Which ads make money? AI-based ad revenue optimization (SK AI SUMMIT presentation)

SK Planet operates a digital advertising platform that displays ads across multiple apps such as OK Cashbag, Syrup, and Orak:, running a real-time bidding structure based on OpenRTB to pursue revenue optimization. When an ad request comes in, bids are sent simultaneously to dozens of Ad Networks, and the highest-priced ad is selected for display—a structure in which traffic, price, and response rate are intertwined in real time.

The core dilemma in ad operations was balancing Minimum Price and Fill Rate. Raising the minimum bid price increases unit price and quality but reduces exposure opportunities, while lowering it increases exposure but can destabilize unit price and quality, requiring the optimal point to be found by examining response rate and price patterns by time of day together. To address this, instead of indiscriminately sending ad requests to all networks, a structure was also designed to selectively request several multiples of the needed quantity, adjusted via Display Weight.

On the operations side, an MLOps pipeline based on AWS cloud was built to extract and refine price and meta information from OpenRTB responses before passing it to the machine learning server. The server performs MAB (Multi-Armed Bandit) while storing auxiliary data in a Data Warehouse, and was designed to revert to initial values or past high-performing parameters if anomaly detection results were poor.

The first attempt on the machine learning side was heuristic-based adjustment. As the number of placement-network combinations grew to the hundreds, manual adjustment hit its limits, leading to the introduction of MAB to adjust bid floors by balancing exploration and exploitation. Performance improved initially, but over time weaknesses emerged: it failed to reflect Context such as time of day, device, and region, and could not sufficiently explain the nonlinear price-revenue relationship.

The next step was applying GMM (Gaussian Mixture Model) to break down the revenue structure into multiple peaks, exploring several revenue clusters simultaneously rather than a single mean. As a result, while MAB tended to concentrate in the 0.5–1.0 range, GMM explored a wider price range of 1.1–2.9 and above, uncovering hidden revenue opportunities.

Ultimately, a GMM blending strategy was introduced, combining Past GMM, which captures long-term historical patterns, with Recent GMM, which reflects trends from the past few hours. This secured structurally stable performance while agilely responding to market changes, and revenue improvement was confirmed without exception across all applied placements. It became clear that ad revenue optimization is not simply about price adjustment, but a problem that must address ecosystem structure, real-time data, and model evolution together.