Amazon charged advertisers their full bid 79% of the time, FTC says

Every auction is a promise about arithmetic. Submit a bid, the seller says, and the rule will decide what is charged. The rule is what makes bidding rational; without it a bid is just a number handed to a counterparty with no obligation to explain what happened next.

On August 31, 2026 the Federal Trade Commission and 22 state attorneys general filed a 181-page complaint alleging that Amazon spent more than seven years describing one rule to advertisers while running another. On the same day, a researcher at Zalando published simulation work indicating that the econometric models used to check platform-reported returns overstate paid search performance by roughly two and a half times. Snap made a global attribution change whose entire premise is that its own conversion counts are no longer trusted on their own. Brussels designated ChatGPT under its risk regime after OpenAI declared 159.1 million European users, and Google opened a licensing programme handing rivals its ranking, query, click and view data. An audio company connected the software that runs radio playlists in 120 countries directly to programmatic demand, without publishing how the resulting auction works.

Six items, one question underneath all of them: how much of the transaction is the buyer permitted to see.

The invented bidder

The complaint in United States v. Amazon.com, Inc., case 2:26-cv-03097 in the Western District of Washington, is not an antitrust case. It is a deception case, brought under Section 5(a) of the FTC Act and the deceptive trade practices statutes of 22 states, and its subject is a single line of pricing logic.

PPC Land reported that the filing describes a two-stage calculation in which Amazon runs a generalised second-price auction and then applies an undisclosed soft reserve on top of the result, converting the auction price into the price actually charged. The distinction matters more than it might sound. A conventional reserve is published before an auction runs, and a bid that fails to clear it does not compete. The mechanism described in the complaint operates after the winner and runner-up have already been determined, and its only ceiling is the winning bid itself.

Amazon's own staff appear to have named the thing accurately. A senior scientist described it as an invented auction participant representing how much Amazon thinks a particular ad slot is worth. A senior vice president in the advertising organisation wrote that the second price is not set by an actual bidder but is a proxy second price that Amazon calculates.

The effect compounds over time, and the complaint tracks it through a single figure: the first price rate, meaning the share of clicks charged at the advertiser's own bid rather than at anything resembling a runner-up. In late September and October 2020 that rate sat at 4 percent. During 2021 it ran between 30 and 40 percent. By 2022 it reached 70 percent. In 2024 it was 79.1 percent. For Sponsored Brands in 2024, winners paid their own bid roughly half the time. Digiday's explainer of the case, published by Seb Joseph on September 1, describes the same trajectory as a surcharge occurring around 80 percent of the time by 2024, up from 30 to 40 percent in 2021.

The mechanism was not deployed all at once. Soft reserve pricing began in Sponsored Brands during the 2018 Christmas period, extended to Sponsored Products second-price auctions in mid-2019, and reached Display Ads in 2023.

What gives the complaint its texture is the internal record of people watching the seams. A Sponsored Brands employee wrote in 2018 that price increases were masked by holiday increases in advertiser demand. The head of Sponsored Products auctions noted that surcharge constraints could be relaxed when there were enough confounders, seasonality among them, to make bid shading difficult to detect. A Sponsored Products employee in 2020 spelled out the exposure plainly: imagine being an advertiser and seeing most clicks charged a CPC equal to the bid specified, because it would alter how bidding was approached.

December 2021 supplies the clearest case study. On December 10 the auction team removed a click-level surcharge constraint, and prices climbed over five days. By December 13 more than 20 agencies and advertisers had escalated. On December 17 the vice president of Sponsored Products convened a war room. On December 20 the team drafted a response attributing the movement to Black Friday and Cyber Monday patterns. When an internal summary was revised on February 9, 2022, the discussion of generalised second-price mechanics and reserve pricing had been removed entirely. A product management colleague objected on February 1 that the revision made it sound as though Amazon Ads had no role in the fluctuations, which was not the case. A director instructed that the document should not mention generalised second price at all.

Prime Day 2023 runs the same play in reverse. On July 11 the auction team reported cost-per-click below expectations. On July 12 it obtained approval to raise the surcharge constraint, on an order form stating the purpose as Prime Day ROAS maintenance. An account manager relayed a grocery client's complaint that CPCs were up more than 90 percent and killing return on ad spend.

