REVIEW 3 major objections 6 minor 12 references
Impact of Rankings and Personalized Recommendations in Marketplaces
T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read Public rankings add zero welfare when supply is limited; personalized recommendations deliver all the value.
desk verdict A clean theory paper with genuinely new asymptotic results; the zero-gain ranking theorem is real but narrower than the abstract claims, needing a scope caveat rather than a rewrite. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the assignment invariance of the common term. Because agents and unit-capacity items are equal in number, every allocation consumes each item exactly once, so $\sum_{k=1}^n q_{\sigma(k)} = \sum_{k=1}^n q_k$ for every policy; this makes the common-quality part of total welfare a constant, which is what forces the public-ranking gain to zero. The second mechanism is the deferred-decisions principle: when agent $k$ chooses, her idiosyncratic values for the remaining items can be treated as freshly drawn, so under Full Information her realized value is at least the maximum of $n-k$ i.i.d. idiosyncratic draws, producing the $\rho\,n^{1/\alpha_\phi}$ scaling. In the uncapacitated setting the defining object is the random variable $Z=(1-\rho)X+\rho Y$; the paper shows it inherits a Pareto tail, with scale $c_Z=(((1-\rho)c_X)^\alpha+(\rho c_Y)^\alpha)^{1/\alpha}$ when the tail exponents match, and this identity carries the entire comparison between rankings and personalization.
What would settle it
In the paper's capacitated model, compute the welfare gap between Only Quality Information and No Information after adding one extra item (n agents, n+1 items) or after making the idiosyncratic term correlated across agents. If the gap becomes nonzero, the balanced-market and i.i.d.-refresh assumptions are what carry the zero result; a direct simulation would show $\Delta^{\mathrm{cap}}_{\emptyset\to q}(n)=0$ for every $n$ when the assumptions hold.
Extended reading notes
Core claim
The central discovery is an exact zero: in a balanced market with $n$ agents and $n$ items of unit capacity, the Only Quality Information regime and the No Information regime have identical expected average welfare, $\Delta^{\mathrm{cap}}_{\emptyset\to q}(n)=0$, for any common-term distribution with finite mean. The reason is that every feasible assignment consumes every item exactly once, so the total common-quality contribution $\sum_y q_y$ is invariant across policies; rankings only redistribute items among agents, and each agent's idiosyncratic draw for her assigned item is a fresh independent sample. In the same setting, personalized recommendations increase welfare by $\rho \cdot C_\phi\, n^{1/\alpha_\phi}$ when the idiosyncratic terms have Pareto tails, and by $\rho\cdot(\ln n)/\lambda_\phi$ for exponential tails, because the agent choosing at step $k$ can effectively select the maximum of $n-k$ fresh idiosyncratic draws. In the uncapacitated setting, public rankings raise welfare by $(1-\rho)$ times the expected maximum of $n$ common-term draws, and personalized recommendations add further gains that depend on which tail is heavier, with equal-tail Pareto distributions giving an incremental gain $(((1-\rho)^\alpha c_q^\alpha + \rho^\alpha c_\phi^\alpha)^{1/\alpha} - (1-\rho)c_q)$ times the maximum scaling.
Load-bearing premise
The result depends on a perfectly balanced market, exactly as many agents as unit-capacity items, so that every item is consumed in every allocation; if supply exceeds demand, rankings can steer people to better items and the zero-gain finding collapses.
Editorial extensions
If this is right
- In centralized college or school admissions with matched applicant and seat counts, publishing a quality ranking will not raise average applicant welfare; the marginal investment is in personalization.
- On content platforms with effectively unlimited supply, public rankings deliver most of the welfare gain when preferences are homogeneous, while personalization is what pays off as heterogeneity crosses a threshold.
- The welfare gain from personalized recommendations in capacitated markets is asymptotically linear in the heterogeneity weight $\rho$ and governed by the tail of the idiosyncratic taste distribution, so heavy-tailed individual preferences are where the tool's value concentrates.
