Tannhäuser Gate

Search

SearchSearch

Recent Notes

See 42 more →

Block 8 · Amazon in Practice: Controls, Disclosures, and the FTC Case

Sep 02, 2026, 23 min read

← Day index · Block 8 of 10 · ← Previous · Next →

Time budget

~75 minutes. Read Amazon’s auction help page and the placement-adjustment page first (10 minutes; they are short and load-bearing). Then read the FTC press release and Amazon’s response side by side. Skim the Amazon Science papers you have not met in earlier blocks - OPTIMUS is the one to read properly. Keep the academic Amazon studies and the vendor table as reference.

From theory to a help page

Seven blocks of theory and the question a practitioner asks is still: what does Amazon actually do? The honest answer is that nobody outside Amazon knows the formula, but a surprising amount is on the record - in help documentation, API references, an SEC filing, an FTC complaint, a corporate rebuttal, and a handful of Amazon Science papers. This block collects that record and reads it with the vocabulary the earlier blocks built. Everything here is a claim by a party with an interest; the block labels who is claiming what.

Amazon’s public description of the auction

Amazon’s current guidance describes a pipeline rather than a formula [1]. A shopper’s search term, browse node, or detail-page context produces candidate ads. Campaign status, product relevance, and product availability then determine eligibility. Among eligible ads, the auction evaluates expected relevance, the maximum CPC bid, context match, and likelihood of engagement, and Amazon states plainly that a lower bid with higher relevance may win [1]. Amazon’s August 2026 response to the FTC adds that eligible ads receive a ranking score based on relevance and bid, and that the formula has increasingly weighted relevance over bid [2]. The shorthand from Block 1 is a reasonable model of this, not a disclosed equation; the feature set, model family, calibration, and weights are not public.

Pricing is a separate layer, and here the public description is richer than textbook GSP. A placement can carry a hard threshold plus other reserves that influence the final price using sale likelihood, predicted ROAS, competing bids, placement, context, and predicted performance. The price may exceed the runner-up bid but will not exceed the advertiser’s adjusted maximum bid [1]. Amazon’s response describes the reserve structure as a hard reserve and a soft reserve: a winning bid above both pays the soft reserve, while a bid above the hard reserve but below the soft reserve can still win and pays its own bid [2]. In the language of Block 2, that is a soft floor - the mechanism Amazon’s own auction-realism paper studied in simulation [3].

Amazon’s best-practices guide is consistent and adds the budget rule [31]. “The final CPC is usually determined by an auction and is based on your adjusted bid plus additional factors”; “the auction considers factors beyond bid to determine both ad placement and final CPC”; the final CPC “will never exceed your maximum adjusted bid.” On budgets: “The daily budget is a daily amount you are willing to spend on a campaign over a calendar month,” “daily budgets are not paced throughout the day,” and a $100 daily budget “may receive up to $3,000 worth of clicks in that calendar month.” Two absences matter as much as the quotes. No fetchable public Amazon page uses the words “second-price”; the GSP-form claim exists only in Amazon’s FTC response [2]. And Amazon DSP bid shading, widely discussed in the industry, is not documented on any fetchable Amazon page, so shading heuristics from the DSP side must not be transferred to Sponsored Products.

Section takeaway. Amazon describes a relevance-weighted ranking with a two-layer reserve on price. Ranking decides who wins; reserves decide how much of the gap between the runner-up and your bid you keep.

Bidding controls, as Amazon documents them

The number an advertiser enters is not the number the auction sees. Three documented layers sit between them [4][5][6].

ControlDocumented behaviorWhat it means for the submitted bid
Dynamic bids - down onlyLowers the bid for clicks less likely to convert; may decrease by up to 100% [4]The base bid is a ceiling, not the bid.
Dynamic bids - up and downRaises the bid for clicks more likely to convert and lowers it otherwise; may move up or down by up to 100% [4]Amazon may spend up to twice the base bid in its own example.
Fixed bidUses the exact bid for all opportunities [4]Not subject to dynamic adjustment, but placement modifiers still apply.
Placement adjustmentApplied separately or with a strategy; Amazon factors both into the real-time bid [5]Base bid, strategy multiplier, and placement multiplier must be logged separately.
Placement ceilingTop of search, rest of search, and product-page adjustments can combine to a 900% increase, i.e. 10x the base bid [6]A placement multiplier can dwarf any base-bid change.

