Pith. sign in

REVIEW 1 major objections 5 minor 131 references

In a production app-store deployment, an adaptive sliding-window policy matches the revenue of the always-five-ads rule while recovering most conversion losses, beating every fixed ad-load benchmark.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Raising sponsored-slot counts raises revenue up to 43% but lowers conversions up to 5%, and a deployed query-level adaptive ad-load policy achieves five-ad revenue with conversion losses closer to low-ad policies.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Strong causal experiment and a useful adaptive algorithm, but the headline deployment claim rests on a confounding comparison between cohorts with different histories, and the regret bound's key independence assumption conflicts with the deployed design. the 1 major comments →

arxiv 2607.14418 v1 pith:WSOFASEG submitted 2026-07-15 cs.LG econ.GNq-fin.EC

Adaptive Ad Load Design for Sponsored Search Markets: Evidence, Theory, and Deployment

Mohammad Rashid , Hema Yoganarasimhan This is my paper
classification cs.LG econ.GNq-fin.EC
keywords ad load designsponsored searchadaptive policyfield experimentsliding-window banditsrevenue–conversion trade-offdynamic regretnon-stationarity
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper establishes that the number of sponsored ads shown per search should be chosen query-by-query and updated as market conditions shift, rather than fixed platform-wide. A 66-day randomized experiment on a large Android app store shows that raising ad slots from one to six lifts advertising revenue by up to 43% but cuts total app installs by up to 5%, with the extra revenue concentrated on queries whose ads already convert well. To exploit this, the authors design LAAL, a model-free rule that estimates each query's recent revenue and conversions in a sliding window and selects the ad load with the highest scalarized reward, wrapped in an exploration layer (e-LAAL) that keeps fixed-cohort arms alive. In a 22-day production deployment serving 22.3 million users, LAAL earned revenue statistically indistinguishable from the always-five-ads policy while delivering roughly 3% more conversions, and it outperformed both the best historical uniform rule and a historical query-dependent static rule. The paper also proves a finite-time dynamic-regret bound for the e-LAAL architecture.

Core claim

The central claim is that a locally adaptive, model-free policy can shift the empirical revenue–conversion frontier outward relative to every fixed uniform or query-dependent static ad-load policy in a non-stationary sponsored-search market. The experiment shows that marginal revenue from additional slots depends sharply on query-level advertiser quality and on within-query changes such as brand-advertiser presence, so no single static rule is near-optimal. LAAL operationalizes this by maintaining sliding-window estimates of mean revenue and conversions for each query–ad-load pair, scalarizing them as λ·Revenue + Conversion with λ≈0.01 calibrated to the experiment's revealed exchange rate, a

What carries the argument

LAAL (Locally Adaptive Ad Load) is a model-free, query-level sliding-window decision rule: for each query and candidate ad load, it computes recent mean revenue and conversion per search over a fixed-length window (3 days in deployment), forms the scalarized reward r = λ·Revenue + Conversion (λ≈0.01, calibrated to the experiment's empirical revenue–conversion slope), and selects the ad load maximizing that reward. A tiny regularization term breaks near-ties toward fewer ads, and queries with fewer than m=100 recent observations fall back to a default of four ads. The e-LAAL deployment architecture splits traffic 90/10: 90% of users run LAAL while 10% are permanently assigned to fixed-ad-load

Load-bearing premise

The regret guarantee assumes that, for a given query's sequence of searches, the experimental cohort assigned to each search is independent across searches; in the deployed system users are permanently assigned to cohorts by hashed device IDs, so repeated searches by the same user are correlated, and without that independence the concentration argument behind the advertised regret bound is not proven.

