(Yeah, I do have some data to back that up - I wrote & benchmarked an autocomplete implementation for a financial software firm. String tries were roughly 10x slower than binary searching an array for a 20k word corpus, which is probably around the size you're dealing with here. They have terrible cache locality - each character makes you follow a pointer, which is probably a cache miss, and the total size of the trie is on the order of 26 * 8 * 20k = 4M. They become a bit better if the number of distinct words in the corpus is small (so that everything fits into cache) or if the number of words is huge (so that your hashtable or array blows the cache anyway).
I thought that an autocompletion widget for stock tickers would be the perfect application for tries: short, dense key space, lots of elements, and a fair likelihood of hash collisions. But apparently not, because our data size just happened to be one where cache effects dominate. I talked to one of the GMail guys later and he was quite surprised, from which I inferred (but it was not stated) that GMail probably uses tries to good effect.