KAIST

<(Bottom from left) M.S candidate Gyurim Hwang, M.S candidate Yeongho Kim, Ph.D. candidate Kyungho Kim, Ph.D.candidate Jongha Lee, M.S candidate Yeonje Choi (Top from left) Undergraduate student Sejin Chung, Researcher Hongseok Lee, Researcher Myeong Ho Song,  Ph.D. candidate Sunwoo Kim, M.S candidate Juyeon Kim, Professor Kijung Shin>

Social media advertising usually requires running multiple ad drafts in practice before determining which ad is effective. Because of this, testing advertisements demands significant time and costs. Furthermore, the criteria for an effective advertisement vary greatly by brand. While some brands prefer person-centered advertisements, others receive better responses from advertisements that emphasize actual usage scenes. However, these effective advertising strategies for each brand are often not clearly defined in the field, which has limited the technology to systematically reflect them and predict advertising performance.

To solve this problem, a research team led by Professor Kijung Shin at KAIST, in collaboration with the AI marketing company MADUP, developed 'ADvisor', an AI technology that predicts advertising performance for each brand.

ADvisor utilizes a generative vision-language model that understands images and text simultaneously to find different advertising success criteria for each brand and predict advertising effectiveness based on them. To achieve this, it not only analyzes the characteristics of the brand but also considers advertising data from other brands with similar tendencies for new brands that do not have sufficient advertising data to derive advertising strategies. Through this process, it can identify distinct advertising success criteria for each brand; for instance, a "strong headline phrase" is analyzed as an important criterion for a specific fashion brand, while "logo exposure" acts as a key element for another brand. Afterward, ADvisor evaluates the advertisement based on the derived criteria for each brand, reviews the evaluation results on its own, and repeatedly compensates for deficiencies to make the final prediction.

The research team verified the technology's performance using data from 10 brands in the beauty, fashion, and platform sectors collected through actual marketing campaigns. As a result, ADvisor recorded up to 7.2% higher performance compared to existing AI advertising prediction models. In particular, in an online A/B test conducted in a real Instagram advertising environment, it achieved an average of 27% better performance in key indicators such as click-through rate (CTR), cost per click (CPC), and return on ad spend (ROAS) than advertisements selected by field marketing experts, proving that it can be utilized in actual marketing decision-making.

Professor Kijung Shin stated, "Predicting advertising performance in advance is the first step for effective advertisement production," adding, "In the future, we will develop our research in a direction where AI directly generates and optimizes advertisements tailored to brand characteristics."

The study, in which Ph.D candidate Kyungho Kim and M.S candidate Yeonje Choi from the KAIST Kim Jaechul Graduate School of AI participated as co-first authors, was published online on April 18 in the Industry Track of ACL 2026, one of the most prestigious international academic conferences in the field of natural language processing. It has been accepted as an oral presentation paper and is scheduled to be presented in the United States this coming July.

※ Paper Title: Pre-Deployment Advertisement Ranking under Data Scarcity via Context-Aware Criteria Generation with VLMs ※ Paper Link: https://openreview.net/forum?id=il84gAzAxx

Meanwhile, this research is an achievement of the project 'EntireDB2AI: Deep Representation Learning and Prediction Source Technology and Software Development Utilizing Entire Relational Databases Comprehensively', supported by the Institute for Information & Communications Technology Planning & Evaluation (IITP).