A job recommender optimized for clicks shows people jobs they will apply for. A job recommender optimized for welfare shows them jobs they should apply for. These are not the same list.
Bied, Caillou, Crépon, Gaillac, Pérennes, and Sebag (arXiv:2603.21699) formalize the gap. A job seeker faces two quantities for each vacancy: how much they'd like the job (utility) and how likely they are to get it (hiring probability). A welfare-optimal recommender ranks vacancies by expected surplus — the product of utility and probability, adjusted for application costs. The optimal ranking integrates both dimensions. Recommenders that optimize for only one — matching to preferences (utility-only) or matching to likelihood (probability-only) — are generically suboptimal.
The problem is deeper than weighting. Optimizing for behavior — what job seekers actually click on, apply to, engage with — does not optimize for welfare. This is the inversion problem: behavior reveals preferences only under specific conditions (complete information, no search costs, rational expectations). When job seekers don't know their chances at a job, they apply based on desire rather than expected surplus. The recommender that best predicts their behavior reinforces their miscalibration rather than correcting it.
Two field experiments with France's public employment service validate the model. The first confirms that both utility and hiring probability influence decisions. The second tests welfare-informed recommendations against existing approaches. The welfare-optimized algorithms substantially outperform — generating outcomes close to the theoretical benchmark.
The through-claim: recommendation systems trained on engagement data optimize for the user's revealed preferences, which are a biased sample of their actual interests filtered through their (incorrect) beliefs about what's feasible. The recommender that knows what you want and what you can get serves you better than the one that knows only what you click on. But the click data is what most systems are trained on, because clicks are observable and welfare is not.