One thing they could do to help alleviate this is favor the Unicode section(s) associated with the visitor's language, as well as punctuation. That way we'd be less likely to get results like "Ogham letter nion: ᚅ" for drawing pi.

I tried the ae ligature, long s, sharp s, Russian zha, Greek capital gamma, and a heart, and of those only the gamma was in the top five (it was #2); zha, ae, and heart weren't even in the list, and long-s and sharp-s were waaay down the list. (Actually, I just tried it with a better-drawn sharp-s and it didn't even turn up.) In some cases the top choices really do resemble the letter I was looking for, so that's fine, but there's a lot that shows up that resembles the canonical form a lot less than my drawing does.

So, not quite ready for prime time, I think. Cool idea, though.

Pound (£) sign was fine for me, even though I actually drew a lira sign (₤ - is has two crossbars). Hash (#) was ridiculous, though:

Tifinagh letter tuareg yazh: ⵌ

Vai syllable pu: ꖛ

Equal and parallel to: ⋕

Viewdata square: ⌗

Another datapoint: I had no trouble with any of those.

Pi took me a couple of goes, and when I drew a somewhat narrow ampersand, it gave me a few erroneous results.

Both were successful, and put the desired character at the top of the list. All the others worked first time.

It would be excellent if it could be made to learn, with users being able to say "that's what I meant"