Technical Summary: Impact of Rankings and Personalized Recommendations in Marketplaces

1. Problem Statement

Decision-making in large markets often occurs under incomplete information and imperfectly formed preferences. Agents rely on information provisioning tools—specifically public rankings (aggregating population-level quality) and personalized recommendations (revealing individual-specific fit)—to navigate choice sets. While these tools are ubiquitous, their aggregate welfare implications across different market environments, particularly regarding the interplay between supply constraints (uncapacitated vs. capacitated) and preference heterogeneity, remain poorly understood.

The central research question is: What are the implications of different information provisioning tools in environments with and without supply-side constraints? Specifically, the paper seeks to quantify the incremental value of personalization over public rankings and the value of rankings over no information, isolating the economic forces of scarcity, congestion, and heterogeneity.

2. Methodology

The authors develop a stylized, parsimonious large-market model to analytically quantify welfare gains.

Model Primitives

• Agents and Items: A market with $n$ agents and $n$ items.
• Utility Structure: An agent $x$'s utility for item $y$ is a weighted sum of two independent components:
  $$u_{xy} = (1 - \rho) q_y + \rho \phi_{xy}$$
  • $q_y$: Common term (population-level quality), drawn i.i.d. from $F_q$.
  • $\phi_{xy}$: Idiosyncratic term (individual-specific fit), drawn i.i.d. from $F_\phi$.
  • $\rho \in [0, 1]$: Heterogeneity parameter. Low $\rho$ implies homogeneous preferences; high $\rho$ implies heterogeneous preferences.
• Information Regimes: Agents observe noisy signals of these components:
  • No Information ($\emptyset$): Agents observe neither component.
  • Only Quality Information ($q$): Agents observe the common term (public rankings) but not the idiosyncratic term.
  • Full Information ($u$): Agents observe both components (personalized recommendations).
  • Intermediate noisy regimes are also considered to quantify the value of imperfect signals.
• Supply Settings:
  • Uncapacitated: Infinite capacity; any number of agents can choose the same item (e.g., digital content).
  • Capacitated: Unit capacity; items are consumed once. Agents choose sequentially based on priority scores (e.g., college admissions, Airbnb).

Analytical Approach

The paper utilizes extreme-value theory to characterize the asymptotic behavior of agent welfare as $n \to \infty$. Welfare is defined as the average realized utility across agents. The authors derive closed-form asymptotic expressions for welfare under Gaussian primitives and extend these results to alternative tail distributions (Pareto, Exponential, Bounded).

3. Key Contributions and Results

A. The Fundamental Role of Capacity Constraints

The paper establishes a stark dichotomy in how information affects welfare depending on supply constraints.

1. Uncapacitated Supply (Infinite Capacity)

In settings without capacity constraints, information improves welfare strictly through better selection.

• Public Rankings: When preferences are relatively homogeneous (small $\rho$), public rankings capture the first-order welfare gains because the common component drives utility.
• Personalized Recommendations: As heterogeneity ($\rho$) increases, the marginal value of personalization grows. In the Gaussian base model, the incremental value of personalization over rankings scales at the extreme-value rate of $\sqrt{2 \ln n}$.
• Mechanism: Agents select the item maximizing their perceived utility. Better signals allow agents to identify items with higher realizations of the underlying distributions.

2. Capacitated Supply (Unit Capacity)

Under capacity constraints, a strict conservation logic applies to the common component.

• Public Rankings (Zero Aggregate Value): Revealing common quality ($q_y$) only reshuffles assignments. Since each item is allocated exactly once, the aggregate common value $\sum q_y$ is fixed regardless of the matching. Therefore, public rankings yield zero aggregate average welfare gains ($\Delta_{\emptyset \to q} = 0$), even if they significantly alter the distribution of who gets which item.
• Personalized Recommendations (Substantial Value): Revealing idiosyncratic fit ($\phi_{xy}$) generates strictly positive aggregate welfare gains. By guiding agents to their specific best fits, personalization mitigates "congestion" on overdemanded high-quality items. This unlocks match-specific value that rankings cannot access.
• Scaling: The welfare gain from personalization scales linearly with $\rho$ and follows the extreme-value rate of the idiosyncratic distribution (e.g., $\sqrt{2 \ln n}$ for Gaussian).

B. Robustness Across Tail Distributions

The authors extend the analysis to non-Gaussian primitives (Pareto, Exponential, Bounded) to test the robustness of these insights.

• Uncapacitated: The scaling of welfare gains depends on the tail behavior of the distributions (polynomial for Pareto, logarithmic for Exponential, $\sqrt{\ln n}$ for Gaussian). A phase transition occurs in Exponential distributions: if the common term has a heavier tail than the idiosyncratic term, personalization offers no leading-order gain unless $\rho$ is sufficiently large.
• Capacitated: The core insight holds: rankings yield zero aggregate welfare gains regardless of tail behavior, while personalization yields substantial gains proportional to the tail of the idiosyncratic distribution.

C. Supply-Side Priorities and Pricing

• Supply-Side Priorities: In a model with "own-group" priorities (e.g., neighborhood school choice), the conservation of common value persists. Rankings reshuffle access within priority groups but do not increase average welfare. Personalization remains valuable by improving fit within the effectively available choice sets.
• Supply-Side Pricing: Introducing strategic pricing reveals a distributional dichotomy:
  • Excess Demand (Short Supply): Sellers are on the short side. Better information allows them to raise prices and extract the newly generated surplus. While total surplus increases with personalization, agent welfare remains unchanged (or improves negligibly) as sellers capture the gains.
  • Excess Supply (Short Demand): Buyers are on the short side. Competition disciplines prices, preventing sellers from extracting the surplus. Consequently, agents capture the full welfare gains unlocked by personalized recommendations.

4. Significance and Claims

The paper claims to provide a clean theoretical framework that isolates the mechanisms driving the value of information tools. Its primary significance lies in:

1. Formalizing the "Congestion" Argument: The paper rigorously demonstrates that in scarce markets, non-personalized quality information (like public rankings) primarily changes sorting (who gets what) rather than aggregate outcomes. This challenges the assumption that better quality signals always improve average welfare in capacity-constrained environments.
2. Quantifying the Value of Fit: It establishes that the marginal value of personalization is driven by the ability to reduce correlated demand for top-tier items, thereby unlocking value through better idiosyncratic fit.
3. Policy Implications: The results suggest that in capacity-constrained markets (e.g., college admissions, housing), resources spent on refining impersonal rankings may have diminishing returns for average welfare. Conversely, investing in tools that help agents discover their idiosyncratic fit (personalized recommendations) can fundamentally raise average match quality.
4. Distributional Effects: The analysis highlights that while information expands total surplus, the distribution of these gains is determined by market tightness (who is on the short side), with sellers capturing gains in tight markets and buyers in loose markets.

The authors note that these findings are derived from a stylized model and abstract from strategic behavior, endogenous entry, and privacy costs, identifying these as directions for future research. However, the core insights regarding the conservation of common value under scarcity and the dominance of idiosyncratic fit in personalized tools are presented as robust across various distributional assumptions and institutional extensions.