Amazon’s worked example is worth memorizing: a $1 base bid with a 50% top-of-search adjustment becomes $1.50; a further 100% adjustment takes it to $3; another 50% takes it to $4.50 [6]. Rest-of-search adjustments were added in January 2024 alongside the existing top-of-search and product-page controls [7]. Note what the up-and-down strategy is: Amazon’s own predicted conversion likelihood applied as a multiplier on your bid. The platform is running a pCVR-weighted bidder on your behalf, in front of its own pCTR-weighted ranker. Any experiment that changes only a base keyword bid and reads CPC is measuring the composition of all of these.

Section takeaway. The same nominal $1 bid is a different auction input depending on strategy, placement, and Amazon’s conversion prediction. Record all of them or the analysis is not identified.

Hourly data and suggested bids

Two pieces of infrastructure make the hourly loop of Block 4 possible from the outside. Amazon Marketing Stream is a push-based service that sends processed performance data to an advertiser’s AWS account via Amazon Data Firehose or SQS, with traffic and conversion data summarized hourly rather than daily and near-real-time messages for entity changes [8]. Amazon names hourly bid automation as an explicit use case: historical hourly traffic can support bids by hour rather than one daily bid [8][9]. Two limits are equally explicit: the stream is processed reporting, not a raw auction tape, and retail data is not included [8].

The Ads API also returns bid suggestions. Theme-based suggestions return three bids per keyword - low, median, high - with weekly click and purchase-order impact metrics for similar products, computed from a group of recent winning bids and refreshed periodically [10]. Amazon’s caveats are the useful part: the metrics are historical guidance rather than expected per-keyword performance, they assume the campaign does not run out of budget, and keyword-level metrics require at least five keywords in the ad group [10]. A suggested bid is a market anchor, not a demand curve, and certainly not a marginal-ROAS estimate.

Section takeaway. You can observe outcomes hourly and you can see where recent winning bids cluster. You cannot observe the ranking score, the reserves, or the competitor bid distribution.

Revenue

Amazon reports advertising services as a segment covering ads sold to sellers, vendors, publishers, authors, and others, recognized as ads are delivered by click or impression [11].

PeriodAdvertising services revenueStatus
FY2022$37.739BCompleted fiscal year [11]
FY2023$46.906BCompleted fiscal year [11]
FY2024$56.214BCompleted fiscal year [11]
FY2025$68.635BCompleted fiscal year [11]
Q1 2026$17.243BQuarterly [12]
Q2 2026$19.809BQuarterly; reported as $19.8B, +26% year over year [12]
H1 2026$37.052BArithmetic sum of Q1 and Q2, not an Amazon-reported figure

Revenue growth does not identify a mechanism. More placements, more advertisers, better relevance, and different reserve policy all raise the same line, and the dispute below is precisely about which one.

The FTC complaint and Amazon’s response

On August 31, 2026, the FTC and 22 states sued Amazon over Sponsored Products pricing [13]. Everything in this section is an allegation or a rebuttal, not an adjudicated fact.

The complaint’s account [13][14]. Amazon represented Sponsored Products as a generalized second-price auction but used reserve mechanisms to charge more than competition would produce. A soft-floor mechanism was used for Sponsored Brands in 2019; “eOPS Based Pricing” was tested for Sponsored Products in 2021; from 2022, eOPS set hidden soft reserves for almost all Sponsored Products search-page ads, applying on 70% of clicks that year. In 2019, advertisers paid less than 50% of their bid on average for top-of-search Sponsored clicks. The “First Price Rate” - the share of clicks where the advertiser paid its own bid - rose from 4% in late 2020 to 52-64% by late 2023, and the FTC’s summary puts the endpoint near 80%. The complaint characterizes this as extracting tens of billions of dollars over more than seven years.