Then there is the vocabulary. In April 2024 a scientist was instructed to strike the word surcharge from an internal conference slide, with guidance to avoid the term in any internal or external communication. In July 2024 the Sponsored Products team queried ChatGPT for alternatives, weighing value adjustment, price refinement, inflated CPC, price exaggeration factor, performance premium and profit extraction lever. It settled on performance premium. In September 2024 a research paper titled "How Do Advertisers React to Changes in Ad Auction Pricing?" was pulled from Amazon's public science page and from the KDD conference site, prompting a senior manager to write that he should have been on top of it.

The deception count rests on what advertisers were told instead. Since at least 2014, the complaint alleges, Amazon represented that winners pay slightly higher, or one penny more, than the second-place bidder. That representation appeared on the website, on a YouTube training channel, in Amazon Ads Academy courses, in webinars for Kindle Direct Publishing authors, in the Sponsored Brands developer guide posted through July 2026, and in pitch decks to Sony, Samsung, Harman JBL and Coach between 2020 and 2023. In February 2024 a customer service agent explained to an advertiser that a two-dollar bid against a one-dollar next bid would be charged 1.01 dollars.

Reporting design closed the loop. Invoices were not itemised by click, reporting aggregated at keyword level, and desktop and mobile placements were combined. A 2019 operating plan referred to auctions with highly obfuscated reporting. An internal memorandum concluded that Sponsored Brands advertisers were unlikely to adjust bids to lower their payout because they had no visibility into the behaviour of individual auctions. The head of Sponsored Products auctions observed in 2020 that dynamic bid adjustments, which can raise a bid by up to 900 percent, mean an advertiser does not even know what bid is being entered.

Amazon rejects the case. The company says nobody got overcharged, that average Sponsored Products cost-per-click was flat in inflation-adjusted terms between 2019 and 2024, that conversion rates rose more than 24 percent for individual advertisers between 2021 and 2025, and that roughly 92 percent of selected Sponsored Products ads in 2024 were not the highest bid, since ranking weighs relevance alongside price. It puts the mean winning bid at around the twelfth bid by amount, estimates 58 percent higher sales and 46 percent better return against a bid-only ranking, and calculates that advertisers saved more than 8 billion dollars between 2021 and 2025 as a result. It characterises the FTC's evidence as a small handful of outdated training materials and stray internal emails, and argues that reserve pricing is common across the industry and that no advertiser ever paid above a submitted bid.

Two dates in the disclosure timeline are worth holding together. Amazon removed the second-price representation from a marketing page in October 2024, after learning of the investigation. It informed the FTC of that removal on August 26, 2026, five days before the complaint landed. The first public reference to reserve pricing appeared on a single Support Center page on October 30, 2025, reachable by search, and was updated in March 2026 to say that the price charged may exceed the runner-up bid but will never exceed the authorised maximum.

Scale explains the interest. Roughly 1.2 million United States advertising customers are covered, more than 500,000 of them small and medium-sized businesses, against a business that grew 26 percent year on year to 19.8 billion dollars in the second quarter of 2026. The 20 billion dollar figure in the complaint arrives with the average surcharge rates, the size of the reserve guardrails and several dollar estimates redacted from the public version.

The FTC seeks a permanent injunction under Section 13(b). The states seek injunctive relief, rescission or reformation of contracts, restitution, disgorgement, civil penalties and costs, with Maryland alone providing for up to 10,000 dollars per violation. The Commission authorised the filing 2-0.

Context sits close by. The FTC opened consumer protection inquiries into both Amazon and Google over search advertising pricing disclosures in September 2025. A federal court ordered Google to disclose material auction changes that same month. Teads filed a complaint in July 2026 detailing manipulation at the level of individual impressions. The Media Rating Council issued draft transparency standards in September 2025, and the IAB Tech Lab published auction definitions in January 2026. What is new here is the venue: this is the first regulatory action aimed squarely at auction mechanics inside a closed retail media environment, where the seller of the inventory, the operator of the auction and the reporter of the results are the same company.

