In this case, the A/B testing is mostly in the form of machine learning in their learn-to-rank scoring algorithm. The machine learning gets the final say as to

In this case, the A/B testing is mostly in the form of machine learning in their learn-to-rank scoring algorithm. The machine learning gets the final say as to if those quotes are important or not. I'm not sure what they do to prevent learn-to-rank from over-optimizing the median case, but it seems to not be enough for the tastes of many power users.