Where do negative examples come from when the user never rejected anything explicitly?

Three sources, from the most honest to the cheapest. Impressed but not clicked is a genuine negative: the item was on screen and the user skipped it. Random catalogue items are cheap noise — fine for training candidate generation, far too easy for a ranker. Hard negatives are popular or near-positive items the model confuses; these drive the actual gains. The classic mistake is training the ranker on random negatives: it learns to separate relevant from arbitrary, whereas production asks it to separate relevant from nearly relevant.