A smaller Amazon story published the same morning describes the other end of the same opacity. PPC Land documented a United Kingdom marketplace seller whose account was suspended for counterfeit violations three days after Amazon had reinstated the listing at the centre of the complaint. The sequence began with a two-week pattern in late July 2026: coordinated buyer accounts placed orders, requested refunds, left one-star feedback carrying scam accusations and shipped to different addresses. A second wave cancelled orders within hours while the one-star ratings stayed attached, and counterfeit complaints arrived alongside. Negative feedback aggregates against an order defect rate with a one percent threshold. A forum moderator replied that he had almost zero ability to assist with buyer-related situations because of limited visibility. At the time of posting the seller was roughly 48 hours into a 72-hour suspension review, having escalated to United Kingdom management five days earlier without response.

Different surface, same architecture: an automated decision, an aggregated report, and no view of the inputs.

The correction factor is 2.5

The obvious response to a platform whose numbers cannot be checked internally is to check them externally, which is what marketing mix modelling exists to do. Research published on August 21 and reported on August 31 argues that the external check has a systematic bias of its own.

Niklas Heusch of Zalando posted "Structural Estimation of Marketing Mix Model Parameters from Geo-Experiments" as arXiv:2608.21128v1. PPC Land reported the headline result: against a synthetic dataset with a known true paid search return of 4.20 times, a standard marketing mix model returned 10.61 times, roughly two and a half times the truth, with a 90 percent credible interval of 6.56 to 14.36 that never contains the correct answer.

The construction is deliberately generous to the model. Data covers 156 weeks and three channels for a simulated online retailer, with four geo-experiments at different spending levels and periods, estimated in PyMC using the No-U-Turn Sampler across four chains and 2,000 sampling iterations, with a Beta(1,3) adstock prior and a Gamma(3,1) saturation prior.

The critical result is not the headline number but the control condition. Given oracle controls, meaning the true confounders handed to the model directly, the estimate improved only to 8.41 times, with an interval of 6.91 to 9.81. Bias persisted. That is the paper's central claim: this is not a measurement problem that a richer dataset closes.

The cause is endogeneity, and the paper's statement of it is short. Marketing budgets are not randomly assigned. Companies raise spending ahead of high-demand periods. Bidding systems chase performance. Budgets respond to prior results. Standard specifications assume errors are independent conditional on observed controls, a condition the paper describes as virtually impossible to satisfy in practice. The model observes correlation between spend and sales that the business itself created by spending when sales were expected.

The proposed alternative treats a geo-experiment as two parallel universes differing only in the tested channel's spend, subtracts control outcomes from treatment outcomes to eliminate the common component, and estimates the response parameters from the difference. Identification then comes from the experimental design rather than from hoping the covariates were sufficient. With two tests the structural estimate landed at 4.31 times, interval 3.87 to 4.75. With four tests it reached 4.14 times, interval 3.79 to 4.48. Underlying parameters were recovered closely: adstock decay 0.19 against a true 0.20, saturation 1.96 against 1.71, effectiveness 0.20 against 0.18.

The practical obstacle is test volume. Google cut its minimum incrementality experiment budget to 5,000 dollars in November 2025, which lowered the entry price but not the design requirement. The method needs tests at different spending levels, a minimum of four dark weeks, and multiple test periods. One or two studies a year is below what structural estimation requires to identify a saturation curve.

A second and distinct problem sits alongside it. Ryan Dew of Wharton, Nicolas Padilla of London Business School and Anya Shchetkina of Wharton published "Your MMM is Broken: Identification of Nonlinear and Time-varying Effects in Marketing Mix Models" as arXiv:2408.07678v1 in August 2024, running 2,187 simulation settings with 100 datasets each. Without a stock variable, a nonlinear data-generating process produced conflation in 81 percent of cases, 27 percent of them major; a time-varying process produced 91 percent and 40 percent; a Hill function produced 83 percent and 46 percent. With high carryover at 0.8 the time-varying figure reached 99 percent. Noise was the strongest driver of conflation in the regressions, at 23.28 and 22.71.