- Because the sum-invariance argument is distribution-free, the zero-gain conclusion for public rankings in balanced capacitated markets holds for arbitrary common-term distributions with finite mean, not just Pareto or exponential tails.
Reading between the lines
- If the balanced-market assumption breaks, say more items than agents so some items go unconsumed, a public ranking would steer demand toward high-quality items that would otherwise be wasted; the zero-gain result is specific to full consumption, which the paper does not test.
- The paper measures average welfare; a ranking could systematically shift utility from one group of agents to another while leaving the average untouched, so distributional effects are an open question the model cannot resolve.
- The linearity in $\rho$ suggests a practical prior: platforms could estimate $\rho$ from the dispersion of users' past choices and use the model's scaling laws to decide whether to build a personalization engine before committing to it.
- The same invariance argument implies the zero result transfers to any two-sided market where all supply is consumed, but introducing outside options, reserve capacities, or endogenous prices could re-create a role for public rankings.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper studies a stylized marketplace model in which agents choose items with utility equal to a weighted sum of a common quality term q_y and an idiosyncratic term φ_xy, under three information regimes: no information, public rankings (revealing only q), and personalized recommendations (revealing both). Two supply settings are considered: uncapacitated supply, where any number of agents may consume an item, and capacitated supply, where each item is matched to at most one agent. The main results characterize the asymptotic welfare gains of moving between regimes when q and φ have Pareto or exponential tails. In the uncapacitated setting, public rankings capture welfare gains that scale with (1−ρ) times the top order statistic of q, and personalized recommendations add further gains whose size depends on the relative tail heaviness of q and φ and on ρ. In the capacitated setting, the paper proves that public rankings yield exactly zero welfare gain relative to no information, while personalized recommendations produce gains that scale with ρ times the top order statistic of φ. The paper also reports partial results for bounded distributions in an appendix.
Significance. If the main results hold, the paper provides crisp asymptotic comparative statics for a question of clear practical importance: when should a platform or policymaker invest in public rankings versus personalized recommendations? The strength of the paper is that the welfare rates are derived from model primitives rather than fitted to data, giving falsifiable predictions about how the value of information tools scales with market size and preference heterogeneity. The zero-gain theorem for public rankings in fully capacitated markets is a striking and clean result, and the proofs of Theorems 1–4 are generally coherent and based on standard extreme-value arguments. The main limitation is that the headline 'rankings are useless under capacity' conclusion is driven by the balanced-market, full-consumption assumption, and the paper currently presents this as a general insight without making that caveat sufficiently prominent. The appendix also contains an inconsistent statement about bounded distributions.
major comments (3)
- [§3.2 (Theorem 3(a)), §4.3.1, and §5] The zero-gain result of Theorem 3(a) is an accounting identity driven by the balanced-market assumption that there are exactly n agents and n unit-capacity items, so that every item is consumed exactly once and the sum of common terms ∑_y q_y is the same in every regime. The proof in Section 4.3.1 explicitly uses this invariance. If supply exceeds demand, so that more items exist than agents, the set of consumed items becomes endogenous and public rankings have strictly positive welfare value because agents select the highest-q items rather than a random subset. The abstract and Section 5 state that in supply-constrained settings public rankings provide 'limited benefit' or 'no value' without adequately restricting this claim to the balanced, full-consumption case. Please qualify the policy conclusions accordingly and, ideally, provide an explicit treatment or a bound for the excess-supply case.
- [Appendix C, Theorem C.0.b and Remark C.0] Theorem C.0.b states the exact limit lim_{n→∞} Δ_{q→u}^{cap}(n) = ρ(b − μ_φ), but the proof cites Lemma 2, which yields only the bounds ρΦ_n − (1−ρ)μ_q − ρμ_φ ≤ Δ_{q→u}^{cap}(n) ≤ ρΦ_n − ρμ_φ. Since Φ_n → b, these bounds give ρb − (1−ρ)μ_q − ρμ_φ ≤ liminf Δ ≤ limsup Δ ≤ ρ(b − μ_φ), and these two quantities are not equal in general. Remark C.0 itself acknowledges a gap between upper and lower bounds. As written, the theorem and its proof are inconsistent: either a complete proof of the claimed equality must be supplied, or the statement should be weakened to an upper bound or a conjecture.
