• Simplest Learning to Rank family.
• Treat each (query, item) independently. Predict an absolute label: relevance score, or click probability. Plain regression / classification.
• Loss is per item (e.g. binary cross-entropy on "clicked or not"). Then sort by predicted score.

Weakness

• The loss knows nothing about ordering. It tries as hard to get item #500's score right as item #1's, but you only care about the top few. It optimizes calibration of every score, not order of the top.

Why still common

• Simple, scales, and the predicted probabilities are useful elsewhere (e.g. click-through-rate for ad auctions, expected revenue). Most CTR prediction is pointwise.

References