Amazon’s account [2]. Amazon has used a form of GSP since 2006; reserve prices are common in advertising auctions; no advertiser ever pays more than its bid. The ranking score has increasingly favored relevance: approximately 92% of selected Sponsored Products ads in 2024 were not the highest bid, and the mean winning bid was about the 128th bid by amount. Amazon estimates that ranking relevance over bid saved advertisers more than $8B from 2021 to 2025. Average winning bids fell 50% - the response states the window as 2019-2025 in one passage and 2019-2024 in another, an inconsistency worth preserving. Sponsored Products conversion rates rose over 24%, and average CPC from 2019 through 2024 was flat after inflation. Amazon says it began testing advanced ML relevance models in 2014 and had them in use across Store advertising by 2019.

The two accounts are not logically exclusive

A relevance-weighted ranker can pick a lower bid as the winner while a soft reserve raises what that winner pays toward its own bid. “92% not the highest bid” is a statement about allocation. “80% pay their own bid” is a statement about pricing. Both can be true of the same auction, which is why the earlier blocks kept the two layers separate. The empirical question the case will turn on is how the reserve is set - and Amazon’s help page says it uses sale likelihood, predicted ROAS, and competing bids [1], which is a learned reserve in the sense of Block 2.

There is a third account neither party makes, and it comes from the collusion literature in Block 6. Zhao and Berman (2025) trained multi-agent RL bidders on Amazon.com data and found that under high consumer search costs the learners coordinate on lower bids, and that raising reserves does not restore platform profit [32]. Decarolis, Goldmanis, and Penta (Management Science 2020) showed that common agencies - one intermediary bidding for many competing advertisers - exploit GSP more than VCG [33]. A third-party tool bidding for hundreds of competing sellers in the same keyword is a common agency. If tacit underbidding by learning agents is real, a rising soft reserve is what a platform’s response to it looks like, and neither the complaint’s story nor Amazon’s needs to be false for that to be true.

Section takeaway. Read the FTC complaint for the alleged reserve mechanics and the First Price Rate series; read Amazon’s response for the relevance-ranking evidence. Then hold both against the mechanism in Block 1: allocation by , price somewhere between runner-up and bid, set by a reserve.

Amazon Science on ads

Amazon publishes more about its ads science than its ads mechanism. The sequence that matters:

  • Campaign keyword augmentation via generative methods (Shi et al., ECNLP 2021) - a seq2seq generator with trie-constrained search for cold campaigns, evaluated with human annotation and online tests [15]. It defines which clauses enter the auction at all.
  • AuctionGym (Jeunen, Murphy, Allison; Amazon Science announcement 2022, KDD 2023 paper) - the bandit and doubly robust framing of learning to bid, and the simulator, covered in Block 4 [16][17]. The announcement is dated 2022; the ACM record is 2023.
  • Advancing Ad Auction Realism (Chen, Nabi, Siniscalchi; AdKDD 2023) - query-dependent values and CTRs, unobserved and changing competitors, partial aggregated feedback, and incompletely specified payment rules, with advertisers modeled as adversarial bandits. It finds that soft floors can improve platform metrics in richer environments and that advertiser value distributions can be inferred from bids [3]. Read against the FTC section above, it is a simulation study of exactly the pricing device in dispute.
  • A probabilistic framework to learn auction mechanisms via gradient descent (Jeunen, Stavrogiannis, Sayedi, Allison; AAAI Web Advertising workshop 2023) - Gumbel-noise allocation under a Plackett-Luce model with a compatible pricing rule [18].
  • Amazon Ads Multi-Touch Attribution (Lewis et al., arXiv 2025) - RCT-calibrated ML attribution over hundreds of thousands of experiments; it sets the value signal a bidder optimizes, covered in Block 7 [19].
  • OPTIMUS (Mondal, Kandregula, Agrawal, Sembium; ECML-PKDD 2026) - offline bid recommendation for manual-targeting campaigns, which the paper says account for nearly 30% of advertisers on major e-commerce platforms. Sales maximization under a budget yields an equalization condition on marginal ROAS across targeting clauses, solved over bid-landscape, CTR, and CVR models trained on auction logs with a search over a Lagrange multiplier. Online A/B tests report 2-6% gains in sales, units, and clicks over production baselines [20]. The abstract does not name the platform.

OPTIMUS is the paper to read in full. It is the same Lagrangian as USCB in Block 4, solved offline rather than by a policy, on data that is plausibly Amazon’s, with an experiment on customer money.

