A platform maximizes revenue by allocating traffic to creators with the highest forward-looking marginal value—sometimes temporarily boosting lagging creators to exploit word-of-mouth—while screening entrants with a capability threshold and capping follower growth; naive heuristics can cut revenue by up to a quarter, but a follower-growth–adjusted rule recovers near-optimal returns.

Main Finding

A revenue-maximizing creator-platform should allocate all viewer traffic to whichever creator has the highest forward-looking marginal value (a “most-valuable-creator-first” rule). That marginal value combines immediate ad revenue and the discounted future value of followers created by that exposure. Under simple (linear) follower dynamics this yields a bang–bang policy that switches once marginal values equalize and then maintains an indifference path; under Bass (word-of-mouth) dynamics the same principle can produce counterintuitive “conditional reversals” where the platform temporarily favors lagging creators to exploit accelerating diffusion. The optimal policy also behaves as a selective gatekeeper—imposing a capability threshold for who receives traffic and capping growth for winners—and outperforms naive heuristics (myopic or fairness-driven) by a large margin (numerical losses up to ~25%). A practical “follower-growth adjusted” heuristic that uses observed growth momentum achieves near-optimal performance.

Key Points

• Problem setup
  • Platform has N heterogeneous creators (capability ρi and follower capacity mi).
  • Platform controls continuous-time allocation Ai(t) of total traffic A (ΣAi = A).
  • Two revenue streams per creator: instantaneous ad revenue ρi Ai(t) and per-follower revenue ρi r Si(t).
  • Followers evolve by dSi/dt = Ai(t) · F(Si; ρi, mi) (independent across creators).
• Two follower-growth specifications
  • Linear: FL = p ρi (1 − Si/mi).
  • Bass (with word-of-mouth): FB = (p ρi + q · Si/mi) (1 − Si/mi). Bass produces S-shaped, non-concave growth.
• Marginal value index
  • Key analytical object: Ψi(t) = ρi [1 + (r/δ) · F(Si(t); ρi, mi)].
  • Ψi(t) is the forward-looking marginal benefit of allocating an extra unit of traffic to creator i.
• Structure of optimal policy
  • “Most-valuable-creator-first”: allocate all traffic to the creator with largest Ψi(t).
  • Under linear growth, the state-space partitions by an affine indifference curve S2 = g(S1); motion is bang–bang until the indifference curve is reached, then interior sharing maintains Ψ1 = Ψ2.
  • Under Bass dynamics, switching curves are more complex and can imply conditional reversals (prioritizing a laggard temporarily to exploit accelerating word-of-mouth growth).
• Market-structure consequences
  • Selective gatekeeping: the optimal policy screens creators via a capability threshold (not simply “bet on the largest now”).
  • Disciplined growth: for qualifying creators, the policy caps follower bases at an optimal ceiling to avoid over-investment in diminishing-return regions.
  • Contrast with heuristics:
    • Myopic targeting of stars → extreme winner-take-all concentration.
    • Naive fairness (equal/minimum exposure to small creators) → inefficient, indiscriminate growth.
• Practical heuristics and value of optimality
  • Simple heuristics can lose substantial revenue (numerically up to ~25%).
  • A “follower-growth adjusted” heuristic (using observed follower growth momentum) is managerial and robust, approximating optimal performance even with imperfect parameter knowledge.

Data & Methods

• Analytical methods
  • Continuous-time infinite-horizon optimal control formulation maximizing discounted total revenue Π = ∫0^∞ e^{-δt} Σi ρi [Ai(t) + r Si(t)] dt subject to follower ODEs.
  • Derivation of closed-form structure for the two-creator case (explicit indifference curve under linear growth; state-dependent switching curves under Bass).
  • Use of the marginal value index Ψi(t) to characterize bang–bang vs interior allocation regimes.
• Numerical experiments
  • Extensive simulation across parameter regimes (ρi, mi, p, q, r, δ, A, N) to:
    • Illustrate state trajectories and conditional reversal phenomena.
    • Compare optimal policy revenue to myopic, fairness, and follower-growth adjusted heuristics.
    • Quantify revenue losses from heuristics and robustness of the proposed practical heuristic.
• Key model parameters
  • p: exogenous conversion rate.
  • q: word-of-mouth strength (Bass only).
  • mi: maximum follower capacity per creator.
  • ρi: capability (converts traffic → revenue and follower growth).
  • r: per-follower contribution (subscription/tip rate).
  • δ: discount rate.
  • A: total available viewer traffic.

Implications for AI Economics

• Recommendation-system design
  • Algorithms should be forward-looking (estimate Ψi) rather than myopically optimizing immediate clicks/revenue. Incorporating the future value of follower growth changes allocation priorities and long-run revenue.
  • Practical implementable rule: prioritize creators with the highest Ψi(t) ≈ ρi [1 + (r/δ) · estimated short-run conversion], or use the follower-growth adjusted heuristic when parameters are uncertain.
• Market structure and platform policy
  • Platforms are not neutral; traffic-allocation policies actively shape creator ecosystems (selection thresholds, controlled scaling of winners).
  • Platform revenue maximization can justify selective promotion (gatekeeping) rather than indiscriminate “fair” exposure—this has consequences for creator diversity and welfare.
• Welfare, regulation, and competition
  • Observed platform choices that favor stars or that intentionally surface niche creators can be rationalized by forward-looking revenue trade-offs; regulators and researchers should consider dynamic effects (follower externalities) when assessing market power or fairness interventions.
  • Trade-off between platform revenue and broader social goals (diversity, creator welfare) is explicit: policies that maximize platform revenue may not maximize creator-side welfare or cultural diversity.
• Managerial recommendations
  • Estimate creator-specific capabilities (ρi) and short-run/word-of-mouth growth parameters (p, q) from platform data; compute Ψi(t) and allocate traffic to the current maximizer.
  • Monitor follower momentum to detect when temporary priority shifts (conditional reversals) are warranted.
  • Implement admission/screening rules (capability thresholds) and controlled capping of follower promotion to avoid inefficient over-investment.
• Directions for AI economics research
  • Extend to stochastic learning of unknown creator capabilities (exploration-exploitation), multi-platform competition, advertiser responses, and welfare analyses comparing platform-optimal vs socially optimal allocations.
  • Empirical calibration and A/B tests: measure p, q, r, and validate the follower-growth adjusted heuristic in production recommender systems.

Limitations to keep in mind: the model assumes deterministic ODE follower dynamics and independent creator growth, known parameters, and an infinite horizon—real platforms face stochasticity, platform competition, strategic creators, and multi-sided contractual constraints, which are natural extensions.