Conflation here means two model families that fit the same data comparably well and disagree about what to do next. In a coffee brand case study drawn from NielsenIQ Retail Scanner and Nielsen AdIntel data, the nonlinear specification recommended 35 percent of spending on spot television and the time-varying specification recommended none, a divergence the authors priced at 227,000 dollars over 14 weeks.

Heusch's problem is that spending is not random. Dew and colleagues show a problem that survives even when it is. Both papers arrive at the same recommendation from opposite directions, which is that incrementality testing is under-used, and both note that the interaction between the two failure modes is unexamined. Marketing mix modelling has been the industry's answer to platform self-reporting since well before Google released Meridian as an open-source implementation, and the answer now carries a documented correction factor of its own.

Snap stops marking its own homework

The platform response to a credibility problem is to hand the counting to somebody else, and on August 31 Snap did precisely that at global scale.

PPC Land reported that Unified Attribution moved from beta to worldwide availability, with Snap citing a 26 percent lower cost per acquisition for a real money gaming advertiser measured through AppsFlyer against SKAdNetwork-optimised campaigns. Mohegan Sun's online casino recorded 89.8 percent higher return on ad spend and 77 percent lower cost per install in week four. Through Adjust, the French financial app Deblock reported a 25 percent lower effective cost per install, 28 percent lower cost per completed onboarding, 43 percent more installs and 31 percent more users reached, while the publisher Codeway reported an 18 percent lower cost per install at 1.8 times the scale.

The mechanism is a transfer of authority. Unified campaigns use mobile measurement partner signals for optimisation and reporting, inheriting the partner's attribution windows with no advertiser customisation available. The setting is fixed at campaign creation and cannot be changed afterwards, so switching method requires rebuilding the campaign. AppsFlyer and Adjust are live; Singular, Kochava and Branch are in progress. Where an app is connected to more than one supported partner, the system automatically designates whichever records the highest install volume as primary, and that designation can change without advertiser action as volumes shift.

Bright Park, principal partner development manager at AppsFlyer, framed the point directly: for the first time, Snap's optimisation engine and advertiser measurement are working from the same data. Adjust's chief executive Andrey Kazakov described a single view of Snap and Adjust performance data. Adrian Mulryan, Snap's vice president of global agency, called the early results meaningful.

Several qualifications belong beside the percentages. Snap's own documentation states four separate times that Unified Attribution is not designed to match partner reporting exactly, and its help text and FAQ describe the optimisation logic differently, one saying optimisation is powered by Snapchat attribution while outcomes align with partner reporting, the other saying partner signals drive both. Conversion reporting defaults to impression time, with modelling applied where privacy limits data. Accounts can end up displaying three concurrent conversion counts: Snapchat's native attribution, the partner-modelled Unified metrics, and optional SKAdNetwork postbacks. Standard and Unified cost per acquisition are not directly comparable, because they are computed differently.

More to the point, every figure quoted comes from an individual advertiser case supplied through Snap or its measurement partners. No aggregate performance across the beta cohort was published, and no independent evaluation was cited. The claims describe attributed outcomes under changed counting rules, which is a different quantity from proven incremental lift, and it is exactly the quantity the Zalando paper warns is contaminated by the correlation between spending decisions and demand.

Commercially the change lands on a business that needs the credibility. Snap reported advertising revenue up 9 percent in the second quarter with ad prices up 10 percent, and its finance chief flagged World Cup demand as non-recurring. The wider pattern is now well established: Meta aligned with partner measurement in November 2025, TikTok added real-time iOS conversion tracking through Kochava in October 2025, Reddit began displaying dual attribution in May 2026, and OpenAI connected ChatGPT Ads to AppsFlyer and Adjust in July 2026. Google, Meta, Moloco and Unity took equity positions in AppsFlyer partly to keep the measurement layer neutral. Platforms are steadily surrendering the definition of a conversion to companies they do not control, because the alternative is permanent argument with the people holding the budget.

Brussels sets the disclosure rules for two very different products

Regulation approaches the same question from the supply side, and two European decisions landed within hours of each other.