Section takeaway. Amazon’s published work covers matching, simulation, mechanism learning, attribution, and offline allocation. None of it discloses the production ranking score or reserve function, and the absence is itself information.

The rest of Amazon’s ads science shelf

Beyond the headline papers, Amazon Science carries a second shelf that maps onto the earlier blocks almost one to one. (The /tag/advertising page returns a 404; the working tags are ad-related-technologies and auction-theory.)

  • Control and allocation. Karlsson (IEEE CDC 2025) designs cascaded feedback control for pacing and bid multipliers under budget and CPC caps, with stability by the circle criterion, following a 2023 hierarchical-control companion [34]. Ge et al. (AdKDD 2024) allocate budget across campaigns with multi-task combinatorial bandits and Thompson sampling on real Amazon data [35]. Nabi, Nassif, Hong, Mamani, and Imbens (Management Science 2022) learn hierarchical priors by empirical Bayes for cold-start bandits; the ads application is not stated on the page [36].
  • Bid recommendation as mechanism. Guo, Li, Nabi, Salhab, and Zhang’s MESOB (KDD 2023 MARBLE workshop) treats bid recommendation as a bi-objective mean-field problem balancing Nash equilibrium against the social optimum [37] - a platform theory of what “suggested bids” could be doing.
  • Experimentation under interference. Jain, Hut, Islam, and Pan (CODE@MIT 2023) document cross-unit spillovers in ads A/B tests [38]; Jain and Appala (AdKDD 2024) build a search-results-page interference network and use graph-clustered randomization to evaluate a paid-search bidding algorithm [39]. These are the Amazon instances of Block 7.
  • Prediction. Chitlangia, Kesari, and Agarwal (AdKDD 2023) scale generative pre-training over user ad-activity sequences [40]; Chaudhuri, Bagherjeiran, and Liu (KDD 2017) rank and calibrate click-attributed purchases [41]; Karra, Zhao, Murray, and Pellegrini (CIKM 2023) nudge click models to attend to position [42]; CTR-BERT (Muhamed et al., NeurIPS 2021 ENLSP workshop) [43] and rationale-guided distillation for relevance (Agrawal, Ahemad, Sembium, COLING 2025) [44] are the LLM-to-lightweight-model path of Block 3.
  • Attribution horizon. Qin (Applied Marketing Analytics 2023) finds upper-funnel ads realize 30-50% of their effect within two weeks against 60-90% for lower-funnel [45], which bears directly on the 14-day window; Pauwels, Schnaidt, and Caddeo (EMAC 2022) estimate the causal impact of display ads on advertiser performance [46].

Two gaps: no Amazon paper on incrementality measurement or Amazon Marketing Cloud was found for 2023-2026, and nothing on the shelf describes the production Sponsored Products ranker.

What academics have measured on Amazon

A separate literature scrapes or models Amazon’s results pages.

  • Farronato, Fradkin, and MacKay (NBER working paper 30894, 2023; a related AEA Papers & Proceedings piece exists) collected 228,281 sponsored and organic results from 3,019 searches by 184 users via a browser extension. Sponsored products appear about seven positions higher, roughly 18%, and Amazon-brand products also receive additional prominence; after observable controls, the Amazon-brand coefficient is about 60% the size of the sponsored coefficient [21].
  • Dash, Ghosh, Mukherjee, Chakraborty, and Gummadi (arXiv 2024, “Sponsored is the New Organic”) scraped four marketplaces across 4,800 searches: nearly 30% of results were sponsored, at least one sponsored result appeared above the top organic result in 85% of cases, and in more than half of instances the top sponsored result was 50% costlier than the top organic one [22]. Preprint; observational; relevance judged by the authors’ reconstruction of organic ranking.
  • Rock, Strauss, O’Reilly, and Mazzucato (SSRN 2023) model “attention share” and find that visual prominence predicts clicks even for products with higher prices or worse ratings [23].
  • Yu (Stanford working paper, 2024) estimates a structural model of consumers, sellers, and the platform from searches, purchases, and auction bids. Removing ads can help consumers and sellers under a fixed commission but hurt them if Amazon re-optimizes the commission; ads help newer products and more differentiated markets most [24].
  • Preuss, Jungbauer, Janssen, and Williams (Economic Journal 2026) model a platform coordinating organic, sponsored, and premium positions with rich data, and find sponsored positions can improve consumer experience even as organic obfuscation raises revenue [25]. The collected source is a Cornell news account; it is a general platform model, not an Amazon estimate.