What would settle it

Check whether LAAL's deployment point lies below the chord connecting the always-five-ads and always-six-ads arms in revenue–conversion space; if a replication places it below that line, the frontier-improvement claim fails. For the theory, examine user repetition within a 3-day window for high-volume queries: if a small number of users account for a large share of a query's searches, the per-search cohort assignments within the window are highly correlated, violating the independence assumption on which the high-probability regret bound rests.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Platforms should treat ad load as a query-level, time-varying decision; the best uniform static load shifted from three ads during the experiment to five ads during deployment, so any fixed rule chosen from historical data can become stale.
  • A small exploration budget (10% static cohorts) is enough to maintain learning and credible counterfactuals while the adaptive arm exploits; the system-level e-LAAL mixture stays close to the pure LAAL arm.
  • The policy adapts without a parametric reward model or observations of latent advertiser states: for brand queries it recommends fewer ads when the brand advertiser is absent and more when it is present, purely from recent outcomes.
  • The frontier gains persist on non-brand queries, so the improvement is not driven solely by brand-query mechanics.
  • LAAL's empirical regret relative to an ex post query-dependent oracle is near zero on the highest-volume queries, where local estimates are most precise.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves implicit that the same architecture could be ported to other sponsored-search or e-commerce surfaces that observe query-level revenue and conversions, since LAAL requires no platform-specific model.
  • The scalarization parameter λ is calibrated to the short-run revenue–conversion frontier; if the platform values conversions differently over the long run, re-running the deployment with different λ values would trace an adaptive Pareto frontier – a direct testable extension.
  • The proven regret bound rests on an independence assumption that the permanent-cohort design does not strictly satisfy; a sharper theory would need to model user-level persistence, which suggests treating the deployment result primarily as empirical evidence rather than as a corollary of the bound.
  • Because users under high ad loads search more, LAAL's conversion gains may compound over longer horizons; measuring engagement and repeated search behavior beyond the 22-day window would test whether the frontier improvement grows or decays.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

1 major / 5 minor

Summary. This paper studies ad-load design in a large Android app-store sponsored-search market. It first reports a 66-day user-level randomized field experiment (≈5.06M users) that exogenously varies the number of sponsored slots from 1 to 6, documenting a revenue–conversion trade-off: revenue per user rises by up to 43%, total conversions fall by up to 5%, and engagement declines by up to 2.2%. It further shows that the trade-off is heterogeneous across query-level ad-conversion segments and within queries over time as brand-advertiser presence changes. Motivated by these findings, the authors propose e-LAAL, an exploration-augmented, query-level sliding-window adaptive policy that combines a greedy adaptive arm (LAAL) with persistent static cohorts. They provide a finite-time dynamic-regret guarantee (Theorem 1, Web Appendix F) and report a 22-day production deployment serving 22.33 million users and 77.60 million searches, in which LAAL achieves revenue statistically indistinguishable from the always-5-ads policy while generating about 3% more conversions, and outperforms both historical and deployment-period static benchmarks.

Significance. If the empirical claims hold, the paper makes a substantial contribution: it provides some of the largest causal evidence on sponsored-search ad-load effects, documents economically meaningful heterogeneity, and demonstrates that a model-free adaptive policy can improve the revenue–conversion frontier at production scale. The field-experiment design is strong: randomized user cohorts, user-level regressions with tight standard errors, IPW normalization, and several robustness checks (non-brand queries, time trends, cohort-invariance tests) are sensible. The deployment is unusually large and the accompanying theoretical analysis, if valid, would be a useful addition to the non-stationary bandit literature. The paper is also candid about several limitations, including the short deployment window and the conditional-cohort-invariance assumption. However, two load-bearing concerns—the causal identification of the deployment comparison and the validity of the regret guarantee under the deployed cohort assignment—need to be addressed before the headline claims can be accepted.

major comments (1)
  1. [§8.2] The “deployment query-dependent empirical oracle” is selected ex post using the same 22-day deployment outcomes that are then used to compute LAAL's empirical regret (Eq. 21). This benchmark is infeasible and overfits noisy query-level outcomes, so the near-zero regret reported for high-volume queries is partly mechanical. The authors acknowledge the oracle is infeasible, and the comparison is useful as a diagnostic, but the claim that LAAL “exhibits near-zero empirical regret” should be interpreted with caution and should be accompanied by a discussion of the oracle's in-sample selection bias. A small addition—e.g., a split-sample or cross-validation construction—would strengthen this section.
minor comments (5)
  1. [§4.2.1, Eq. (5)] The query-level regressions cluster standard errors at the query level, but treatment is assigned at the user level and each user contributes multiple searches. Clustering at the user level (or two-way clustering) would be more conservative and should be considered, particularly for the heterogeneity estimates.
  2. [§6.5] The text says “We now discuss the three main parameters” but then lists five algorithmic hyperparameters (h, m, η, δ, i_def) in addition to λ and ε. Please adjust the wording for clarity.
  3. [Web Appendix F] The notation δ_sm is introduced in the appendix but not clearly connected to the main-text δ in Eq. (12). Use a single symbol to avoid confusion.
  4. [Figures 8b and 12] These figures would benefit from explicit axis labels and units (e.g., share of searches on the y-axis). Currently the y-axes appear unlabeled.
  5. [References] Some references are incomplete (e.g., “R. P. Leme et al. A field guide to personalized reserve prices” has no venue or year). Please complete the bibliography.

