Better ROAS Does Not Prove a Fair Ad Auction
Executive summary
The Federal Trade Commission and 22 states sued Amazon on August 31, alleging that Amazon secretly altered the pricing of Sponsored Ads auctions through a soft reserve that caused advertisers to pay substantially more than the competitive second price they expected. Amazon denies deception and advertiser harm. It argues that its relevance-weighted auction delivers better performance, that inflation-adjusted cost per click was flat from 2019 through 2024, and that conversion improved 24% from 2021 through 2025.
The case is unresolved, and the allegations should not be treated as findings. But the dispute exposes an operating problem that exists regardless of the legal outcome: marketplace advertisers can observe bids, clicks, attributed sales and return on ad spend while remaining unable to observe the counterfactual clearing price. A campaign can become more productive even as the platform captures more of the value created.
That makes retail-media buying a governance problem, not simply a performance-marketing problem. Merchants need controls for auction-rule changes, bid shading, incrementality, contribution margin and data retention. Better ROAS is useful evidence about a campaign. It is not evidence that an auction is transparent, competitively priced or allocating surplus in the advertiser's favor.
Market context
Amazon sells Sponsored Products, Sponsored Brands and display placements alongside commerce activity on its store. These ads are unusually close to a transaction: the platform can connect a search query, product impression, click and purchase inside one environment. That closed loop improves targeting and measurement, and it gives the auction operator information that an individual seller cannot reproduce.
The FTC's complaint, filed in federal court in Washington, says more than one million brands and sellers were affected and more than 500,000 small and medium-sized advertisers participated in the relevant auctions. The agency alleges that Amazon represented its auctions as generalized second-price auctions, where a winner pays just enough to beat the next-ranked bidder, but added an undisclosed soft reserve beginning in 2019.
Amazon's response describes a different mechanism and conclusion. It says relevance increasingly determines which ad wins, the highest bid is usually not selected, and a soft reserve estimates the value of a placement. When the winning bid exceeds the reserve, the advertiser pays the reserve; when it clears a hard floor but not the soft reserve, the advertiser can receive the placement and pay its own bid. Amazon says the bid is always a maximum and the advertiser never pays more than it submitted.
The legal questions concern disclosure, representations, causation and harm. The management question is simpler: what can an advertiser infer from the metrics it receives? The answer is less than most dashboards imply.
Findings
Finding 1
The auction format changes the rational bid, even when the same ad wins.
In a conventional second-price auction, a bidder can submit something close to the value of the placement because the price is determined by the next-ranked competitor. In a first-price auction, the winner pays its own bid and therefore has an incentive to shade the bid below its estimate of value. A reserve introduces a platform-set floor or reference price that can matter even when no competing advertiser submitted that amount.
The FTC alleges that the share of Sponsored Products wins where the advertiser paid its own bid rose sharply as Amazon used its surcharge. Amazon does not dispute that an advertiser can pay its bid when the bid falls below the soft reserve; it disputes the characterization, disclosure allegations and conclusion that advertisers were harmed.
| Observation year | Alleged share of Sponsored Products wins priced at the advertiser's own bid |
|---|---|
| 2021 | 30%-40% |
| 2022 | 70% |
| 2024 | Approximately 80% |
Source: FTC press release and complaint, filed August 31, 2026. Period: selected annual observations from 2021-2024. Unit: share of winning Sponsored Products auctions, as alleged by the plaintiffs. Transformation: none; the 2021 range and 2024 approximation are retained as reported. Limitations: these are complaint allegations, not adjudicated facts; public materials do not provide the underlying auction-level dataset or confidence intervals.
The trend matters because an advertiser trained on a second-price mental model may bid differently from one told that its own bid can frequently become the price. Automated bidding does not eliminate this issue. An optimizer can respond to observed cost and conversion, but it still learns inside rules set by the platform. If the clearing rule changes, the optimizer may settle at a new equilibrium without identifying how much of the change came from competition, relevance, reserve pricing or the platform's own objective.
Our inference is that the most valuable missing metric is not another attribution ratio. It is the gap between the price paid and the price that would have cleared under the represented or expected rule. Advertisers cannot calculate that counterfactual from campaign reports alone.
Finding 2
Amazon's performance evidence can be true without resolving the transparency question.
Amazon says the average inflation-adjusted cost per click for Sponsored Products search ads was flat from 2019 through 2024, while conversion for individual Sponsored Products advertisers increased more than 24% from 2021 through 2025. It also says average winning bids fell 50% from 2019 through 2025, about 92% of selected Sponsored Products ads in 2024 were not the highest bid, and relevance produced at least 46% better return on ad spend in its 2026 estimate than a highest-bid allocation would have.
These claims support a credible defense of relevance ranking. Showing a useful product rather than the largest nominal bid can improve shopper experience and advertiser productivity. A reserve can also be a legitimate auction design feature when its role is understood.
But the evidence answers a different question. Flat average real cost per click does not show what the same clicks would have cost without a platform-set reserve. Higher conversion does not reveal whether the advertiser retained the improvement or shared more of it with the platform. A falling winning bid can coexist with a higher realized price relative to the next-ranked competitor. Portfolio averages can also hide category, keyword, event-day and advertiser-level dispersion.
| Amazon-reported metric | Period | Reported change | What it supports | What it cannot establish |
|---|---|---|---|---|
| Average Sponsored Products cost per click, inflation-adjusted | 2019-2024 | Flat | Average real click cost did not rise | Counterfactual clearing price or advertiser-level distribution |
| Conversion for individual Sponsored Products advertisers | 2021-2025 | More than +24% | Ads became more productive on average | Whether reserve pricing captured part of that gain |
| Average winning bid | 2019-2025 | -50% | Relevance allowed lower bids to win | Price relative to the next-ranked bid |
| Selected ads that were not the highest bid | 2024 | Approximately 92% | Ranking was not bid-only | Whether pricing and disclosures matched advertiser expectations |
Source: Amazon's August 31, 2026 response. Units: percentage change or share, as labeled. Transformation: none. Limitations: company-reported aggregates use different periods and denominators; the underlying data are not public; Amazon's 2026 ROAS comparison is a modeled estimate rather than an observed market benchmark.
