It wasn't immediately clear what relevant meant, but I don't doubt they have thought deeply about that My real comment though is that to some extent, the more

It wasn't immediately clear what relevant meant, but I don't doubt they have thought deeply about that My real comment though is that to some extent, the more accurate the results are in terms of "relevance", the worse the outliers, and the worse the user understanding of how the query went wrong. For better or worse, the nice thing about keyword matching and boolean queries is you understand exactly why something did or didn't match. If your results are coming from a neural network optimized for relevance, it's much tougher to form a mental model of how the search responds and to adjust it. (This is not a new idea, nor is it specific to google, but I think it does offer an explanation of how google has reduced irrelevant results even as there are more and more complaints about search getting worse. Nobody will complain a keyword search doesn't do what they wanted...)