Circularity Check

0 steps flagged

No significant circularity: deployment result is a measured outcome, and fitted tuning parameters do not force the headline comparison.

full rationale

The paper's central deployment claim—LAAL delivering revenue essentially equal to T5 (36.13 vs. 36.11 per user) while increasing conversions (2.127 vs. 2.064 per user) in Table 6—is a direct measurement from the deployed cohorts, not a quantity reconstructed from fitted inputs. The scalarization weight λ=0.01 is calibrated from the first experiment's uniform-policy frontier (§6.5) and is then used to define the reward metric in the benchmark comparisons, but the headline trade-off improvement is stated in raw revenue and conversion units and would hold independently of λ. The default ad load i_def is fit from historical low-volume queries, but it only governs the 21% fallback searches and is not used to construct the LAAL-versus-static revenue/conversion comparison (§6.5, Appendix G). The dynamic regret bound in Appendix F is derived from explicit assumptions (E1–E4) and standard non-stationary bandit arguments, not imported from a self-citation or a uniqueness theorem. The paper contains self-citations, but none is load-bearing for the main empirical or theoretical claims. The inherited static cohorts do raise a legitimate causal-identification concern about prior 66-day exposure, but that is a selection/confounding threat rather than a circularity in the derivation chain. No step reduces a prediction to its own inputs by construction.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The central empirical results rest on standard randomized-experiment assumptions plus a tested cohort-invariance condition. The theoretical regret bound adds an independence assumption (E2) that is not satisfied by the persistent-user-cohort design. The adaptive policy itself introduces no new physical or market entities; its free parameters are calibration and hyperparameter choices, the most consequential being lambda and i_def.

free parameters (7)
  • lambda (revenue-conversion scalarization) = 0.01
    Calibrated from the 66-day experiment's static revenue-conversion frontier (Sec 6.5) and used in the objective and benchmarks.
  • h (sliding-window length) = 3 days
    Hand-chosen to balance responsiveness and variance (Sec 6.5); affects which outcomes LAAL uses.
  • m (minimum window volume) = 100
    Hand-chosen threshold before greedy selection; 21% of deployed searches fall back to default when below it (Sec 7.1).
  • i_def (fallback ad load) = 4
    Chosen from historical static-cohort data as the best uniform policy on low-volume queries (Web Appendix G).
  • eta (tie-break regularizer) = 1e-3
    Hand-set to favor fewer ads on near-ties (Sec 6.5).
  • delta (smoothing constant) = 1e-6
    Hand-set numerical stability constant (Sec 6.5).
  • epsilon (exploration mass) = 0.10
    Chosen ex ante for the deployment; determines static-arm traffic and enters the regret optimization (Sec 6.1).
axioms (5)
  • domain assumption Conditional cohort-invariance: pooled LAAL and static observations are comparable after conditioning on query, time, and realized ad load.
    Stated in Sec 6.2 and tested in Web Appendix E; needed for LAAL's pooled reward estimates.
  • ad hoc to paper Assumption E2: each search's cohort is drawn independently across searches; realized query-level static support satisfies (A5)-(A6).
    Used in the Appendix F regret proof; inconsistent with persistent user-level assignment.
  • domain assumption Assumption E5: regular query rate, so each h-length window contains Theta(Nh/T) searches of a given query.
    Used to derive the optimized regret rate in Appendix F; not empirically verified.
  • domain assumption Random assignment via hashed device IDs yields exchangeable user cohorts.
    Basis for causal identification in Sec 3.2.
  • domain assumption The ad market is sufficiently thick so realized ad load matches intended treatment condition.
    Stated in Sec 3.2 footnote; needed for treatment-effect interpretation.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive Ad Load Design for Sponsored Search Markets: Evidence, Theory, and Deployment." pith.science (2026). https://pith.science/paper/WSOFASEG

