The tech industry's flashy recommendation systems have created unrealistic expectations. TikTok, Spotify, and Netflix dominate headlines with their sophisticated deep learning models, but most practitioners actually work with simpler gradient-boosted tree approaches. Until recently, I thought my experience with basic tabular models was unusual—now I believe this represents the norm rather than the exception.

What separates these giants from the rest? The framework reveals two critical dimensions: observable outcomes versus catalog stability, and preference subjectivity. IKEA exemplifies the first dimension perfectly—when customers buy one sofa over another, the signal is clear and unambiguous. This creates a strong baseline that's difficult to improve upon. In contrast, platforms like Tinder and Yelp struggle with upper-funnel signals where position bias clouds true preference data. When users click a restaurant simply because it appeared first, you lose the anchor of reliable quality measurement.

Catalog churn presents another challenge. Real estate platforms like Zillow and secondhand marketplaces like Vinted face extreme inventory turnover where items disappear upon purchase. This forces reliance on simplistic sorting methods like "newest first" rather than data-driven leaderboards. Despite these challenges, gradient-boosted trees remain the pragmatic choice across most recommendation systems. They efficiently predict engagement probabilities using engineered features like location, time, and device type—proving that sophisticated personalization doesn't always require complex deep learning architectures.