Block 1 · The Mechanism: GSP, Reserves, and Learned Auctions

← Day index · Block 1 of 10 · ← Previous · Next →

Time budget

~90 minutes. Read Edelman-Ostrovsky-Schwarz ^[1] first and work the pricing formula by hand. Skim Varian ^[2] for the symmetric-equilibrium bounds and Lahaie-Pennock ^[3] for squashing. Spend a full twenty minutes on the two Amazon primary documents - the auction help page ^[6] and the FTC response ^[8] - and read the FTC complaint ^[9] only for its mechanics. Read Wilkens-Cavallo-Niazadeh ^[21] in full. Keep the learned-mechanism papers ^[10]-[15] and ^[28]-[39] as reference; you will meet them again in Block 6.

Why start with the mechanism §

Everything downstream in this reading day - the prediction models, the bidding agents, the measurement - is a response to one object: the rule that turns a set of bids and a set of predicted probabilities into an ordering of ads and a price for each click. If you misunderstand the rule, you will misread every number. The August 2026 FTC dispute is the cleanest demonstration of this: Amazon and the Commission looked at the same auction logs and produced two mutually incompatible descriptions of what the mechanism is ^[8][9]. Neither is a textbook GSP. Both are legible only if you know GSP well enough to see where each departs from it.

This block builds the mechanism in layers: GSP, quality weighting, reserves, Amazon’s disclosed departures, then the frontier where the mechanism is itself a trained model. The map note Amazon Ads, Deeply compresses this to a page.

1. GSP: the rule and why it holds together §

The stage §

A position auction sells $K$ ordered slots on a page. Slot $k$ has a click-through factor ${\alpha }_k$ with ${\alpha }_1>{\alpha }_2>\cdots >{\alpha }_K$, and advertiser $i$ has a private per-click value $v_i$. Varian’s formalization ^[2] models the value of slot $k$ to advertiser $i$ as $v_i{\alpha }_k$: value per click times clicks the slot delivers. Bidders submit a per-click bid $b_i$. The mechanism ranks bids, gives slot $k$ to the $k$-th highest bidder, and charges each winner the bid of the advertiser just below them. That is the generalized second-price auction (GSP), which Edelman, Ostrovsky and Schwarz ^[1] named and analyzed in 2007, drawing on the Google and Yahoo systems of the mid-2000s.

The price formula in the plain version is

$p_k=b_{k+1}+\epsilon ,$

the minimum bid that would have held slot $k$ against the next bidder.

It is not VCG, and it does not need to be §

The natural comparison is the Vickrey-Clarke-Groves mechanism, where truthful bidding is dominant. GSP is not truthful: Edelman et al. give three-bidder examples where under-bidding drops you a slot at a much lower price and raises your payoff ^[1]; Varian makes the same point ^[2]. So why did the market not collapse into permanent shading games?

The answer in both papers is an equilibrium-selection argument. Edelman et al. define locally envy-free equilibria - profiles where no bidder wants to swap slot and price with the bidder immediately above - and prove that GSP has such equilibria, that the one with the lowest revenue coincides slot-for-slot and payment-for-payment with the VCG outcome, and that a natural ascending “generalized English auction” converges to exactly that outcome ^[1]. Varian’s symmetric Nash equilibria are the same set from a different angle, and he uses them to derive upper and lower revenue bounds that he then compares against observed Google prices ^[2]. The takeaway is not that GSP bidders are truthful. It is that the stable outcomes of GSP are pinned to VCG payments, so a bidder with good models can bid near value and lose little.

Quality weighting and the squashing knob §

Real sponsored-search engines never ranked on bid alone. They ranked on $b_i\cdot q_i$, where $q_i$ is a predicted quality term, usually predicted click-through rate. The pricing rule follows from the same “hold your rank” inequality:

$p_k=\frac{b_{k+1}{q}_{k+1}}{q_k}+\epsilon .$

Lahaie and Pennock ^[3] generalized this in 2007 to a one-parameter family of ranking rules, $b_i\cdot q_i^s$ for $s\in [0,1]$, called squashing. At $s=1$ you have pure quality-weighted ranking; at $s=0$, rank-by-bid. Their revenue analysis, run on Yahoo bid and click data, has a result that is still under-appreciated: neither extreme is revenue-optimal in equilibrium, because the optimal $s$ depends on how advertiser values correlate with click-through rates ^[3]. The ranking rule is not just an allocation device. Because it enters the price denominator, it also decides how much of each bid ever reaches the market.

