Recommender systems are usually judged on accuracy: did the user click, watch, or buy? But a review of research on evaluation metrics argues that accuracy alone tells an incomplete story. Serendipity—the ability to surface items the user did not know they wanted—is often neglected, even though it can drive long-term engagement.
The review draws on more than 10 papers to examine how diversity, novelty, surprise, and serendipity are measured. These qualities are related but not interchangeable: each captures a different way recommendations can move beyond the obvious. They are frequently treated as secondary to accuracy, despite their potential to improve user experience.
The takeaway is not that accuracy is unimportant, but that evaluation needs to account for experience. If a recommender only optimises for closeness to past behaviour, it may keep users comfortable but bored. Serendipity and similar measures offer a counterweight—and a path toward systems that genuinely delight.