The short summary of the implementation they used seems to be based on a very influential paper from 2002, Optimizing Search Engines with Clickthrough Data.  They do a pairwise transform on clickstream data and feed it into an SVM. Interestingly this model is not widely used in industry, the paper was influential but the exact methods have fallen out of favor for ensembles of decision tree's along with explicit labels rather than pairwise click data. Their reasoning behind this seems reasonable though, their content is highly silo'd and it's quite rare to have queries against the same content be repeated any reasonable number of time that would allow for labeling via click models or humans.