I trust their statistical significance calculations... but at a glance, the distribution barely looks different from what I'd expect if the turkers picked at random (either on purpose or because search tastes are at some point arbitrary).

It'd be interesting to include such a graph, where all ratings are drawn at random (but with the same slightly-vs.-much proportions) for visual comparison.

Also: what would happen if all the individual queries on which the preference isn't statistically significant were discarded, or repeated until the preference becomes significant? (Or is "6-8 workers" enough for significance?)