Bayesian model reveals why users stop scrolling in search results

The paper models a user facing a ranked list of search results, where each item's relevance is unknown but correlated. The user pays a fixed cost per inspection and decides sequentially whether to continue or stop with the best found so far. The optimal policy is a standout rule: stop when the current best find surpasses the posterior mean of an average item by a threshold that depends on how many items have been inspected. This leads to a collapse of the dynamics into a one-dimensional Markov chain, yielding a closed-form recursion for the full distribution of inspection depth.

The model uncovers three mechanisms—trust, commit, and cut-losses—that explain why users stop early. It also delivers a novel learning-to-rank likelihood: observed depth imposes survival inequalities on latent relevance scores, and the Gaussian probability of those inequalities is differentiable with respect to any feature-based relevance prediction. This opens the door for end-to-end training of ranking models that directly account for user stopping behavior.