The same formula explains the headline arithmetic of the FTC dispute before you read a single filing. If a platform’s models make $q$ more accurate and the true $q$ of good products rises relative to their competitors’, the winning bid needed to hold a slot falls, even at constant competition. That is what “average winning bids fell 50%” while “conversion rates rose 24%” ^[8] looks like from inside the formula.

Truthful cousins and efficiency bounds §

GSP has a truthful sibling. Aggarwal, Goel and Motwani ^[17] gave the laddered auction in 2006: the truthful counterpart of weighted GSP, revenue-equivalent to GSP’s envy-free equilibrium when click-through rates are separable into an advertiser factor and a slot factor. So the non-truthfulness of GSP is not buying the platform revenue; Milgrom ^[18] explains what it is buying. GSP is a “tight” simplification of running a second-price auction per slot: restricting the message space to a single bid removes the bad equilibria of the richer mechanism without adding any new ones. The simplicity is the feature.

Two results bound how bad GSP outcomes can be for the classical quasi-linear bidder. Caragiannis, Kaklamanis, Kanellopoulos, Kyropoulou, Lucier, Paes Leme and Tardos ^[19] prove the price of anarchy is at most 1.282 under full information and at most 2.927 in the Bayesian setting, and - the clause that matters for this reading day - the bounds extend to coarse correlated equilibria, which is what no-regret learning algorithms converge to. Bidders that learn rather than reason still land inside those bounds. Athey and Ellison ^[20] add the consumer: click-through rates are endogenous to how shoppers search down the page, and a reserve price can raise consumer welfare by excluding low-quality advertisers, which is a welfare argument for reserves quite separate from the revenue one in Section 2.

Then the result to put in bold for any ROAS-driven marketplace. Wilkens, Cavallo and Niazadeh ^[21] show that for value maximizers - bidders who maximize conversions or sales subject to a return constraint rather than maximizing surplus - GSP is the truthful auction. The mechanism the field spent a decade apologizing for as “not VCG” is exactly incentive-compatible for the kind of bidder that now dominates Amazon Sponsored Products. Block 2 builds on this.

Finally, Bergemann, Dütting, Paes Leme and Zuo ^[22] give the mechanism-design reason the quality term must be calibrated, not merely well-ordered: if the platform’s disclosed click-through predictions are required to be calibrated, the only consistent mechanism is rank by bid times an unbiased CTR. A model that ranks correctly but overstates probabilities by 15% is not a ranking error; it is a pricing error of 15% in every auction. That is the bridge to Block 3.

Section takeaway. GSP is a rank-by-score, pay-the-minimum-to-hold-your-rank rule; its stable equilibria coincide with VCG payments; its inefficiency is bounded (1.282 / 2.927) even for learning bidders; it is truthful for value maximizers; and the quality weight $q$ sits in the price denominator, so whoever owns the calibrated quality model controls the effective price of every bid.

2. Reserve prices: theory says yes, the field says “it depends” §

A reserve price is the platform’s own bid. Myerson’s 1981 result that a well-chosen reserve is part of the revenue-optimal auction for a single item extends, with work, to position auctions, and the sponsored-search platforms adopted keyword-level reserves early. The best evidence on what reserves actually do is the Ostrovsky-Schwarz field experiment at Yahoo ^[4]. A randomly selected treatment group of keywords received theory-based reserves computed from estimated value distributions; the control group kept the flat ten-cents-per-click reserve.

