Potential Generalized Second Price (PGSP) Auction for Augmented Advertising

This paper introduces an augmented advertising model that bundles multiple ads into a two‑step landing page to increase ad inventory in information‑flow products. The system consists of a "guide ad" shown on the main feed and a "second‑step ad" displayed on a dedicated landing page after the user clicks the guide ad.

The authors identify three key challenges: (1) the asynchronous retrieval of second‑step ads makes it hard to optimize guide‑ad ranking at scale; (2) pay‑per‑click (PPC) billing creates a "free‑riding" problem where guide‑ad owners pay for clicks that do not lead to further engagement; (3) existing GSP and VCG auctions ignore the potential value of the second‑step page, preventing welfare‑optimal outcomes.

To address these issues, a two‑stage auction framework is proposed. The first stage is a "guide‑ad" auction using a new Potential Generalized Second Price (PGSP) mechanism. PGSP orders guide ads by their expected social welfare, which incorporates the estimated revenue from the associated second‑step page, and shifts the billing point to the second click on the guide ad, eliminating free‑riding. The authors prove that PGSP admits a non‑empty symmetric Nash equilibrium.

In the second stage, the "second‑step ad" auction follows the standard GSP rule, as its design remains unchanged under the given guide‑ad ordering.

Estimating the virtual bid for each guide ad (i.e., the expected second‑step revenue) is cast as a regression problem. Features include user identifiers, ad attributes (brand, price, sales), and predicted click‑through rates (pCTR, pCVR). A lightweight multi‑layer neural network is trained, with log‑eCPM of second‑step ads as the target label.

Extensive offline experiments on Taobao’s homepage feed data show that PGSP improves average click‑through (CPS), revenue per stream (RPS), and reduces average CPC compared to GSP and VCG. Online A/B tests over three weeks confirm higher page views, clicks, GMV, and overall ad revenue while maintaining lower CPC.

The study concludes that the augmented advertising format combined with the PGSP two‑stage auction yields superior user experience and monetary gains, validated both theoretically and empirically.