@misc{pith2026260714418,
  author       = {Pith},
  title        = {Pith review of: Adaptive Ad Load Design for Sponsored Search Markets: Evidence, Theory, and Deployment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WSOFASEG}},
  note         = {Machine review of arXiv:2607.14418}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Ad-load design is a central supply-side decision in sponsored search: more sponsored slots can raise revenue, but may crowd out organic results and degrade user outcomes. We study this trade-off using a large-scale randomized field experiment on an Android app store, where over five million users are exposed to one through six sponsored slots. Increasing ad load raises revenue by up to 43%, but reduces total search conversions by up to 5% and daily engagement by up to 2.2%. These average effects mask substantial heterogeneity: additional slots generate large revenue gains for high-ad-conversion queries, but little or negative marginal revenue for low-conversion queries. The trade-off also shifts within query as advertiser composition changes, such as brand-advertiser presence. Motivated by these findings, we design and deploy a novel adaptive algorithm -- exploration-augmented Locally Adaptive Ad Load (e-LAAL). e-LAAL combines LAAL, a model-free query-level decision rule that updates ad-load recommendations using recent outcomes, with static exploration arms that maintain support and provide fixed-policy counterfactual benchmarks. We provide a finite-time dynamic-regret guarantee for the e-LAAL architecture. In a platform-level production deployment serving 22.3 million users and 77.6 million searches, e-LAAL improves the empirical revenue--conversion trade-off relative to deployed static benchmarks and outperforms uniform and historical query-dependent static benchmarks.

Figures

Figures reproduced from arXiv: 2607.14418 by Hema Yoganarasimhan, Mohammad Rashid.

Figure 1
Figure 1. Figure 1: Comparison of single-ad versus three-ad sponsored search layouts. [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Treatment effects by Ad CVR segment. 4.2.2 Heterogeneity Over Time We now examine how the revenue-conversion trade-off varies over time using the presence or absence of brand advertising as a salient time-varying market-state. In this analysis, we focus on brand queries – queries which refer to a specific brand/app, e.g., WhatsApp. We first split the 8,561 brand queries in our data into two groups: • Brand… view at source ↗
Figure 3
Figure 3. Figure 3: Revenue-Conversion tradeoff based on brand activity. Each point corresponds to an experimental [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Distribution of Brand-On searches within Brand-Active Queries [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The effect of brand advertising on left: conversion and right: revenue. The gaps equal to the brand [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Sliding-window update for two searches of the same query [PITH_FULL_IMAGE:figures/full_fig_p021_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Illustration of LAAL’s sliding-window adaptation. As the 3-day window shifts, estimated rewards [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Performance of LAAL (and system-level e-LAAL). [PITH_FULL_IMAGE:figures/full_fig_p028_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Revenue–conversion trade-offs across query segments. [PITH_FULL_IMAGE:figures/full_fig_p029_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Distribution of ad loads recommended by LAAL across query segments. [PITH_FULL_IMAGE:figures/full_fig_p029_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Performance of LAAL and e-LAAL on non-stationary queries (brand entry/exit) [PITH_FULL_IMAGE:figures/full_fig_p029_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Distribution of LAAL ad-load recommendations under [PITH_FULL_IMAGE:figures/full_fig_p030_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Revenue–conversion comparison of LAAL with static benchmarks. The figure shows the static [PITH_FULL_IMAGE:figures/full_fig_p034_13.png] view at source ↗
Figure 13
Figure 13. Figure 13: We see that the historical query-dependent benchmark improves the empirical revenue–conversion [PITH_FULL_IMAGE:figures/full_fig_p035_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Average empirical regret per search of LAAL relative to the deployment query-dependent empirical [PITH_FULL_IMAGE:figures/full_fig_p036_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Revenue–conversion across fixed ad-load cohorts, LAAL, and e-LAAL on [PITH_FULL_IMAGE:figures/full_fig_p038_15.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

131 extracted references · 12 canonical work pages · 1 internal anchor

  1. [1]

    Abbasi-Yadkori, D

    Y. Abbasi-Yadkori, D. P \'a l, and C. Szepesv \'a ri. Improved algorithms for linear stochastic bandits. In Advances in Neural Information Processing Systems 24, pages 2312--2320, 2011. URL https://proceedings.neurips.cc/paper/2011/hash/e1d5be1c7f2f456670de3d53c7b54f4a-Abstract.html

    2011

  2. [2]

    Agrawal and N

    S. Agrawal and N. Goyal. Thompson sampling for contextual bandits with linear payoffs. In Proceedings of the 30th International Conference on Machine Learning, volume 28 of Proceedings of Machine Learning Research, pages 127--135. PMLR, 2013. URL https://proceedings.mlr.press/v28/agrawal13.html

    2013

  3. [3]

    Ads on the app store

    Apple Ads . Ads on the app store. https://ads.apple.com/app-store, 2026. Accessed: 2026-06-02

    2026

  4. [4]