Section takeaway. Sponsored slots buy position, position buys attention, and whether that helps or hurts shoppers depends on relevance and on what the platform does with its commission. The ranking-by-relevance story and the rent-extraction story are both consistent with these data.

Third-party tools

The automation vendors that operate on Marketing Stream describe their bidders in public, though never reproducibly. Every figure below is a vendor claim from the vendor’s own page.

VendorDescribed mechanismDisclosed inputsVendor-reported result
PerpetuaContextual, conversion-based bidding with goal presets (growth, profitability, brand defense) [26]Objectives; automated campaign executionNone on the cited page
PacvueRules-based automation plus AI bidding [27]Dayparting, pacing, Marketing Stream, placements, Buy Box, inventory, pricing“10%+ ROAS lift, 25%+ sales” (vendor claim)
QuartileHourly, single-keyword bidding adjustments [28]Marketing Stream, Amazon Marketing Cloud, scheduled rules“+41% ROAS average” (vendor claim)
AdbrewAI insights plus custom rule automation [29]Bids, budgets, placements, targetsCase studies: 31.9% lower ACOS, 30.86% lower spend (vendor claims)
TeikametricsProprietary forecasts of AOV and CVR feed an optimal-bid calculation [30]ACOS limit, bid modifier, product and goal data“Up to 50% better bid-to-value accuracy, up to 30% net revenue” (vendor claims)

Two patterns are worth noting. Quartile’s hourly loop is the outside-in version of Zhao et al.’s control-by-model design in Block 4, running on the exact data cadence Marketing Stream exposes. Teikametrics’ forecast-to-bid pipeline is OPTIMUS without the optimality condition or the experiment. Neither page says how the tool’s submitted bid interacts with Amazon’s own dynamic bidding, which is the question that determines whether two optimizers are fighting over the same multiplier.

Section takeaway. Vendor pages are evidence about control surfaces and data cadence. They are not evidence about causal lift.

How to read a Sponsored Products CPC

Put the block together as a decomposition. When an advertiser sees a CPC in a report, seven things happened to produce it:

  1. Base bid. The number entered on the keyword or target.
  2. Strategy multiplier. Down-only, up-and-down, or fixed. Under up-and-down, Amazon’s own predicted conversion likelihood scales the bid by a factor between 0 and 2 [4].
  3. Placement multiplier. Top of search, rest of search, or product page, compounding to as much as 10x [6].
  4. Adjusted maximum bid. The product of the three above. Amazon states the price will never exceed this [1].
  5. Ranking score. The adjusted bid combined with predicted relevance and engagement; the ad wins or loses here, and Amazon says 92% of winners in 2024 were not the top bid [2].
  6. Reserve layers. A hard reserve and a soft reserve, set from sale likelihood, predicted ROAS, and competing bids. Above both, you pay the soft reserve; between them, you pay your bid [1][2]. The FTC alleges the soft reserve sits at or near the winner’s bid most of the time [13].
  7. CPC. Somewhere between the runner-up’s effective price and the adjusted maximum bid, and then averaged in the report across placements, hours, and strategies.

One more constraint sits outside the list. The daily budget is averaged over the calendar month and is not paced intraday [31], so a campaign can spend several times its daily budget in an hour and go dark until midnight. That is why third-party tools run their own pacers, and why hourly Marketing Stream data exists as a product.

A bidding agent that treats step 1 as its action and step 7 as its reward is optimizing through five layers it cannot see. The practical implication, which the The Ad Auction, From the Inside lab lets you feel, is that the informative experiments vary steps 1 through 3 one at a time and log the placement-level CPC, not the account average.