Read the results carefully, because they are more honest than their citations. Treatment effects concentrated on high-volume keywords, keywords whose optimal reserve was relatively high, and keywords with relatively few bidders ^[4]. The headline normalized figure of +12.85% treatment revenue shrinks to roughly 8% when a single control keyword is dropped; after trimming the top 0.1% of keywords by volume the estimate is +3.8% and not statistically significant; and on the long tail the effect is about -2.5% ^[4][5]. Revenue per search moved -1.45%, also not significant ^[4]. The paper circulated as a working paper from 2009, appeared at EC 2011 ^[23], and was finally published in the Journal of Political Economy in 2023 ^[5], so you will see it cited with three dates.

The practice moved on to personalized reserves - a different floor per bidder, set from that bidder’s history. Paes Leme, Pál and Vassilvitskii ^[24] wrote the field guide: the problem of choosing per-bidder reserves from samples, which is NP-hard in general but admits good approximations, and the pitfalls of doing it on live data. Mohri and Muñoz Medina ^[25] gave the learning-theoretic treatment, reserve selection as a regression problem with generalization bounds. Derakhshan, Golrezaei and Paes Leme ^[26] gave an LP-based algorithm for personalized reserves in second-price auctions with an approximation guarantee. Read Amazon’s disclosed reserve, set from your predicted sale likelihood and ROAS ^[6], as a personalized reserve in exactly this sense.

Two lessons transfer. Reserves bite where competition is thin - the long tail where a Sponsored Products bidder is often the only serious contender. And theory gives the sign of a reserve’s effect, not its magnitude under strategic response; that needs experiments (Block 7).

Section takeaway. Reserve prices are the platform bidding against you; they matter most on thin keywords; and the one clean field experiment found the aggregate effect statistically indistinguishable from zero.

3. What Amazon actually says about its auction §

Amazon publishes more about its auction than is generally assumed, mostly in help documentation, and the August 2026 lawsuit forced a further layer into the open.

The disclosed pipeline §

Amazon’s auction help page describes a pipeline, not a formula ^[6]. A shopper’s search term, browse node, or detail-page context generates candidate ads. Campaign status, product relevance, and product availability determine eligibility. Eligible ads are then evaluated on expected relevance, maximum CPC bid, context match, and predicted likelihood of engagement, and the page states plainly that a lower bid with higher relevance may win over a higher bid ^[6]. In the language of Section 1, this is a rank-by-score system with a learned $q$, but Amazon does not disclose that the score is multiplicative, what features enter $q$, or how $q$ is calibrated. Treat ”$b\cdot q$” as the right intuition, not as Amazon’s published equation.

Pricing: hard reserve, soft reserve, and the ceiling §

Amazon’s best-practices guide says the final CPC “is usually determined by an auction and is based on your adjusted bid plus additional factors,” and that it “will never exceed your maximum adjusted bid” ^[27]. No fetchable Amazon help page uses the phrase “second-price”; the claim that the auction is GSP in form comes from Amazon’s FTC response ^[8]. The same guide contains a sentence every pacing discussion in Block 2 should start from: “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” ^[27] - Amazon paces the month, not the day.

Pricing is described as a separate layer, and here Amazon’s account is materially richer than textbook GSP. The help page says 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’s bid but cannot exceed the advertiser’s adjusted maximum bid ^[6]. Amazon’s response to the FTC adds the structure ^[8]: there is a hard reserve below which an ad cannot win, and a soft reserve above it. A winning bid above both reserves pays the soft reserve (or the runner-up, if higher). A bid above the hard reserve but below the soft reserve can still win, and pays its own bid.

Note what that last clause does. In classical GSP, a winner’s payment is determined by other bidders. In this rule, when the soft reserve sits above your bid, your bid becomes your price - a first-price outcome inside a second-price frame. The two conditions for paying your own bid are entirely in the platform’s hands: where the soft reserve is set, and whether your bid landed below it.