The contrarian conclusion is not that ROAS is meaningless. It is that ROAS can rise while advertiser surplus falls relative to an unavailable alternative. If better relevance lifts conversion by more than price rises, a merchant can be better off in absolute terms and still receive a smaller share of the incremental value. Performance and auction fairness are separate tests.
Finding 3
Retail-media procurement lacks the controls applied to other opaque suppliers.
Merchants often manage sponsored search through daily budgets, target ROAS and platform automation. Those controls are useful but circular: the platform supplies the inventory, runs the auction, attributes the sale and recommends the bid. A seller can optimize within the system without independently validating the system.
This concentration is most consequential when advertising becomes necessary to preserve visibility on the same marketplace where the merchant sells. The practical choice may not be advertise or do not advertise. It may be how much margin to transfer back to the platform to defend rank, launch a product or reach a branded query.
The FTC alleges that Amazon increased surcharges more aggressively on high-volume days. Amazon says advertisers adjust quickly, citing that 80% of bid changes on clicked Sponsored Products search ads from 2019 through 2024 occurred within one day of a previous change. Both statements, if accurate, point to the same operational reality: auction conditions move faster than monthly finance review.
Merchants therefore need a control stack that treats retail media like an unobservable procurement market. That includes immutable exports of bid and placement data, change logs for automated strategies, contribution-margin ceilings, holdout tests and escalation when realized prices separate from competitive proxies. The goal is not to reverse-engineer every auction. It is to know when the platform's optimization is no longer aligned with the merchant's economics.
Implications for operators
First, separate campaign productivity from auction governance. Continue tracking conversion and ROAS, but add gross profit after ad spend, new-to-brand contribution, repeat purchase and organic displacement. A sale that would have occurred without the ad should not justify the same bid as an incremental sale.
Second, retain granular data. Export search-term, bid, placement, click, conversion and attributed-sales records at the shortest useful interval. Preserve campaign-setting changes and automation rules. Without a historical record, merchants cannot distinguish a market shift from a platform-rule shift.
Third, run controlled bid-shading experiments. On matched keywords or products, vary bid caps and observe win rate, placement, conversion and marginal gross profit. The purpose is not to identify a universal discount. It is to estimate how much bid can be removed before economic value falls.
Fourth, set a marginal contribution cap. Target ROAS is a revenue ratio; it can remain attractive while fulfillment, returns, marketplace fees and promotion erase profit. Give bidding systems a ceiling derived from product-level contribution margin and the expected incrementality of the placement.
Fifth, ask for rule-change disclosure. Large advertisers and agencies should make auction mechanics, reserve use, material pricing changes, data access and audit rights part of commercial negotiations. Smaller sellers may lack negotiating leverage, which makes shared agency analysis and regulator-provided transparency more important.
Finally, diversify measurement. Compare marketplace attribution with experiments, first-party customer data where permitted and finance-level cohort outcomes. No single source will produce a perfect answer, but a platform's own dashboard should not be the only evidence used to validate spending on that platform.
Risks & open questions
The complaint has not been adjudicated. Amazon may establish that its disclosures were sufficient, that advertisers understood the bid as a maximum, that reserves are standard, or that the FTC's damages theory fails. Public reporting does not provide the auction-level data required to reproduce either party's economic analysis.
The plaintiffs' estimate of tens of billions of dollars in alleged extraction is consequential but remains an allegation. Amazon's claims of flat real CPC, higher conversion and billions in advertiser savings are also company analyses. The measures cover different periods, and none supplies the central counterfactual: prices, placements and advertiser behavior under a fully disclosed alternative rule.
The thesis would weaken if detailed evidence shows that advertisers were consistently informed of the reserve mechanism in the interfaces they actually used, adjusted bids accordingly, and paid prices close to competitive alternatives after controlling for relevance and placement quality. It would strengthen if auction-level discovery shows systematic price increases unrelated to competition, materially different effects on smaller advertisers or deliberate suppression of information needed for rational bidding.
There is also a wider policy question. Requiring a platform to publish every pricing parameter could make auctions easier to manipulate. Useful transparency may therefore mean clear rule categories, change notice, outcome diagnostics and auditable controls rather than disclosure of real-time reserve values.
Appendix / methodology notes
This report reviewed the FTC's August 31, 2026 press release and 181-page complaint, Amazon's response published the same day, and publicly accessible explanations of the parties' auction claims. All claims about Amazon's conduct, intent, affected advertisers and financial impact are attributed to the complaint and are not presented as findings of fact. Amazon's performance metrics are labeled as company-reported.
The tables preserve the periods and units used by each source. They should not be combined into a damages estimate because the denominators, auction populations and methodologies are not public. No synthetic auction data are used.
A decisive chart would require anonymized auction-level records containing the advertiser bid, ranking score, next-ranked eligible bid, hard reserve, soft reserve, realized price, placement, query, timestamp, conversion outcome and subsequent refund or return. With those inputs, analysts could compare realized prices with alternative clearing rules, estimate advertiser surplus by cohort and test whether effects differ for small sellers, branded queries and high-volume shopping days.