Reading list for this block

  1. Amazon Ads, Sponsored Products across retailers (auction help page) - 5 min. The pipeline and the reserve language.
  2. Amazon Ads, Adjust Sponsored Products bids (placement adjustments) - 5 min. The 900% ceiling and the compounding example.
  3. FTC press release and complaint, August 2026 - 20 min. Read the eOPS timeline and the First Price Rate series.
  4. Amazon, Response to the FTC’s lawsuit, August 2026 - 15 min. The 92%, 128th-bid, and $8B claims and the hard/soft reserve description.
  5. Mondal et al., OPTIMUS, ECML-PKDD 2026 - 20 min. The marginal-ROAS equalization and the A/B design.
  6. Chen, Nabi, Siniscalchi, Advancing Ad Auction Realism, AdKDD 2023 - 10 min. Skim for the soft-floor result.
  7. Farronato, Fradkin, MacKay, Self-Preferencing at Amazon, NBER 2023 - reference. The cleanest external measurement of what a sponsored slot buys.
  8. Marketing Stream and bid-suggestion API documentation - reference. What you can and cannot observe.

Questions to carry forward

  • If the soft reserve is set from predicted ROAS and competing bids, is it a learned mechanism in the RegretNet sense, and what objective would it be trained on?
  • Under up-and-down bidding, Amazon multiplies your bid by its pCVR estimate and your bidder multiplies it by yours. When the two disagree, whose estimate should win, and how would you find out?
  • “92% not the highest bid” and “80% pay their own bid” are consistent. What single logged quantity would distinguish a relevance-driven market from a reserve-driven one?
  • OPTIMUS reports 2-6% gains for manual-targeting campaigns. What does that imply about the gap between a good offline Lagrangian and the production baseline?
  • Which of the seven CPC layers can a third-party tool actually observe, and is the rest identifiable from hourly data at all?

