The web is unique as a "hypothesis developing" aid in translation, but less than ideal for (in)validating hypotheses.

Many translators will be familiar with this sense of uncertainty: how can we be sure that these short extracts we are reading on screen:

• come from reliable texts;
• have been written by linguistically competent (native-speaking?) authors;
• actually deal with the subject of wine: maybe the word wine appears a long way from room/ambient temperature, and the two expressions are not in fact part of the same text...
  

We are not suggesting that corpora are free from these problems (and indeed one could argue that a certain amount of uncertainty is part and parcel of translation). But the fact that in a corpus we have selected texts prior to searching for their contents should make us feel more confident about our findings. We know that the texts in our corpus are reliable, comparable, well-written etc, because we have evaluated them before deciding to include them in our corpus.