The European Commission designated ChatGPT a Very Large Online Search Engine on August 31, 2026 under reference IP/26/1772, after OpenAI declared 159.1 million average monthly users in the European Union, well above the 45 million threshold. Reddit, at 57.2 million, and Roblox, at 46.6 million, were designated Very Large Online Platforms in the same decision, bringing the total number of designated services to 28 and covering 262.9 million declared users between the three. Henna Virkkunen, executive vice-president for tech sovereignty, security and democracy, said the designations mean the three services will be held to a higher standard of scrutiny and accountability.

Compliance is due at the end of December 2026, four months from notification. OpenAI Ireland Limited is the main establishment, which puts day-to-day supervision with Ireland's Coimisiún na Meán. Reddit Netherlands B.V. falls to the Netherlands Authority for Consumers and Markets. Roblox has no European Union establishment and has appointed a legal representative in the Netherlands.

Three obligations bear directly on advertising. Article 26 requires each advertisement to disclose the advertiser's identity, the basis on which it is presented and the parameters used to select its recipient. Article 28(2) prohibits advertising based on profiling where the service is aware with reasonable certainty that the recipient is a minor. Article 39 requires a public repository holding creative content, advertiser identity, campaign period, targeting parameters and reach metrics, retained for a year after last display. Annual independent audits and vetted researcher data access under Article 40 apply on top, along with systemic risk assessments before critical-impact functionality ships.

The timing is what makes this consequential rather than procedural. Digiday reported on August 31 that OpenAI's advertising business reached a one billion dollar annualised run rate in under 200 days, a figure derived by multiplying current monthly revenue of roughly 83 million dollars by twelve, and therefore a snapshot rather than money booked across a year. Self-serve access opened the same day to eligible advertisers across the 31 European markets where ads launched earlier in August, in approved categories and subject to policy checks, alongside the existing agency and technology partner routes. Dave Dugan, vice-president of global ad solutions, said expanding self-serve access across European markets opens the opportunity to businesses of every size. Seb Joseph and Krystal Scanlon note that sustaining a path to OpenAI's stated 100 billion dollar 2030 target implies compound annual growth of 216.2 percent.

So the same product now runs contextual sponsored placements to Free and Go users in Europe, which PPC Land reported as arriving earlier in August, and must within four months publish a searchable repository describing how each of those placements selected its audience. Enforcement history suggests the repository requirement is not decorative. The Commission fined X 120 million euros on December 5, 2025 over transparency failures including its advertisement repository, then fined Temu 200 million euros on May 28, 2026 over illegal product risk assessment failures roughly two years after designation, and fined AliExpress 550 million euros on July 20, 2026.

Reddit and Roblox arrive with advertising businesses of their own. Reddit reported 762 million dollars of second-quarter advertising revenue, up 64 percent year on year, against 130 million daily active users, and restricted chat and advertising personalisation for European users aged 13 to 15 from June 24, 2026. Roblox sells immersive video and billboard formats through Magnite, PubMatic and Google for users aged 13 and over, and appointed SuperAwesome as its sole third-party contextual advertising partner for under-13s globally in June 2026.

The second European decision runs in the other direction, taking data out of a gatekeeper rather than requiring it to be published. Search Engine Roundtable documented the terms of Google's European Search Dataset Licensing Program, which makes anonymised ranking, query, click and view data from European Economic Area search activity available to qualifying rival search engines. The obligation traces to Google's designation as a gatekeeper on September 6, 2023 and to Article 6(11) of the Digital Markets Act, with the Commission adopting detailed compliance measures on July 16, 2026. PPC Land covered the underlying order requiring Google to hand rivals its search data when it was set.

Eligibility is narrow. An applicant must qualify as an online search engine provider under Digital Markets Act definitions, operate in the European Economic Area with services directed at European Economic Area users, not be operated or controlled by a non-European Economic Area state actor, not be controlled by a sanctioned entity, and either run for two consecutive years with at least 50,000 monthly average European Economic Area users or, if founded within two years, have secured at least 50 million euros of capital. Pricing follows fair, reasonable and non-discriminatory terms limited to the incremental cost of making the data available plus a specified rate of return. Recipients must accept purpose limitations, data segregation, access controls, detailed logging, third-party audit and reporting, and restrictions on international transfers. Applications go through a Google Form.

