Spotify knows what I played. Netflix knows what I watched. Neither necessarily knows why I chose it—or what I want tonight.
That is why recommendation is hard: behavior is evidence, not intent. A click can mean curiosity, habit, compromise, background noise, or regret. Then systems train on their own suggestions, turning yesterday’s guesses into tomorrow’s profile.
AI can help, but the interface cannot become an interview. Every question spends effort and raises the expectation that the result will improve. If it still misses, the system may lose more trust than if it had made a low-confidence suggestion first. The better shift is selective conversation: ask only when one answer will materially change the slate; otherwise show a diverse starting point, expose the assumptions, and make correction effortless.
Good recommendation may be less prediction and more shared discovery.