Optimal Auction Mechanisms, Reserve Prices, Dynamic Pricing, and Budget Pacing in Online Advertising

Online advertising has become a dominant monetization model, driving the rise of major ad platforms like Google, Facebook, ByteDance, Alibaba, Baidu, and Tencent, and making ad revenue a key component of overall earnings.

The optimal auction mechanism is built on a virtual valuation function Y(v) that is monotonic under regularity assumptions; the item is allocated to the highest virtual valuation and the winner pays the second‑highest virtual valuation, which corresponds to a second‑price auction. The virtual valuation also represents the seller’s marginal revenue, linking economics and mechanism design.

A market reserve price (r) is the seller’s minimum acceptable price, equal to the seller’s cost. If the highest virtual valuation is below r, the item is retained and the auction does not occur, affecting social welfare and efficiency.

Dynamic reserve prices adapt r for each advertiser based on their estimated valuation distribution. By setting Y = r and solving v - (1‑F(v))/f(v) = r, the platform can compute a personalized reserve price that maximizes expected revenue, provided the valuation density f(v) is known.

Price‑compression factors modify the eCPM formula to eCPM = 1000 × pCTR × bid × p, where p (0 < p < 1) scales the influence of the bid versus the predicted click‑through rate. Increasing p gives more weight to price, decreasing p emphasizes click probability, allowing platforms to steer auction dynamics.

LinkedIn’s budget‑pacing method splits a day into K time slots, defines a total budget B, and compares cumulative actual spend S(t) with planned spend A(t). The participation rate P_t is adjusted multiplicatively: if spend is too high, P_t = P_{t‑1} × (1‑R); if spend is too low, P_t = P_{t‑1} × (1 + R), where R is a small adjustment factor (e.g., 0.1). This simple feedback loop smooths spend without considering performance metrics.

Yahoo’s approach adds a quality‑layer dimension. Requests are divided into L layers based on predicted performance (e.g., click‑through rate). Each layer l has its own participation rate r_{l,t}. The platform adjusts rates layer‑by‑layer: if cumulative spend exceeds the plan, rates are lowered starting from the highest‑quality layer; if spend lags, rates are raised starting from the lowest‑quality layer. Formulas compute the required adjustment to keep the spend curve aligned with the budget plan.

When a cost‑per‑click (CPC) goal is imposed, Yahoo introduces an effective CPC metric ecpc(i) for each layer i, calculated as the total cost divided by total clicks for layers i…L. The algorithm iteratively adjusts layer participation rates so that ecpc(i) ≤ goal, using proportional scaling based on the difference between actual and target spend.

Practical settings include choosing a time‑slot length of 1–3 minutes for smoother control, initializing participation rates low (e.g., 0.1) to allow learning, and setting the adjustment factor R around 0.1. Accurate traffic forecasts are essential for allocating budget across slots and for cold‑start scenarios.

In summary, LinkedIn’s method is simple and effective for pure spend control, while Yahoo’s layered approach balances spend and performance but requires more complex modeling. The choice between them depends on platform priorities, data availability, and the need to respect performance constraints.