Bid adjustments: what “your bid” means §

The bid that enters the auction is itself a function of the base bid. Dynamic bidding “down only” can lower the bid by up to 100% for clicks Amazon predicts are less likely to convert; “up and down” can raise it by up to 100% for likely conversions and lower it for unlikely ones; “fixed” uses the exact bid, though manual placement adjustments still apply ^[7]. Placement adjustments for top-of-search, rest-of-search (added January 2024), and product pages can each be set up to 900%, and all adjustment types can combine to a 900% increase - ten times the base bid ^[7]. Amazon’s own worked example: a $1 bid with a 50% top-of-search adjustment becomes $1.50; a further 100% adjustment takes it to $3; another 50% to $4.50 ^[7]. Any analysis of CPC that does not log base bid, strategy, placement multipliers, and actual CPC separately is conflating four different objects.

Two incompatible accounts of the same logs §

On August 31, 2026, the FTC and 22 states sued Amazon over Sponsored Products pricing ^[9]. What follows are allegations and responses; nothing has been adjudicated.

The complaint’s central claim is that Amazon represented Sponsored Products as a generalized second-price auction while using reserve mechanisms to charge more than competition alone would produce. It alleges that a soft-floor mechanism was used for Sponsored Brands in 2019, that “eOPS-based pricing” was tested for Sponsored Products in 2021, and that from 2022 eOPS set hidden soft reserves on almost all Sponsored Products search-page ads, applying to roughly 70% of clicks that year ^[9]. It alleges that advertisers paid on average less than half their bid for top-of-search clicks in 2019, and that the share of clicks priced at the advertiser’s own bid - the complaint’s “First Price Rate” - rose from about 4% in late 2020 to 52-64% by late 2023 ^[9]. The FTC’s press summary compresses this to advertisers paying their own winning bid close to 80% of the time ^[9].

Amazon’s response ^[8] does not dispute the existence of reserves; it disputes their characterization. It says Amazon has used a form of GSP since 2006, that reserve prices are standard across the industry, and that no advertiser ever pays more than its bid. It then makes the relevance argument with numbers: approximately 92% of Sponsored Products ads selected in 2024 were not the highest bid; the mean winning bid was roughly the 128th-highest bid by amount in its auction; relevance-over-bid saved advertisers more than $8 billion from 2021 to 2025; average winning bids fell 50% (one passage says 2019-2025, another 2019-2024 - the inconsistency is in the source); Sponsored Products conversion rates rose more than 24%; and inflation-adjusted average CPC from 2019 through 2024 was flat ^[8].

Here is the point of reading both. The two stories are not logically exclusive. A relevance-weighted ranking can select a lower bidder and a soft reserve can then set that bidder’s price at its own bid. The 92% figure is about selection; the 80% figure is about price formation. Section 1 shows how both fit: $q$ lives in the ranking and the price denominator, while a reserve is a floor applied after. Whether the combination is deceptive is a legal question; whether it is a GSP is definitional. That the layers are distinct is a fact about the mechanism.

A worked example §

Three advertisers compete for one top-of-search slot. Adjusted bids and predicted quality terms:

Advertiser	Adjusted bid $b_i$	Quality $q_i$	Score $b_i{q}_i$
A	$2.00	0.020	0.040
B	$3.00	0.010	0.030
C	$1.20	0.030	0.036

Rank by score: A, then C, then B. The highest bidder, B, loses - the 92% phenomenon in miniature. Under quality-weighted GSP, A’s price is the minimum needed to stay above C:

$p_A=\frac{b_C{q}_C}{q_A}=\frac{1.20\times 0.030}{0.020}=\$1.80.$

Now layer on Amazon’s disclosed reserve structure. Suppose the hard reserve for this placement is $0.50 and the soft reserve, set from predicted sale likelihood and ROAS, is $1.95. A’s bid ($2.00) clears both, so A pays $\max (p_A,r_{\text{soft}})=\max (1.80,1.95)=1.95$ dollars. The runner-up no longer determines the price; the reserve does. Move the soft reserve to $2.10 instead: A’s bid is above the hard reserve but below the soft one, so under the stated rule A still wins and pays its own bid, $2.00. The same three bidders, the same $q$’s, and the price moved from $1.80 to $1.95 to $2.00 as a single platform parameter moved. That parameter is what the complaint calls eOPS and what Amazon calls a reserve ^[8][9]. Play with exactly this in the first widget of The Ad Auction, From the Inside.

