Work out recommendation coverage instantly with clear inputs, formula shown and shareable results.
Catalogue coverage is the share of items a recommender ever surfaces. Popularity-biased models score well on accuracy while recommending a tiny head of the catalogue, which starves the long tail of the impressions it needs to accumulate interaction data — a feedback loop that makes coverage fall further over time. Coverage is therefore a health metric that accuracy alone cannot reveal.
Coverage
coverage = distinct items recommended / catalogue size; impressions per covered item = slots served / distinct items
It depends on the business, but below roughly 20 percent most of the catalogue is commercially dead. Track the trend rather than an absolute target.
Add exploration such as epsilon-greedy or Thompson sampling to the ranker, apply popularity debiasing in training, and reserve slots for cold-start and long-tail items.