Interviews in waves of five to eight. Enough to see a pattern, few enough that you can change something between waves.
After each wave, revise the artefact and test again. Research that never changes anything is theatre.
In practice: Makerz Network
One of the standing research inputs is the product's own search logs — what people typed into the search box, and which of those searches came back with nothing.
It is the cheapest research available: nobody has to agree to an interview, and it records what someone actually wanted at the moment they wanted it rather than what they recall wanting a fortnight later. What it cannot tell you is why. That is what the interviews are for.