Do not infer the production formula from any of this

The help pages give the shape of the pipeline; the litigation gives contested statistics. Neither discloses the feature vector, the model, the calibration, or the reserve function. Everything numeric above is a schematic.

Section takeaway. Amazon’s public account is a relevance-ranked allocation with a two-tier reserve laid over a second-price payment, plus multipliers of up to 10x on the bid that enters; the FTC and Amazon disagree about what to call it, but they agree on the layers.

4. When the mechanism is itself a model §

The research frontier has moved past hand-set scoring rules and hand-set reserves. Five lines of work matter.

RegretNet. Dütting, Feng, Narasimhan, Parkes and Ravindranath ^[10] reframed auction design as constrained learning: an allocation network and a payment network, trained to maximize revenue with a penalty on ex post regret, the gain a bidder could get by misreporting. It recovers Myerson’s optimal auction in settings where the answer is known and produces high-revenue mechanisms where no analytic answer exists. Incentive compatibility becomes something you measure after training rather than something you prove. For a platform, this is the license to treat the auction as an optimizable object.

Deep GSP. Zhang et al. at Alibaba ^[11] kept the GSP skeleton - rank by a score, charge the minimum-to-hold-rank - but replaced the $b\cdot q$ score with a neural network over bids and features, trained to optimize a constrained mixture of platform objectives (revenue, CTR, CVR, user experience) with Lagrange multipliers. Because the score is monotone in the bid, the pricing rule still makes sense. This is the direct descendant of Lahaie-Pennock’s squashing: the ranking rule as a tunable, now learned, knob.

Neural Auction. Liu et al. ^[12] went further at Taobao with a differentiable relaxation of the sort operation, so that the whole allocation-and-payment pipeline is trained end-to-end on advertiser utility, platform revenue, and user experience jointly. They report large-scale offline experiments and an online A/B test in which the learned mechanism beat the incumbent industry mechanisms, and the system was deployed ^[12]. Of all the learned-mechanism papers this is the one with production evidence.

Probabilistic mechanism learning at Amazon. Jeunen, Stavrogiannis, Sayedi and Allison ^[13] take a different route to differentiability: inject Gumbel noise into the allocation so that the winner is drawn from a Plackett-Luce distribution over scores, derive a pricing rule compatible with that stochastic allocation, and train by gradient descent. They report gains over classical mechanisms even with nondeterministic allocation. This is a workshop paper (AAAI 2023 Workshop on AI for Web Advertising), and it is not a description of the production Sponsored Products auction - but it tells you what Amazon’s auction scientists consider a live design direction, and randomization is exactly what the autobidding theory in Block 2 recommends.

The RegretNet family. The journal version is JACM 2024 ^[28]; the follow-ups fix one weakness each: permutation equivariance ^[29], a learned loss in place of the Lagrangian ^[30], transformers for variable bidder and item counts ^[31][32], and MenuNet’s exactly-strategyproof single-bidder menus ^[33]. Two papers take incentives seriously rather than approximately: certified strategyproofness via robustness verification ^[34], and learning inside randomized affine maximizer auctions, a class that is exactly strategyproof ^[35]. Bai, Xie and Wang ^[36] is the 2018 industrial precursor to Deep GSP. The pattern: deployed systems - Deep GSP, Neural Auction, Amazon’s Gumbel mechanism - learn the score and keep GSP-style payments; exact strategyproofness exists only in restricted classes nobody has deployed.

Boosted Second Price. The KDD 2021 paper ^[14] studies a second-price auction with per-bidder additive boosts, with monopoly reserves; its abstract reports up to a 6% revenue improvement over standard second price (the exact conditions were not verified from the full text - treat the number as a pointer, not a result). Boosts are the platform-side twin of the advertiser-side “boosts” analyzed in Deng et al. ^[16], and they matter because they are the simplest learnable perturbation of GSP.

