I've pondered this, too. Why not a DMOZ 2.0 type concept, where you seed the index with high quality sites and grow from there. Freshness won't matter as much for most topics, so crawling X times per day probably won't be necessary. Maybe have user defined flags to indicate the types of results they want (large sites, small sites, high authority, technical sites-- maybe different sort by features) with backlinks/domain prominence as a small factor, but also using NLP to determine authoritativeness of said content so as to facilitate less-linked site discovery.