References

  1. Amazon Ads. Sponsored Products across retailers (auction and pricing guidance). advertising.amazon.com
  2. Amazon. Amazon’s response to the FTC’s lawsuit regarding Sponsored Ads. August 2026. aboutamazon.com
  3. Chen, Nabi, Siniscalchi. Advancing Ad Auction Realism: Practical Insights & Modeling Implications. AdKDD 2023. amazon.science · arXiv:2307.11732
  4. Amazon Ads. Bidding strategies for Sponsored Products (dynamic bidding help). advertising.amazon.com · guide
  5. Amazon Ads API. Bid controls. advertising.amazon.com
  6. Amazon Ads. Adjust Sponsored Products bids (placement adjustments). advertising.amazon.com
  7. Amazon Ads. Rest of Search Bid Adjustment for Sponsored Products. January 2024. advertising.amazon.com
  8. Amazon Ads API. Amazon Marketing Stream - Overview. advertising.amazon.com
  9. Amazon Ads. Amazon Marketing Stream: Measure campaign performance in real time. advertising.amazon.com
  10. Amazon Ads API. Theme-based bid suggestions - Quick-start guide. advertising.amazon.com
  11. Amazon.com, Inc. Form 10-K for fiscal year ended December 31, 2025. sec.gov
  12. Amazon. Q2 2026 earnings: CEO Andy Jassy on what’s driving Amazon Ads growth. aboutamazon.com
  13. FTC. FTC, States Sue Amazon Over Secret Ad Surcharge Scheme. Press release, August 2026. ftc.gov
  14. FTC et al. v. Amazon. Complaint for Permanent Injunction, Monetary Judgment, Civil Penalty Judgment, and Other Relief. ftc.gov
  15. Shi, Rao, Wu, Zhang, Wang. Campaign Keyword Augmentation via Generative Methods. ECNLP at ACL-IJCNLP 2021. amazon.science
  16. Amazon Science. Amazon scientists win best-paper award for ad auction simulator. 2022. amazon.science
  17. Jeunen, Murphy, Allison. Off-Policy Learning-to-Bid with AuctionGym. KDD 2023. doi.org · code
  18. Jeunen, Stavrogiannis, Sayedi, Allison. A Probabilistic Framework to Learn Auction Mechanisms via Gradient Descent. AAAI Workshop on AI for Web Advertising 2023. amazon.science
  19. Lewis, Zettelmeyer, Gordon, Garib, Hermle, Perry, Romero, Schnaidt. Amazon Ads Multi-Touch Attribution. arXiv 2025. arXiv:2508.08209
  20. Mondal, Kandregula, Agrawal, Sembium. OPTIMUS: Optimal Offline Bidding Strategy for Manual Targeting Advertising Campaigns. ECML-PKDD 2026. amazon.science
  21. Farronato, Fradkin, MacKay. Self-Preferencing at Amazon: Evidence from Search Rankings. NBER Working Paper 30894, 2023. nber.org · AEA P&P
  22. Dash, Ghosh, Mukherjee, Chakraborty, Gummadi. Sponsored is the New Organic: Implications of Sponsored Results on Quality of Search Results in the Amazon Marketplace. arXiv 2024. arXiv:2407.19099
  23. Rock, Strauss, O’Reilly, Mazzucato. Behind the Clicks: Can Amazon Allocate User Attention as it Pleases? SSRN 2023. ssrn.com
  24. Yu. The Welfare Effects of Sponsored Product Advertising. Stanford working paper / SSRN 2024. ssrn.com
  25. Preuss, Jungbauer, Janssen, Williams. Search Platforms: Big Data and Sponsored Positions. Economic Journal 2026. doi.org · Cornell Chronicle
  26. Perpetua. eCommerce Advertising Software. perpetua.io
  27. Pacvue. Pacvue for Amazon. pacvue.com
  28. Quartile. Quartile for Amazon PPC. quartile.com
  29. Adbrew. Amazon PPC Automation and Optimisation Software. adbrew.io
  30. Teikametrics. Intro to the Teikametrics Bidder. help.teikametrics.com
  31. Amazon Ads. Best practices for your Sponsored Products ads. advertising.amazon.com
  32. Zhao, Berman. Algorithmic Collusion in Sponsored Search Auctions (multi-agent RL on Amazon.com data). arXiv 2025. arXiv:2508.08325
  33. Decarolis, Goldmanis, Penta. Marketing Agencies and Collusive Bidding in Online Ad Auctions. Management Science 2020. doi.org
  34. Karlsson. Multivariable Feedback Control for Multi-Constraint Optimization in Online Advertising. IEEE CDC 2025. amazon.science
  35. Ge et al. Multi-Task Combinatorial Bandits for Budget Allocation. AdKDD 2024. amazon.science
  36. Nabi, Nassif, Hong, Mamani, Imbens. Bayesian Meta-Prior Learning Using Empirical Bayes. Management Science 2022. amazon.science
  37. Guo, Li, Nabi, Salhab, Zhang. MESOB: Balancing Equilibria & Social Optimality. KDD 2023 MARBLE workshop. amazon.science
  38. Jain, Hut, Islam, Pan. Cross-Unit Spillovers in A/B Testing: Empirical Evidence from Ads. CODE@MIT 2023. amazon.science
  39. Jain, Appala. SERP Interference Network and Its Applications in Search Advertising. AdKDD 2024. amazon.science
  40. Chitlangia, Kesari, Agarwal. Scaling Generative Pre-training for User Ad Activity Sequences. AdKDD 2023. amazon.science
  41. Chaudhuri, Bagherjeiran, Liu. Ranking and Calibrating Click-Attributed Purchases in Performance Display Advertising. KDD 2017. amazon.science
  42. Karra, Zhao, Murray, Pellegrini. Nudging Neural Click Prediction Models to Pay Attention to Position. CIKM 2023. amazon.science
  43. Muhamed et al. CTR-BERT: Cost-effective Knowledge Distillation for Billion-Parameter Teacher Models. NeurIPS 2021 ENLSP workshop. pdf
  44. Agrawal, Ahemad, Sembium. Rationale-Guided Distillation for E-Commerce Relevance Classification. COLING 2025. amazon.science
  45. Qin. Lengthen Your Attribution Window: Which Digital Ads Have Most Long-Term Impact? Applied Marketing Analytics 2023. amazon.science
  46. Pauwels, Schnaidt, Caddeo. Causal Impact of Digital Display Ads on Advertiser Performance. EMAC 2022. amazon.science

Part of A Day on Amazon Ads. This block is the ground truth for the mechanism sketched in Section 1 of Amazon Ads, Deeply; the FTC figures quoted there come from the sources above.

Graph View

Backlinks