Soft floors, and why simulators need realism §

Finally, Amazon’s own Chen, Nabi and Siniscalchi ^[15] ask what happens to mechanism design conclusions when the auction is modeled realistically: query-dependent slot values and click rates, competitors that are unobserved and changing, partial aggregated feedback, and partially specified payment rules. They model advertisers as adversarial-bandit learners who know nothing about the mechanism and find two things. Soft floors - minimum scores rather than minimum prices - can improve platform metrics even when bidders are drawn from the same population; and advertiser value distributions can be inferred from observed bids ^[15]. The relevance to the FTC dispute is immediate: soft floors are the technical name for the object the complaint alleges and Amazon’s response describes. Zeithammer ^[40] had already shown, in a Management Science analysis of soft floors in display auctions, that a soft floor is not a reserve in disguise: it changes equilibrium bidding, and can lower revenue relative to a well-chosen hard reserve. Chen et al.’s finding that soft floors can help in richer environments is a result about the richer environment, not a reversal. The paper is a 2023 AdKDD workshop paper, so read it as research direction, not a production spec.

Mechanisms for generated text §

The newest frontier is the auction for a slot inside an LLM’s answer rather than on a results page. Dütting, Mirrokni, Paes Leme, Xu and Zuo ^[37] design a token auction in which advertisers bid to influence the generated text, token by token, and characterize the incentive-compatible aggregation rules (WWW 2024 best paper). Soumalias, Curry and Seuken’s MOSAIC ^[38] auctions influence over the whole generated output. Hajiaghayi, Lahaie, Rezaei and Shin ^[39] build ad auctions for LLMs on retrieval-augmented generation: the allocation is bid times RAG relevance, which is relevance-weighted GSP for generated text - the mechanism of Section 1, one interface later. If Amazon’s shopping assistant sells placements, this is the theory it will be sold under.

Section takeaway. The mechanism is now a trained object at the frontier - learned scores (Deep GSP), learned allocation (Neural Auction), stochastic allocation (Amazon’s Gumbel mechanism), learned floors - and the one production result (Taobao) says learned mechanisms beat hand-set ones; what a bidder faces is increasingly a model of the bidder.

What to carry into Block 2 §

The quality weight sits in the price denominator; reserves are a separate layer that can make a click first-price; GSP is truthful for value maximizers; and the mechanism is drifting toward being learned against a model of the bidders. Block 2 asks what happens when the bidders are learned models too.

Reading list for this block §

1. Edelman, Ostrovsky, Schwarz - Internet Advertising and the Generalized Second-Price Auction, AER 2007. 25 minutes. The definitions, the non-truthfulness example, and the locally-envy-free theorem. Skip the generalized English auction proof.
2. Varian - Position Auctions, IJIO 2007. 10 minutes. Sections on symmetric Nash equilibria and the revenue bounds.
3. Lahaie & Pennock - Revenue Analysis of a Family of Ranking Rules for Keyword Auctions, EC 2007. 10 minutes. The squashing parameter and the correlation result.
4. Ostrovsky & Schwarz - Reserve Prices in Internet Advertising Auctions: A Field Experiment, JPE 2023. 10 minutes. Design, then the results table, then the heterogeneity discussion.
5. Amazon Ads help - Sponsored Products across retailers (auction description) and Adjust Sponsored Products bids. 10 minutes. Read for the pipeline and the reserve language.
6. Amazon - Response to the FTC’s Lawsuit Regarding Sponsored Ads, Aug 2026, and the FTC complaint. 15 minutes. Read the mechanics paragraphs of both, side by side.
7. Liu et al. - Neural Auction, KDD 2021. 10 minutes. Architecture figure and the online A/B section.
8. Wilkens, Cavallo, Niazadeh - GSP: The Cinderella of Mechanism Design, WWW 2017. 10 minutes. The value-maximizer truthfulness theorem and its proof sketch.
9. Bergemann, Dütting, Paes Leme, Zuo - Calibrated Click-Through Auctions, WWW 2022. 5 minutes. Statement of the calibration result; pair with Block 3.
10. Chen, Nabi, Siniscalchi - Advancing Ad Auction Realism, 2023. Reference. Return to it in Block 7.

