Recommendation algorithms deliberately skew feeds to cultivate monetizable content: platforms subsidize exposure for talented creators and screen out low-ability ones to trigger effort-based network effects, and shifts in ad-targeting profitability or privacy rules alter the balance between algorithmic amplification and direct creator payments.

Main Finding

The platform’s revenue-maximizing recommendation algorithm optimally reallocates scarce user attention between ads (monetization) and organic content (to incentivize creators). To leverage positive effort–attention network effects, the platform deliberately recommends (subsidizes) some content that is ex post suboptimal for the viewer. These attention subsidies are targeted: low-ability creators are excluded, medium-ability creators receive within-category exposure subsidies, and high-ability creators receive additional cross‑category reach or direct monetary payments when efficient. Private information (unobserved creator ability and viewer valuation) amplifies these distortions, reduces participation and effort, and narrows content variety. Improvements in ad monetization shift the platform away from cross‑category exposure subsidies toward monetary payments and more accurate personalization. Regulations that limit profiling/personalization (e.g., DMA/DSA) can therefore reduce content diversity and creator participation.

Key Points

• Attention is a scarce, monetizable resource with two roles: (i) generating ad revenue and (ii) providing exposure that incentivizes creators to exert costly effort and raise content quality (network externality).
• Creators: vertical ability θ ∈ [0, θ̄], choose effort e at cost e; quality q = θ e; payoff = u(X) − e where X is attention received.
• Viewers: horizontal category j and valuation v for quality in that category; utility from consuming one unit of content of quality q in preferred category = v q − c (attention cost c); ads yield −c and platform earns z per unit of ad attention.
• Platform problem: mechanism design / screening under private vertical types (θ, v). Platform observes categories and realized content quality q, but not θ or v. Platform commits to an attention allocation (recommendation) rule to maximize ad revenue subject to incentive and participation constraints.
• Attention subsidies: to induce effort and strengthen network effects the platform recommends some content even when v q < c for the individual viewer — i.e., viewers are allocated content they would not ex post prefer — because doing so raises creators’ expected reach and therefore future quality for viewers who value the category highly.
• Targeting of distortions:
  • Low-ability creators are excluded (too costly to incentivize).
  • Medium-ability creators get exposure within their category.
  • High-ability creators may get additional cross‑category reach (algorithmic amplification).
• Private information tightens screening: platform more narrowly targets high-value viewers and high-ability creators, raising entry/participation thresholds and reducing effort and engagement relative to complete-information benchmark.
• Implementation: a simple nested/ordered feed (infinite-scroll + stopping rule) can implement the optimal allocation: lower-ranked items get less attention and lower match quality for given viewer — a testable prediction.
• Extensions:
  • Allowing direct monetary transfers to creators: for each category there is typically an ability cutoff; below it creators are incentivized via exposure, above it marginal incentives are provided through monetary payments while exposure is held fixed.
  • Comparative statics: higher ad revenue per attention (better ad targeting, market power) → platform substitutes monetary payments for exposure-based incentives and improves recommendation accuracy.
• Policy trade-offs:
  • Privacy/personalization restrictions (DMA/DSA) limit ability to target and exploit complementarities → fewer creators included, especially in niche categories; less content variety; some viewers lose access to content they value.
  • Rules reducing ad-targeting profitability can increase the platform’s reliance on attention-based subsidies (more algorithmic amplification).

Data & Methods

• This is a theoretical/mechanism-design paper (no empirical dataset).
• Model architecture:
  • Two-sided continuum model with N horizontal categories.
  • Heterogeneous creators (mass αi per category) with ability θ ~ Fi(·) (strictly increasing hazard rate).
  • Heterogeneous viewers (mass βj per category) with valuation v ~ Gj(·) (strictly increasing hazard rate).
  • Key primitives: q = θ e; creator utility u(X) (increasing) minus effort cost e; viewer payoff v q 1{i=j} − c per attention unit; ad revenue z per attention unit.
  • Information: platform observes categories and realized q, but not θ or v.
• Analytical approach:
  • Characterize profit-maximizing recommendation mechanism under complete and incomplete information.
  • Use screening and incentive-compatibility constraints to derive optimal attention allocation rules; characterize which creators are included and how attention is dispersed across and within categories.
  • Show existence of a simple, natural implementation via a nested ordered feed where viewers consume sequentially and choose a stopping point.
  • Extend the mechanism to allow monetary transfers to creators and analyze comparative statics with respect to z (ad revenue per attention) and information constraints (personalization capability).
• Key assumptions to note:
  • Viewers only get value from their preferred category (payoff from other categories normalized to zero).
  • Attention cost c constant across uses and viewers (heterogeneity captured via v).
  • Platform observes realized quality q but not the decomposition (θ,e).
  • Creators value total attention (abstracting from audience composition), though the paper discusses extensions.

Implications for AI Economics

• Algorithmic allocation is an active instrument: Recommendation systems do more than match supply and demand — they are tools for incentive provision on the supply side (creators). Economic evaluation of recommender systems must account for these second-order production responses.
• Trade-off between monetization and production incentives: Platforms internalize a tension between immediate ad revenue and longer-run supply quality; algorithmic distortions (subsidized exposure) are optimal when exposure efficiently raises creator effort.
• Personalization and privacy regulation are economic levers: Limits on profiling reduce the platform’s ability to finely target attention, weakening network effects and shrinking content diversity and scale. Regulators should weigh privacy/autonomy benefits against potential losses in content variety and creator income pathways.
• Advertising-market structure shapes platform incentive design: Improvements in ad-targeting or increases in ad value make attention more valuable to monetize, encouraging platforms to rely more on direct payments (monetary transfers) to creators and to produce more accurate recommendations.
• Empirical predictions and tests:
  • Lower-ranked (lower-exposure) content should show lower viewer match quality conditional on viewer type.
  • Platforms that experience exogenous improvements in ad monetization (or increases in z) should increase direct creator payments and improve recommendation accuracy; conversely, privacy regulation that reduces targeting should correlate with greater algorithmic amplification and reduced niche content reach.
  • Niche categories and smaller audiences are disproportionately harmed by restrictions on personalization: fewer creators survive and cross-category reach shrinks.
• Broader research directions:
  • Empirically quantify the trade-off between ad revenue and supply-side incentives using platform-level shocks (e.g., policy changes, ad-monetization product introductions).
  • Extend models to heterogeneous audience composition value (creators value audience type) and rivalrous attention across overlapping categories.
  • Study welfare implications and optimal regulation when accounting for consumer privacy, creator surplus, advertiser surplus, and platform profits.