Producing a good ordering rather than a good score for each item individually. What matters in search and recommendation is which things appear at the top, so the model should be trained on the ordering directly rather than on predicting each item’s relevance in isolation.
Technical view
Pointwise approaches regress relevance per item, pairwise approaches learn preferences between pairs, and listwise approaches optimise a ranking measure over the whole list. Evaluation uses normalised discounted cumulative gain, mean reciprocal rank or precision at k, all of which weight the top of the list heavily. Gradient boosted trees with a ranking objective remain a very strong baseline, and rerankers in retrieval systems are the same idea applied to documents.
Related
Part of AI Terminology Index, section: Classical Machine Learning.