First-hand information for everyone
Main finding: When a platform offers both ad‑supported and ad‑free plans, advertising revenue can pay for useful service for consumers who would receive zero service under a pure‑subscription benchmark. The authors call this effect “ad‑funded quality democratization.” (This can happen even after accounting for production costs and lost premium payments.)
How the authors studied it: They model a monopolistic large‑language‑model (LLM) platform that publicly commits to a menu of plans. Each plan specifies response quality q, advertising intensity α (how much advertising users see), and a subscription price T. Consumers have a one‑dimensional private type θ and choose a plan or an outside option. The platform maximizes expected profit (subscription payments plus advertising revenue minus quality production cost), subject to incentive‑compatibility (IC — users choose the menu item intended for their type) and individual rationality. The analysis reduces to pointwise maximization of a “virtual surplus” and a derived “virtual advertising profit.” (Short form of methods and key equations used for explicit results are given below.)
Concrete results (selected and explicit): - Pure‑subscription benchmark. With the baseline specification, the pure‑subscription quality is q_sub(θ)=max{0,2θ−1}. Thus types θ ≤ 1/2 receive zero quality under pure subscription. (This is the benchmark that advertising can improve on.)
- Advertising can fund positive quality for low types. With advertising, the pointwise optimal choice for low types satisfies q*(θ)=max{0,(2θ−1)(1−δ)+λ} when the platform sets maximal ad load α=1 for those types. A positive quality‑driven advertising premium λ can make q* positive for some types with negative virtual value (that is, some who would get zero under pure subscription). This is the formal statement of ad‑funded quality democratization.
- Market segmentation and thresholds. Under the baseline functional forms the optimal menu can split the market into up to four segments. These segments are: (i) a zero‑quality ad‑supported baseline; (ii) positive quality with full advertising (low types); (iii) partial advertising (middle types); and (iv) ad‑free service (high types). They arise because advertising affects both the revenue a plan generates and the effective quality (and thus the information rents) that keep higher‑valuation users from switching down. The analysis gives closed‑form critical thresholds. With the baseline forms the positive‑quality threshold is θ_pos = 1/2 − λ/[2(1−δ)]. The platform begins reducing ad intensity above θ1 = 1/2 + λ/(2δ). The ad‑free threshold θ2 solves (2θ2−1)[(2θ2−1)δ − λ] = r̄ and can be written θ2 = [2δ + λ + sqrt(λ^2 + 4δ r̄)]/(4δ).
- Utility polarization and pricing effects. Introducing partly ad‑supported and ad‑free plans weakens effective quality in middle plans. That reduces alternatives available to premium users. The platform can then raise the price of unchanged ad‑free service. Some fully ad‑supported users can gain utility. Partly ad‑supported and ad‑free users can lose utility. Thus broader access via advertising can coexist with “utility polarization.” An intermediate cap on advertising exposure can maximize total consumer surplus by preserving revenue that finances service while limiting advertising‑induced degradation.
- Effects of cheaper compute and high ad returns. Lower compute costs make higher quality cheaper to supply. The platform may both raise supplied quality and reduce ad intensity. But if quality rises enough, the remaining advertised interactions become more valuable and aggregate advertising revenue can increase. With sufficiently large advertising returns to quality, advertising‑funded upgrades can push all plans to the same technological ceiling and full advertising exposure. When that happens, tiers become identical and cannot sustain different prices; they merge.
Key baseline functional forms and parameters (what the explicit results use): - Consumer types are uniform: θ ~ U[0,1]. - Compute cost: c(q)=β q^2 /2. Baseline β=1, giving a quality ceiling q̄ = 1/β = 1. - Effective (experienced) quality: g(q,α)=q(1−δα) where δ∈(0,1) is the proportional quality loss per unit advertising intensity. This reduced form captures channels such as interference or privacy frictions without modeling them separately. - Advertising revenue: R(q,α)=λ q α + r̄(α − α^2/2). The parameter λ≥0 is the quality‑driven advertising premium; r̄ captures quality‑independent baseline ad returns.
Important caveats and limitations the authors state: - The quality‑linked advertising premium λ is treated as a modeling primitive. The cited platform practices motivate it, but do not directly establish a causal effect that raising quality raises advertising returns. - The analysis maintains simplifying assumptions that limit generality: one‑dimensional private information (single type θ), full commitment to posted menus, and a monopolistic platform. The authors discuss these maintained assumptions and extensions in appendices. - Results rely on parameter regularity to ensure an endogenous ad‑free premium tier. The paper imposes δ∈(0,1), λ∈(0,δ), and r̄∈(0,δ−λ). Other parameter regions can change allocations (for example, when λ + r̄ > δ even the highest type may receive advertising). - The functional forms for g and R are reduced‑form for tractability. They are approximations that summarize multiple operational channels rather than modeling each channel separately.
What this means in practice: The model shows that advertising can extend useful LLM service to users who would otherwise get none under subscription alone. But it also shows trade‑offs. Advertising can weaken mid‑tier service and change pricing, producing winners and losers. Limits on advertising exposure or changes in compute cost interact with these trade‑offs in non‑trivial ways. The quantitative boundaries of these effects depend on the model’s parameters and primitives, which the authors highlight as important for interpretation and further empirical work.