Questions to carry forward §

• If the ranking score is learned end-to-end (Deep GSP, Neural Auction), what does “the minimum bid to hold your rank” mean, and does the locally-envy-free argument survive?
• Amazon’s response says the mean winning bid was the ~128th-highest by amount. What does that imply about the number of eligible ads per query, and hence about the tail?
• A soft reserve set from predicted ROAS is a reserve that depends on the platform’s model of your conversion rate. Who has the information advantage in that pricing, and what does a bidder learn from being charged its own bid?
• Ostrovsky-Schwarz found reserves bite on thin keywords. Where in an Amazon account is competition thinnest, and is that where a bidder should expect the platform’s floor rather than a competitor to set price?
• Wilkens et al. say GSP is truthful for value maximizers. Amazon’s FTC response says the auction is GSP in form. If both hold, is bid-shading by a ROAS-constrained Amazon bidder ever rational, and where does the soft reserve break the argument?
• Randomized allocation (Jeunen et al.) removes the deterministic “you won or you didn’t” feedback. What does that do to a bidder trying to learn the landscape in Block 4?

References §

1. Edelman, Ostrovsky, Schwarz. Internet Advertising and the Generalized Second-Price Auction: Selling Billions of Dollars Worth of Keywords. AER 2007. aeaweb.org
2. Varian. Position Auctions. International Journal of Industrial Organization 2007. sciencedirect.com
3. Lahaie, Pennock. Revenue Analysis of a Family of Ranking Rules for Keyword Auctions. EC 2007. doi.org/10.1145/1250910.1250918
4. Ostrovsky, Schwarz. Reserve Prices in Internet Advertising Auctions: A Field Experiment. Working paper (PDF dated 2016). stanford.edu
5. Ostrovsky, Schwarz. Reserve Prices in Internet Advertising Auctions: A Field Experiment. Journal of Political Economy 131(12), 2023. doi.org/10.1086/725702
6. Amazon Ads. Sponsored Products across retailers (auction, eligibility, and pricing description). advertising.amazon.com
7. Amazon Ads. Adjust Sponsored Products bids (placement adjustments, 900% ceiling, worked example) and Bid controls (dynamic bidding). advertising.amazon.com, API docs
8. Amazon. Amazon’s Response to the FTC’s Lawsuit Regarding Sponsored Ads. Aug 2026. aboutamazon.com
9. FTC. FTC, States Sue Amazon Over Secret Ad Surcharge Scheme (press release, Aug 2026) and Complaint for Permanent Injunction (PDF). ftc.gov, complaint
10. Dütting, Feng, Narasimhan, Parkes, Ravindranath. Optimal Auctions through Deep Learning. ICML 2019. proceedings.mlr.press
11. Zhang et al. Optimizing Multiple Performance Metrics with Deep GSP Auctions for E-commerce Advertising. WSDM 2021. dl.acm.org, arXiv:2012.02930
12. Liu et al. Neural Auction: End-to-End Learning of Auction Mechanisms for E-Commerce Advertising. KDD 2021. arXiv:2106.03593
13. Jeunen, Stavrogiannis, Sayedi, Allison (Amazon). A Probabilistic Framework to Learn Auction Mechanisms via Gradient Descent. AAAI 2023 Workshop on AI for Web Advertising. amazon.science
14. Boosted Second Price Auctions. KDD 2021. dl.acm.org
15. Chen, Nabi, Siniscalchi (Amazon). Advancing Ad Auction Realism: Practical Insights & Modeling Implications. AdKDD 2023. amazon.science, arXiv:2307.11732
16. Deng, Mao, Mirrokni, Zuo. Towards Efficient Auctions in an Auto-bidding World. WWW 2021. arXiv:2103.13356
