Advertising Targeting: From Glory to Sunset – Technical Reflections and Future Directions

The author presents a personal and technical summary of several years of work on advertising targeting, describing it as a once‑dominant vertical in the ad‑tech ecosystem that is now approaching its sunset.

Historically, the shift from portal‑slot selling to flow‑distribution made targeting the core engine of ad delivery, likened to selling specific cuts of meat rather than whole carcasses, and yielded over 30% lift in conversion for many advertisers, especially mid‑tier ones.

From a product perspective, the targeting stack evolved through several stages: initial single‑machine Python binary classifiers for intent tags, simple CTR models with handcrafted features, then a migration to XGBoost with a custom xgboost4j‑spark bridge to enable large‑scale parallel prediction.

Subsequently, the team transitioned to deep neural networks, adopting multi‑task learning, wide‑&‑deep architectures, and attention mechanisms to weight user tags more intelligently, while preserving a parent‑child tag hierarchy that allows both precise and high‑volume targeting.

Cost‑aware optimization was incorporated by normalizing diverse conversion costs and applying loss weighting, effectively balancing ROI and conversion volume across different ad categories.

Beyond traditional targeting, the article explores user‑recall techniques using DMP seed users, tag‑based similarity, and DSSM‑style models to expand audience pools, emphasizing the distinction between ad‑centric and user‑centric recall sampling.

Finally, the author reflects on the diminishing role of explicit targeting as full‑recall and intelligent ranking improve, suggesting that while the explicit targeting paradigm may fade, the underlying modeling insights, tag embeddings, and system engineering remain valuable for future ad‑tech challenges.