NDCG (Normalized Discounted Cumulative Gain) measures ranking quality: it sums the relevance of the results but discounts each by how far down the list it sits, then normalizes against the ideal ordering — so 1.0 means a perfect ranking.
Worked example: two systems can return the same relevant documents, but the one that ranks the best result at position 1 instead of position 8 scores higher, because the discount penalizes burying good results. Gotcha: NDCG needs graded relevance labels (how relevant, not just yes/no) and a cutoff (NDCG@10), and it says nothing about results below the cutoff — so pick the cutoff to match what users actually see.