The estimations from an equality where there's no parameter sniffing is based on the average density of the column. The estimations for an inequality where there's no parameter sniffing is, if I recall, 30% of the table. The latter is far too high for seek + key lookup to be efficient, the former may be small enough.