- [§4.3.2, proof of Lemma 2] The 'deferred decisions' argument in Lemma 2 considers the maximum over n − k idiosyncratic draws, but when agent k makes a choice there are n − k + 1 remaining items. The off-by-one error is asymptotically immaterial for the theorems that use the lemma, but the lemma statement and proof should use the correct number of remaining draws.
minor comments (6)
- [§3.1, Theorem 1(a)] The limit in Theorem 1(a) is written with Δ_{∅→φ}^{uncap}; this should be Δ_{∅→q}^{uncap} throughout the statement and discussion.
- [§3.1, after Theorem 1] The sentence 'Theorem (3.a) quantifies the improvement from No Information to Only Quality Information' refers to a theorem in the capacitated section; it should say Theorem (1.a).
- [Figure 3 caption] The caption writes the normalizing constant as c_q Γ(1 − α_q); it should be c_q Γ(1 − 1/α_q).
- [§4.2.1, proof of Theorem 1(a)] The proof says the last equality follows because σ_q(1) is uniformly random; the more precise reason is that φ is independent of q and is i.i.d. across items, so the distribution of the selected φ is still P_φ.
- [Appendix B.1] The derivation of exponential-tail results from Pareto-tail results via the double limit α = ln n, c = ln n/λ invokes the Moore-Osgood theorem, but the hypotheses for interchanging the limits are not verified. Since separate proofs are provided in Appendices B.3 and B.4, this passage should be presented as an intuition rather than a formal derivation.
- [Throughout] There are several typographical errors, including 'asympotitic rates' in Section 1.1 and 'user bahavior' and 'idiosynratic' in Section 1.2; these should be corrected.
Circularity Check
No circularity: the zero-gain result for public rankings is a transparent implication of the balanced-market assumption, and all welfare scalings are derived from primitives rather than fitted to data.
full rationale
This is a pure theory paper with no data fitting, no calibration, and no post-hoc exclusions. Each claimed result is derived from model primitives (rho, tail parameters, distributional assumptions) via explicit proofs. Theorem 3(a) computes both regimes and shows that Delta^cap_{emptyset->q}=0 because, with n agents and n unit-capacity items, every item is consumed exactly once, so sum q_{sigmaemptyset(k)} = sum q_k = sum q_{sigma_q(k)}, while E[phi_{k sigmaemptyset(k)}] = E[phi_{k sigma_q(k)}] = mu_phi by independence of the idiosyncratic terms. This is an algebraic consequence of the explicitly stated balanced-market/full-consumption assumption, not an input disguised as an output; the invariance is exhibited in equations (5)-(6). The concern that the policy claim may overreach to excess-supply settings is a scope/correctness issue, not a circularity issue. There is no fitted parameter called a prediction, no uniqueness theorem imported from the authors' prior work, and no ansatz smuggled in via self-citation. The only self-citation (Baswana et al. 2019, co-authored by Kanoria) is used to motivate the serial-dictatorship model of Indian college admissions; it is not load-bearing for any theorem and does not raise the circularity score. The paper also transparently flags its own limitations, including the independence assumption and the gap between bounds for bounded distributions, which further supports a non-circular reading.
Assumptions & free parameters
free parameters (4)
- rho
- alpha_q and alpha_phi (Pareto tail exponents)
- c_q and c_phi (Pareto scale parameters)
- lambda_q and lambda_phi (exponential rates)
assumptions (7)
- domain assumption Agent utility is u_xy = (1-rho)q_y + rho phi_xy with q_y common and phi_xy idiosyncratic, all independent.