17. Aggarwal, Goel, Motwani. Truthful Auctions for Pricing Search Keywords. EC 2006. doi.org/10.1145/1134707.1134708
18. Milgrom. Simplified Mechanisms with an Application to Sponsored-Search Auctions. Games and Economic Behavior 2010. doi.org/10.1016/j.geb.2008.12.003
19. Caragiannis, Kaklamanis, Kanellopoulos, Kyropoulou, Lucier, Paes Leme, Tardos. Bounding the Inefficiency of Outcomes in Generalized Second Price Auctions. Journal of Economic Theory 2015. doi.org/10.1016/j.jet.2014.04.010, arXiv:1201.6429
20. Athey, Ellison. Position Auctions with Consumer Search. Quarterly Journal of Economics 2011. doi.org/10.1093/qje/qjr028
21. Wilkens, Cavallo, Niazadeh. GSP: The Cinderella of Mechanism Design. WWW 2017. doi.org/10.1145/3038912.3052687
22. Bergemann, Dütting, Paes Leme, Zuo. Calibrated Click-Through Auctions. WWW 2022. arXiv:2105.09375
23. Ostrovsky, Schwarz. Reserve Prices in Internet Advertising Auctions: A Field Experiment. EC 2011. doi.org/10.1145/1993574.1993585
24. Paes Leme, Pál, Vassilvitskii. A Field Guide to Personalized Reserve Prices. WWW 2016. doi.org/10.1145/2872427.2883071
25. Mohri, Muñoz Medina. Learning Theory and Algorithms for Revenue Optimization in Second-Price Auctions with Reserve. ICML 2014. proceedings.mlr.press
26. Derakhshan, Golrezaei, Paes Leme. Linear Program-Based Approximation for Personalized Reserve Prices. Management Science 2022. doi.org/10.1287/mnsc.2020.3897
27. Amazon Ads. Best Practices for Your Sponsored Products Ads (final CPC, adjusted bid ceiling, daily budgets not paced, monthly cap). advertising.amazon.com
28. Dütting, Feng, Narasimhan, Parkes, Ravindranath. Optimal Auctions through Deep Learning: Advances in Differentiable Economics. JACM 2024. doi.org/10.1145/3630749
29. Rahme, Jelassi, Bruna, Weinberg. A Permutation-Equivariant Neural Network Architecture for Auction Design. AAAI 2021. arXiv:2003.01497
30. Rahme, Jelassi, Weinberg. Auction Learning as a Two-Player Game (ALGnet). ICLR 2021. arXiv:2006.05684
31. Duan et al. A Context-Integrated Transformer-Based Neural Network for Auction Design (CITransNet). ICML 2022. arXiv:2201.12489
32. Ivanov et al. Optimal-er Auctions through Attention (RegretFormer). NeurIPS 2022. arXiv:2202.13110
33. Shen, Tang, Zuo. Automated Mechanism Design via Neural Networks (MenuNet). AAMAS 2019. arXiv:1805.03382
34. Curry, Chiang, Goldstein, Dickerson. Certifying Strategyproof Auction Networks. NeurIPS 2020. arXiv:2006.08742
35. Curry, Sandholm, Dickerson. Differentiable Economics for Randomized Affine Maximizer Auctions. IJCAI 2023. arXiv:2202.02872
36. Bai, Xie, Wang. Practical Constrained Optimization of Auction Mechanisms in E-Commerce Sponsored Search Advertising. 2018. arXiv:1807.11790
37. Dütting, Mirrokni, Paes Leme, Xu, Zuo. Mechanism Design for Large Language Models. WWW 2024 (best paper). arXiv:2310.10826
38. Soumalias, Curry, Seuken. Truthful Aggregation of LLMs with an Application to Online Advertising (MOSAIC). NeurIPS 2025. arXiv:2405.05905
39. Hajiaghayi, Lahaie, Rezaei, Shin. Ad Auctions for LLMs via Retrieval Augmented Generation. NeurIPS 2024. arXiv:2406.09459
40. Zeithammer. Soft Floors in Auctions. Management Science 2019. doi.org/10.1287/mnsc.2018.3164

---

Part of A Day on Amazon Ads. Map note: Amazon Ads, Deeply. Next: Block 2 · Auction Theory in the Autobidding World.