    Athey and G

    S. Athey and G. Ellison. Position auctions with consumer search. Quarterly Journal of Economics, 126 0 (3): 0 1213--1270, 2011. doi:10.1093/qje/qjr028

    doi:10.1093/qje/qjr028 2011

  5. [6]

    Carri \'o n, J

    A. Carri \'o n, J. Wang, H. S. Nair, W. Xu, and Y. Zhang. Blending advertising with organic content in e-commerce: A virtual bids optimization approach. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), volume 37, pages 4141--4149. AAAI Press, 2023. doi:10.1609/aaai.v37i4.25629. Also available as arXiv preprint arXiv:2105.13556

    Pith/arXiv arXiv 2023

  6. [7]

    Cavenaghi, G

    E. Cavenaghi, G. Sottocornola, F. Stella, and M. Zanker. Non-stationary multi-armed bandit: Empirical evaluation of a new concept drift-aware algorithm. Entropy, 23 0 (3): 0 380, 2021. doi:10.3390/e23030380

    work page doi:10.3390/e23030380 2021

  7. [9]

    W. C. Cheung, D. Simchi-Levi, and R. Zhu. Learning to optimize under non-stationarity. In Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics (AISTATS), volume 89 of Proceedings of Machine Learning Research, pages 1079--1087. PMLR, 2019

    2019

  8. [10]

    W. Chu, L. Li, L. Reyzin, and R. Schapire. Contextual bandits with linear payoff functions. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, volume 15 of Proceedings of Machine Learning Research, pages 208--214. PMLR, 2011. URL https://proceedings.mlr.press/v15/chu11a.html

    2011

  9. [11]

    K. Deb, K. Sindhya, and J. Hakanen. Multi-objective optimization. In Decision sciences, pages 161--200. CRC Press, 2016

    2016

  10. [13]

    Garivier and \'E

    A. Garivier and \'E . Moulines. On upper-confidence bound policies for switching bandit problems. In International Conference on Algorithmic Learning Theory (ALT), pages 174--188, 2011. doi:10.1007/978-3-642-24412-4_16. URL https://doi.org/10.1007/978-3-642-24412-4_16

    work page doi:10.1007/978-3-642-24412-4_16 2011

  11. [14]

    Goldfarb and C

    A. Goldfarb and C. E. Tucker. Privacy regulation and online advertising. Management Science, 57 0 (1): 0 57--71, 2011. doi:10.1287/mnsc.1100.1246

    arXiv 2011

  12. [15]

    A. Goli, J. Huang, D. Reiley, and N. M. Riabov. Measuring consumer sensitivity to audio advertising: a long-run field experiment on pandora internet radio. Quantitative Marketing and Economics, pages 1--31, 2025 a

    2025

  13. [17]

    D. G. Horvitz and D. J. Thompson. A generalization of sampling without replacement from a finite universe. Journal of the American Statistical Association, 47 0 (260): 0 663--685, 1952. doi:10.1080/01621459.1952.10483446

    arXiv 1952

  14. [18]

    Internet advertising revenue report: Full-year 2025 results

    Interactive Advertising Bureau . Internet advertising revenue report: Full-year 2025 results. https://www.iab.com/wp-content/uploads/2026/04/IAB_PwC_Internet_Ad_Revenue_Report_Full_Year_2025_April_2026.pdf, Apr. 2026. Accessed: 2026-06-01

    2025

  15. [19]

    L. Jain, Z. Li, E. Loghmani, B. Mason, and H. Yoganarasimhan. Effective adaptive exploration of prices and promotions in choice-based demand models. Marketing Science, 43 0 (5): 0 1002--1030, 2024

    2024

  16. [20]

    A. Kim, S. Balachander, and K. Kannan. On the optimal number of advertising slots in a generalized second-price auction. Marketing Letters, 23 0 (3): 0 851--868, Sept. 2012. doi:10.1007/s11002-012-9193-2

    work page doi:10.1007/s11002-012-9193-2 2012

  17. [21]

    Kohavi, A

    R. Kohavi, A. Deng, B. Frasca, R. Longbotham, T. Walker, and Y. Xu. Trustworthy online controlled experiments: Five puzzling outcomes explained. In Proceedings of the 18th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 786--794, 2012

    2012

  18. [22]

    Lahaie and D

    S. Lahaie and D. M. Pennock. Revenue analysis of a family of ranking rules for keyword auctions. In Proceedings of the 8th ACM Conference on Electronic Commerce (EC), pages 50--56, 2007. doi:10.1145/1250910.1250918

    arXiv 2007

  19. [23]

