Coverage
A recommender can post excellent accuracy numbers while only ever recommending the same few thousand items. Coverage is the metric that notices — it asks what share of the catalogue a system is actually willing to put in front of anyone.
Two measurements under one name
Catalogue coverage is the fraction of distinct items that appear in at least one user's top-N list across an evaluation run: 4,000 distinct titles recommended from a 50,000-title catalogue is 8%. User coverage is the complement — the fraction of users the system can produce a recommendation for at all, which is where cold start becomes a number. Gini index or entropy over the recommendation-frequency distribution sharpens the first one, asking how evenly exposure is spread, not merely how many items received any.
Why accurate models score badly
A model trained on logged implicit feedback learns that popular items are the safe prediction, so it concentrates exposure on the head of the distribution. That maximises Precision@k while leaving the long tail unseen, and the damage compounds: items that are never shown collect no interactions, so the next training round has even less reason to retrieve them. Coverage is how that feedback loop is caught offline, before it surfaces months later as a stagnant catalogue.
Using it in practice
Coverage is read alongside accuracy, never instead of it — a random recommender scores 100% and is useless. It belongs in the same report as NDCG for every candidate model, and it is the number that justifies re-ranking for diversity or adding exploration. On MovieLens, implicit ALS and BPR differ far more in coverage than in hit rate.
Coverage measures how much of the catalogue a recommender actually recommends — the fraction of items appearing in at least one user's top-N list, and, in its second sense, the fraction of users who can be served at all. It exposes the failure accuracy metrics hide: a model that concentrates exposure on popular items scores well while leaving the long tail invisible. Coverage is read alongside Precision@k, never instead of it.
Think of a bookshop with fifty thousand titles that, week after week, only ever puts the same three hundred on the display table. Coverage is the simple check that catches this: out of everything the service could show people, how much does it ever actually show?
It matters because a recommender can look very good and still be very narrow. Suggesting the same proven hits to everyone produces plenty of clicks, so the usual scorecards are happy — but most of the library is never offered to anybody, the people who made those items get no audience, and anyone with unusual taste is poorly served. Coverage puts a number on that blind spot.