- domain assumption Balanced market: n agents, n items, each item has unit capacity (or total capacity equals number of agents), so every item is consumed exactly once.
- domain assumption Agents choose sequentially in priority order, myopically maximizing perceived utility with random tie-breaking.
- domain assumption The common and idiosyncratic terms have non-negative support, finite mean, and Pareto tails (Definition 1) or exponential tails (Definition 2).
- standard math Standard extreme value and asymptotic results: E[max of n i.i.d. Pareto draws] ~ c Gamma(1-1/alpha) n^(1/alpha), and E[max of n exponential draws] ~ ln n / lambda.
- standard math Tail of a weighted sum of independent heavy/light-tailed variables is dominated by the heavier tail (Lemma 1 and Proposition B.3).
- standard math Deferred decisions principle: drawing phi values for remaining items at the time an agent chooses is equivalent to drawing all phi in advance.
Cite this review
Pith. "Pith review of Impact of Rankings and Personalized Recommendations in Marketplaces." pith.science (2026). https://pith.science/paper/G6AVODI7
@misc{pith2026250603369,
author = {Pith},
title = {Pith review of: Impact of Rankings and Personalized Recommendations in Marketplaces},
year = {2026},
howpublished = {\url{https://pith.science/paper/G6AVODI7}},
note = {Machine review of arXiv:2506.03369}
}
read the original abstract
Individuals often navigate several options with incomplete knowledge of their own preferences. Information provisioning tools such as public rankings and personalized recommendations have become central to helping individuals make choices, yet their value proposition under different marketplace environments remains unexplored. This paper studies a stylized model to explore the impact of these tools in two marketplace settings: uncapacitated supply, where items can be selected by any number of agents, and capacitated supply, where each item is constrained to be matched to a single agent. We model the agents utility as a weighted combination of a common term which depends only on the item, reflecting the item's population level quality, and an idiosyncratic term, which depends on the agent item pair capturing individual specific tastes. Public rankings reveal the common term, while personalized recommendations reveal both terms. In the supply unconstrained settings, both public rankings and personalized recommendations improve welfare, with their relative value determined by the degree of preference heterogeneity. Public rankings are effective when preferences are relatively homogeneous, while personalized recommendations become critical as heterogeneity increases. In contrast, in supply constrained settings, revealing just the common term, as done by public rankings, provides limited benefit since the total common value available is limited by capacity constraints, whereas personalized recommendations, by revealing both common and idiosyncratic terms, significantly enhance welfare by enabling agents to match with items they idiosyncratically value highly. These results illustrate the interplay between supply constraints and preference heterogeneity in determining the effectiveness of information provisioning tools, offering insights for their design and deployment in diverse settings.
Figures
Figures from the paper (4 more)
Reference graph
Works this paper leans on
-
[1]
Now, define α = ln n, c = ln n/λ and consider the double limit x → ∞ and n → ∞. Now we have that 1 = lim x→∞ P(X > x ) exp(−λx) = lim x→∞ lim n→∞ P(X > x ) (1 + λx/ ln n)− ln n = lim x→∞ lim n→∞ P(X > x ) (1 + x/c)−α = lim n→∞ lim x→∞ P(X > x ) (1 + x/c)−α , where the interchange of limit follows from the Moore-Osgoo d theorem ( Taylor, 1985). As an illust...
work page 1985
-
[2]
Principle of Deferred Decisions
Note that ∆ cap q→u(n) = A Wcap u (n) − A Wcap q (n). We will begin by characterizing the social welfare in the Only Quality Information regime. In the Only Quality Information regime, the agents base their decisions solely on the common term ( qy). Therefore, we have that the agent k will choose the item with common term value q(k:n) (recall that X(k:n) ...
