Modeling and Optimizing Real‑Time Bidding for Xiaohongshu "Fries" Advertising

Xiaohongshu’s "Fries" advertising product allows creators and merchants to promote notes via a simple mobile interface. Although the user experience is straightforward, the platform faces significant challenges in allocating commercial traffic among competing ad formats while maximizing overall revenue for both advertisers and the platform.

From a strategic perspective, the commercial technology team modeled the Fries bidding process, derived a theoretically optimal solution, and combined it with practical considerations to design a simple yet effective control scheme that quickly achieved the desired business performance.

The business flow involves users selecting target exposure, duration, audience, and budget, after which the system automatically runs the campaign. Fries ads compete with other ads in the real‑time bidding (RTB) pipeline, encompassing the three classic stages of recall, estimation, and auction.

Key client‑side requirements include smooth pacing, predictable conversion volume, and guaranteed delivery. Platform‑side concerns focus on handling under‑delivery and avoiding loss‑making bids, since each Fries order consumes a portion of the platform’s ad inventory that could otherwise generate revenue.

By formalizing the problem, the team defined variables such as target delivery D, order revenue B, CPM cost M = B/D, and conversion cost constraint C. The optimal bidding formula was derived from a linear programming model and its dual, yielding a per‑impression bid that satisfies multiple constraints (exposure, profit margin, conversion cost).

Because real‑time traffic is volatile, the solution incorporates online feedback: the bid parameters are continuously adjusted based on observed performance. The three‑parameter control problem (exposure, profit, conversion) was reduced to a two‑parameter (slope a, intercept b) linear function that partitions the traffic space. By fixing the slope to a historically optimal value and adjusting the intercept with a simple PID controller, the system can closely approximate the optimal solution while remaining easy to implement.

Practical heuristics include capping bids at a multiple of the CPM cost M to avoid excessive spend, and filtering out low‑value traffic (v < M/C) to simplify the auction. These rules, together with the PID‑based pacing strategy, address client‑side smoothness, delivery guarantees, and platform profit objectives.

The approach balances rapid deployment with solid business impact, while acknowledging that more sophisticated methods such as reinforcement learning or model predictive control could further improve performance in the future.

At the end of the document, the Xiaohongshu Search Advertising Algorithm team posted a recruitment notice for algorithm engineers, outlining required qualifications and application contacts.