At the start, it would suck, but if the search engine reached 98% accuracy rate, you don't need more results. The engine could even learn when results are in order - for example, asking "where is mongolia" will give you a single answer, but "tv review sites" will give you a list of search results.
So bootstrapping is a challenge. Will users wait out the 'training' period?
(Or similarly, if a forward-thinking group is willing to help train, are their choices representative enough of what the impatient masses eventually want?)
I think it's a promising area for experimentation.
I am going to model this mathematically soon, but with a very small sample size, you can quickly reach a pretty accurate result. And in general, most results are pretty clear - number of states in the u.s will always be the same, no matter how forward thinking you are.
Such an engine has to start small and be trained for a few months, otherwise users will think it sucks. As it grows, it adapts.
I bet one could actually do this very easily using yahoo BOSS. Want to take a stab at it? If anyone wanted to work on this, I'd help, so long it was written in python.