One regime forces a platform to publish how its advertising selects an audience. The other forces a platform to share the signals that make its ranking work. Both are answers to the same complaint that arrives in the Amazon filing from an entirely different direction.

Radio playlists become an inventory feed

The last story concerns a channel where the auction has not been visible because there has not been one.

PPC Land reported on August 31 that Triton Digital connected its programmatic marketplace to RCS playout software, opening broadcast inventory at stations across 120 countries to automated buyers. RCS, whose brands include RCS, Media Monitors, Mediabase and Florical Systems, supplies the software that decides second by second what a radio station transmits, including where the commercial breaks fall. Live demonstrations run at the International Broadcasting Convention in Amsterdam between September 1 and 14, 2026.

John Rosso, Triton Digital's president and chief executive, said radio has always earned trust at a scale digital aspires to replicate, and described the work as laying groundwork for a future in which every dollar of audio advertising moves through the same connected infrastructure. Susan Larkin, RCS chief executive and president, put the pitch to broadcasters in one line: new revenue without new complexity.

The commercial logic is a gap between consumption and money. Audio accounts for roughly 31 percent of consumer media time against approximately 9 percent of advertising budgets. Closing part of that gap by making over-the-air breaks purchasable inside the same platforms that buy streaming audio and podcasts is the entire proposition, and iHeartMedia is furthest along in pursuing it, targeting roughly 200 million dollars of programmatic revenue in 2026 against 135 million in 2025. Its Audio and Media Services Group, which houses RCS, reported second-quarter revenue up 18.8 percent to 80.5 million dollars and segment adjusted EBITDA up 54.6 percent to 36.7 million.

That last detail carries a disclosure of its own. Triton Digital and RCS are both iHeartMedia subsidiaries, and the announcement describes the arrangement as a partnership without noting the shared parent, which also operates more than 860 stations competing for the same demand. The build-out has been methodical: Viant became the first buyer of iHeartMedia over-the-air inventory in December 2025, StackAdapt integrated broadcast radio in November 2025, and Amazon DSP added Triton audio supply across more than 80 countries, with an Amazon DSP integration for broadcast inventory expected in the fourth quarter of 2026.

What has not been published is the part a buyer would need. No station count. No launch markets. No implementation schedule beyond a description of rolling out. No revenue split between Triton, RCS and broadcasters. No list of which demand-side platforms can reach the inventory. No floor prices, auction structure or CPM guidance. Crucially, the announcement leaves open whether this supports real-time bidding against individual breaks or operates through scheduled allocations, which are different products wearing the same word.

Four demand-side platforms already held 85 percent of global programmatic spend in the first quarter of 2026. A new inventory class arriving inside that concentration, priced by mechanics nobody has described, is the situation the Amazon complaint documents at the far end of its life cycle.

Also noted

• August 31: Google is testing shopping ads inside AI Mode in two new placements, an inline carousel within the generated response and a single ad unit at the foot of the page, spotted by Brodie Clark through the SERPalerts account and published without Google comment. Search Engine Roundtable
• August 31: Optimizely launched Virtual Teammates, AI personas that take a job title and a slot on the organisational chart inside its Opal platform, offered in roles including chief of staff, SEO and AI search analyst, marketing analyst, personalisation strategist and conversion rate optimisation manager, with no pricing or adoption figures disclosed. AdExchanger
• August 31: Microsoft Advertising began previewing a redesigned interface with new navigation and reworked tables, first reported by Hana Kobzová, with users able to revert to the classic experience and Microsoft noting the preview may change before general availability. Search Engine Roundtable
• August 31: BetMGM formalised a two-year working relationship with independent agency Cape as creative and brand strategy agency of record without running a pitch, launching a Jon Hamm campaign called Mullet Over against estimated net media spend of 160 million dollars. Adweek
• September 1: Evian, Sephora and Lavazza built pop-ups across New York City around the United States Open, which runs from August 23 to September 13, with Lavazza taking over the USTA Fan Experience at Hudson Yards from August 30 to September 5 and Sephora running shade matching at Finals FanFest from September 12. Digiday