work page 2008
-
[7]
Optimizing rank ings for recommendation in match- ing markets
25 Yi Su, Magd Bayoumi, and Thorsten Joachims. Optimizing rank ings for recommendation in match- ing markets. In Proceedings of the ACM Web Conference 2022, pages 328–338,
work page 2022
-
[8]
Accessed: 2024-10-03. Zhehao Zhang, Ryan A Rossi, Branislav Kveton, Yijia Shao, Di yi Yang, Hamed Zamani, Franck Dernoncourt, Joe Barrow, Tong Yu, Sungchul Kim, et al. Perso nalization of large language models: A survey. arXiv preprint arXiv:2411.00027,
arXiv 2024
-
[11]
Define ¯c = (1 + ǫ)c and c = (1 − ǫ)c
Since X has an exponential tail with parameters c > 0 and λ > 0, there exists a constant x0 ≥ 0 such that for all x ≥ x0, we have that (1 − ǫ)c exp(−λx) ≤ P(X > x ) ≤ (1 + ǫ)c exp(−λx) (B.1) Next we want to bound the tail distribution for X(n:n) which is maximum of n i.i.d copies of X. Define ¯c = (1 + ǫ)c and c = (1 − ǫ)c. Using ( A.2) and ( B.1), we have...
work page 2019
-
[12]
Define Z = (1 − ρ)X + ρY . Let Z(n:n) = max {Z1, Z2, . . . , Zn} App-7 where Z1, Z2, . . . , Zn are i.i.d copies of Z. Then we have that lim n→∞ E[Z(n:n)] ln n = max { 1 − ρ λq , ρ λϕ } Proof of Proposition B.3. Fix ǫ > 0 and let ρ ∈ (0, 1). Define ˜X ≜ (1 − ρ)X and ˜Y ≜ ρY . We have that ˜X and ˜Y have exponential tails with parameters ( cX , λX/(1 − ρ)) a...
work page 2019
-
[2008]
Hiring under co ngestion and algorithmic mono- culture: Value of strategic behavior
23 Jackie Baek, Hamsa Bastani, and Shihan Chen. Hiring under co ngestion and algorithmic mono- culture: Value of strategic behavior. arXiv preprint arXiv:2502.20063,
-
[2016]
Online assortment optimizatio n for two-sided matching platforms
Ali Aouad and Daniela Saban. Online assortment optimizatio n for two-sided matching platforms. Management Science, 69(4):2069–2087,
Show all 12 references
-
[2017]
Accessed: 2024-07-31
URL https://news.gallup.com/reports/208535/gallup-strad a-education-ecs-education-regret.aspx . Accessed: 2024-07-31. 24 Julien Grenet, YingHua He, and Dorothea K¨ ubler. Preferenc e discovery in university admissions: The case for dynamic multioffer mechanisms. Journal of Poli...
2024
-
[2018]
Monoculture in matching markets
Kenny Peng and Nikhil Garg. Monoculture in matching markets . Advances in Neural Information Processing Systems, 37:81959–81991, 2024a. Kenny Peng and Nikhil Garg. Wisdom and foolishness of noisy m atching markets. arXiv preprint arXiv:2402.16771, 2024b. Alvin E Roth and Maril...
-
[2024]
Maxim Naumov, Dheevatsa Mudigere, Hao-Jun Michael Shi, Jia nyu Huang, Narayanan Sundara- man, Jongsoo Park, Xiaodong Wang, Udit Gupta, Carole-Jean W u, Alisson G Azzolini, et al
Ac- cessed: 2024-10-03. Maxim Naumov, Dheevatsa Mudigere, Hao-Jun Michael Shi, Jia nyu Huang, Narayanan Sundara- man, Jongsoo Park, Xiaodong Wang, Udit Gupta, Carole-Jean W u, Alisson G Azzolini, et al. Deep learning recommendation model for personalization an d recommendation...
2024 arXiv
-
[2025]
Centralized admissions for engineering c olleges in india
Surender Baswana, Partha Pratim Chakrabarti, Sharat Chand ran, Yashodhan Kanoria, and Utkarsh Patange. Centralized admissions for engineering c olleges in india. In Proceedings of the 2019 ACM Conference on Economics and Computation, pages 323–324,
2019
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.