    Lattimore and C

    T. Lattimore and C. Szepesv \'a ri. Bandit algorithms. Cambridge University Press, 2020

    2020

  20. [24]

    R. P. Leme et al. A field guide to personalized reserve prices, 2016. arXiv preprint

    2016

  21. [25]

    L. Li, W. Chu, J. Langford, and R. E. Schapire. A contextual-bandit approach to personalized news article recommendation. In Proceedings of the 19th international conference on World wide web, pages 661--670, 2010

    2010

  22. [26]

    G. Liao, Z. Wang, X. Wu, X. Shi, C. Zhang, Y. Wang, X. Wang, and D. Wang. Cross DQN : Cross deep Q network for ads allocation in feed, 2021. URL https://arxiv.org/abs/2109.04353. Accepted by WWW '22

    Pith/arXiv arXiv 2021

  23. [29]

    Ostrovsky and M

    M. Ostrovsky and M. Schwarz. Reserve prices in internet advertising auctions: A field experiment. Journal of Political Economy, 131 0 (12), 2023. doi:10.1086/725702

    work page doi:10.1086/725702 2023

  24. [30]

    Rafieian and H

    O. Rafieian and H. Yoganarasimhan. Targeting and privacy in mobile advertising. Marketing Science, 40 0 (2): 0 193--218, 2021. doi:10.1287/mksc.2020.1235

    arXiv 2021

  25. [31]

    Rafieian, A

    O. Rafieian, A. Kapoor, and A. Sharma. Multiobjective personalization of marketing interventions. Marketing Science, 44 0 (2): 0 457--477, 2025

    2025

  26. [32]

    Russac, C

    Y. Russac, C. Vernade, and O. Capp \'e . Weighted linear bandits for non-stationary environments. In Advances in Neural Information Processing Systems 32, pages 12040--12049, 2019. URL https://papers.nips.cc/paper/9372-weighted-linear-bandits-for-non-stationary-environments

    2019

  27. [33]

    Sagtani, M

    H. Sagtani, M. G. Jhawar, R. Mehrotra, and O. Jeunen. Ad-load balancing via off-policy learning in a content marketplace, 2023. URL https://arxiv.org/abs/2309.11518

    Pith/arXiv arXiv 2023

  28. [35]

    A. Sayedi. Real-time bidding in online display advertising. Marketing Science, 37 0 (4): 0 553--568, 2018. doi:10.1287/mksc.2017.1083

    arXiv 2018

  29. [37]

    Simonov, C

    A. Simonov, C. Nosko, and J. M. Rao. Competition and crowd-out for brand keywords in sponsored search. Marketing Science eJournal, 2017. URL https://api.semanticscholar.org/CorpusID:947920

    2017

  30. [38]

    Tr \`o v \`o , S

    F. Tr \`o v \`o , S. Paladino, M. Restelli, and N. Gatti. Sliding-window thompson sampling for non-stationary settings. Journal of Artificial Intelligence Research, 68: 0 311--364, 2020. doi:10.1613/jair.1.11407

    work page doi:10.1613/jair.1.11407 2020

  31. [39]

    H. R. Varian. Position auctions. International Journal of Industrial Organization, 25 0 (6): 0 1163--1178, 2007. doi:10.1016/j.ijio.2006.10.006

    work page doi:10.1016/j.ijio.2006.10.006 2007

  32. [40]

    B. Wang, Z. Li, J. Tang, K. Zhang, S. Chen, and L. Ru. Learning to advertise: How many ads are enough? In J. Z. Huang, L. Cao, and J. Srivastava, editors, Advances in Knowledge Discovery and Data Mining: 15th Pacific-Asia Conference, PAKDD 2011, Shenzhen, China, May 24--27, 2011, Proceedings, Part II, volume 6635 of Lecture Notes in Computer Science, page...

    work page doi:10.1007/978-3-642-20847-8_42 2011

  33. [41]

    Y. Wang, L. Tao, and X. X. Zhang. Recommending for a multi-sided marketplace: A multi-objective hierarchical approach. Marketing Science, 2024

    2024

  34. [43]

    K. C. Wilbur, L. Xu, and D. Kempe. Correcting audience externalities in television advertising. Marketing Science, 32 0 (6): 0 892--912, 2013

    2013

  35. [44]

