How does recommendation differ from search and from plain classification?

Classification has a ground-truth label per item; search has an explicit query. Recommendation has neither: the query is implicit (the user plus context) and there is no single correct answer — only an ordering that satisfies the user better or worse. Three consequences: the output is a ranked list of K items, so metrics are position-aware; training data is generated by the previous model and is therefore biased (you only observe feedback on what you showed); and true quality is measured online, with offline evaluation as an approximation.