the top row of results, while a “Sponsored” tag suppressed selection and an “Overall Pick” label boosted it. Sensitivity to actual quality varied wildly: a 0.1-point rating advantage produced failure rates ranging from 0 percent for one Gemini model to 71.7 percent for GPT-4o. Telling agents to disregard position did little.
- A model upgrade from a preview to an official Gemini release flipped position bias from the bottom of the page to the top, reshuffling simulated market share overnight without any seller changing a thing.
- A September 2026 paper (arXiv:2609.17989) examined delegated shopping directly: consumer-delegated agents penalize sponsored listings, but naming a platform as the principal “significantly attenuates this sponsorship penalty,” and stricter disclosure wording like “Sponsored” versus “Promoted” “fails to eliminate” the bias.
Read those against the FTC filing and an uncomfortable synthesis forms. Agents are position-biased readers of a shelf where position was sold in an auction that allegedly rigged its own prices, on a platform whose sponsored results skew costlier and lower rated than organic ones. A human shopper at least sees the small grey “Sponsored” label, and roughly a third of them can’t even distinguish it from organic. An agent summarizing results into a tidy recommendation table may launder paid placement into what reads like an impartial verdict, dropping the label entirely. The 2024-vintage manipulator had to trick a human’s eye. The 2026-vintage manipulator only has to win an auction.
The structural lesson: separate the ranking from the evidence
Neither litigation nor disclosure fixes this. A courtroom can weigh whether a surcharge was deceptive after the fact, years later. It cannot verify, before a purchase, whether the first result a shopper or an agent sees earned its position. And disclosure labels have now been shown not to survive the summarization layer.
What survives is architecture. If the shelf is bought, trust has to come from somewhere the seller and the platform cannot write:
- Review-quality scoring over review-count volume. GoBuy’s Smart Score (0-100) is computed from the quality and authenticity of reviews, not how many a listing accumulated, so a slot-17 gem and a slot-1 ad compete on evidence.
- Fake review filtering before scoring. Manipulated reviews are detected and filtered out, then authentic ones weighted up, so the substrate an agent or a human reasons over is clean.
- A short list, not a shelf. GoBuy surfaces only the top 7 products per query. Seven is small enough that every slot must be defensible, which is the opposite incentive of an auction that monetizes every position.
- An evidence layer agents call before checkout. GoBuy’s MCP server at gobuy.ai/api/mcp exists precisely for this: an AI agent consults an independent trust signal before completing a purchase, instead of trusting the ranking it was handed.
- Time-tested verification. The GoBuy Verified badge requires a score of 80+ sustained over 90 days, which no auction, no launch-day ad blitz, and no review farm can shortcut.
The FTC’s unsealed complaint may ultimately cost Amazon billions. It will not refund the overpayments hidden in a decade of first-page prices, and it will do nothing at all for the agent that recommends a sponsored listing this Friday because it sat at the top of a rigged, or at minimum rented, shelf. The shelf was never neutral. Now that machines are learning to read it, neutrality has to be built somewhere else.
Before you buy, or before you let your agent buy: check the evidence, not the placement. Start at gobuy.ai, install the Chrome extension to see the trust panel directly on Amazon product pages, and if you build agents, wire them to gobuy.ai/agent-docs so every purchase decision consults trust data that no auction can buy.