    M. Yan, J. Xu, A. A. Tiwana, and M. Chatterjee. Ads allocation in feed via constrained optimization. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), pages 3299--3307. ACM, 2020. doi:10.1145/3394486.3403391

    arXiv 2020

  36. [45]

    Yao and C

    S. Yao and C. F. Mela. A dynamic model of sponsored search advertising. Marketing Science, 30 0 (3): 0 447--468, 2011. doi:10.1287/mksc.1100.0626

    arXiv 2011

  37. [46]

    Z. Ye, D. J. Zhang, H. Zhang, R. Zhang, X. Chen, and Z. Xu. Cold start to improve market thickness on online advertising platforms: Data-driven algorithms and field experiments. Management Science, 69 0 (7): 0 3838--3860, 2023

    2023

  38. [47]

    Yoganarasimhan

    H. Yoganarasimhan. Search personalization using machine learning. Management Science, 66 0 (3): 0 1045--1070, 2020

    2020

  39. [48]

    Yoganarasimhan, E

    H. Yoganarasimhan, E. Barzegary, and A. Pani. Design and evaluation of optimal free trials. Management Science, 69 0 (6): 0 3220--3240, 2023

    2023

  40. [49]

    S. Yuan, J. Wang, B. Chen, P. Mason, and S. Seljan. An empirical study of reserve price optimisation in real-time bidding. In Proceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD), pages 1897--1906. ACM, 2014. doi:10.1145/2623330.2623357

    arXiv 1906

  41. [51]

    Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics , series=

    Contextual Bandits with Linear Payoff Functions , author=. Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics , series=. 2011 , publisher=

    2011

  42. [52]

    Marketing Science , volume=

    Correcting audience externalities in television advertising , author=. Marketing Science , volume=. 2013 , publisher=

    2013

  43. [53]

    Ads on the App Store , year =

  44. [54]

    Proceedings of the 18th ACM SIGKDD international conference on Knowledge discovery and data mining , pages=

    Trustworthy online controlled experiments: Five puzzling outcomes explained , author=. Proceedings of the 18th ACM SIGKDD international conference on Knowledge discovery and data mining , pages=

  45. [55]

    Internet Advertising Revenue Report: Full-Year 2025 Results , year =

    2025

  46. [56]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Fast Video Classification via Adaptive Cascading of Deep Models , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=. 2017 , url=

    2017

  47. [57]

    Computational Economics , year=

    Competitive Pricing Using Model-Based Bandits , author=. Computational Economics , year=. doi:10.1007/s10614-024-10816-w , url=

    work page doi:10.1007/s10614-024-10816-w

  48. [58]

    International Conference on Algorithmic Learning Theory (ALT) , pages=

    On Upper-Confidence Bound Policies for Switching Bandit Problems , author=. International Conference on Algorithmic Learning Theory (ALT) , pages=. 2011 , doi=

    2011

  49. [59]

    2020 , publisher=

    Bandit algorithms , author=. 2020 , publisher=

    2020

  50. [60]

    Horvitz, D. G. and Thompson, D. J. , title =. Journal of the American Statistical Association , year =

  51. [61]

    Management Science , volume=

    Design and evaluation of optimal free trials , author=. Management Science , volume=. 2023 , publisher=

    2023

  52. [62]

    Marketing Science , volume=

    Effective adaptive exploration of prices and promotions in choice-based demand models , author=. Marketing Science , volume=. 2024 , publisher=

    2024

  53. [63]

    Advances in Neural Information Processing Systems 24 , pages=

    Improved Algorithms for Linear Stochastic Bandits , author=. Advances in Neural Information Processing Systems 24 , pages=. 2011 , url=

    2011

  54. [64]

    Proceedings of the 30th International Conference on Machine Learning , series=

    Thompson Sampling for Contextual Bandits with Linear Payoffs , author=. Proceedings of the 30th International Conference on Machine Learning , series=. 2013 , publisher=

    2013

  55. [65]

    Advances in Neural Information Processing Systems 27 , pages=

    Stochastic Multi-Armed-Bandit Problem with Non-stationary Rewards , author=. Advances in Neural Information Processing Systems 27 , pages=. 2014 , url=

    2014

  56. [66]

    Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics , series=

    Learning to Optimize under Non-Stationarity , author=. Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics , series=. 2019 , publisher=

    2019

  57. [67]

    Proceedings of the 19th international conference on World wide web , pages=

    A contextual-bandit approach to personalized news article recommendation , author=. Proceedings of the 19th international conference on World wide web , pages=

  58. [68]

    Advances in Neural Information Processing Systems 32 , pages=

    Weighted Linear Bandits for Non-Stationary Environments , author=. Advances in Neural Information Processing Systems 32 , pages=. 2019 , url=

    2019

  59. [69]

    Proceedings of the tenth ACM international conference on web search and data mining , pages=

    Unbiased learning-to-rank with biased feedback , author=. Proceedings of the tenth ACM international conference on web search and data mining , pages=

  60. [70]

    Management science , volume=

    Advertiser prominence effects in search advertising , author=. Management science , volume=. 2018 , publisher=

    2018

  61. [71]

    Marketing Science , year=

    Recommending for a multi-sided marketplace: A multi-objective hierarchical approach , author=. Marketing Science , year=

  62. [72]

    Marketing Science , volume=

    Multiobjective personalization of marketing interventions , author=. Marketing Science , volume=. 2025 , publisher=

    2025

  63. [73]

    arXiv preprint arXiv:2504.13444 , year=

    Balancing Engagement and Polarization: Multi-Objective Alignment of News Content Using LLMs , author=. arXiv preprint arXiv:2504.13444 , year=

    arXiv

  64. [74]

    Decision sciences , pages=

    Multi-objective optimization , author=. Decision sciences , pages=. 2016 , publisher=

    2016

  65. [75]

    Management Science , volume=

    Search personalization using machine learning , author=. Management Science , volume=. 2020 , publisher=

    2020

  66. [76]

    Journal of marketing research , volume=

    Location, location, location: An analysis of profitability of position in online advertising markets , author=. Journal of marketing research , volume=. 2011 , publisher=

    2011

  67. [77]

    Proceedings of the 2008 international conference on web search and data mining , pages=

    An experimental comparison of click position-bias models , author=. Proceedings of the 2008 international conference on web search and data mining , pages=

    2008

  68. [78]

    2014 , url=

    Position Effects in Search Advertising: A Regression Discontinuity Approach , author=. 2014 , url=

    2014

  69. [79]

    Proceedings of the eleventh ACM international conference on web search and data mining , pages=

    Position bias estimation for unbiased learning to rank in personal search , author=. Proceedings of the eleventh ACM international conference on web search and data mining , pages=

  70. [80]

    The Power of Rankings: Quantifying the Effect of Rankings on Online Consumer Search and Purchase Decisions , author=. Mark. Sci. , year=

  71. [81]

    Consumer Search Activities and the Value of Ad Positions in Sponsored Search Advertising , author=. Mark. Sci. , year=

  72. [82]

    Behavioral Marketing eJournal , year=

    Examining the Impact of Ranking on Consumer Behavior and Search Engine Revenue , author=. Behavioral Marketing eJournal , year=

  73. [83]

    What Makes Them Click: Empirical Analysis of Consumer Demand for Search Advertising , urldate =

    Przemyslaw Jeziorski and Ilya Segal , journal =. What Makes Them Click: Empirical Analysis of Consumer Demand for Search Advertising , urldate =

  74. [84]

    1978 , publisher=

    Optimal search for the best alternative , author=. 1978 , publisher=

    1978

  75. [85]

    Quantitative Marketing and Economics , pages=

    Measuring consumer sensitivity to audio advertising: a long-run field experiment on Pandora internet radio , author=. Quantitative Marketing and Economics , pages=. 2025 , publisher=

    2025

  76. [86]

    Memory, and Conversion Outcomes (June 21, 2024) , year=

    The Role of Digital Ad Clutter in Ad Viewership, Memory, and Conversion Outcomes , author=. Memory, and Conversion Outcomes (June 21, 2024) , year=

    2024

  77. [87]

    Proceedings of the 16th international conference on World Wide Web , pages=

    Predicting clicks: estimating the click-through rate for new ads , author=. Proceedings of the 16th international conference on World Wide Web , pages=

  78. [88]

    Marketing Science eJournal , year=

    Competition and Crowd-Out for Brand Keywords in Sponsored Search , author=. Marketing Science eJournal , year=

  79. [89]

    and Li, Sai-Ming and Lewis, Randall A

    Reiley, David H. and Li, Sai-Ming and Lewis, Randall A. , title =. Proceedings of the 11th ACM Conference on Electronic Commerce , pages =. 2010 , isbn =. doi:10.1145/1807342.1807390 , abstract =

    arXiv 2010

  80. [90]

    International Conference on Machine Learning , pages=

    Pareto regret analyses in multi-objective multi-armed bandit , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.