"Show unstable platforms" + "Sort by number of features" is quite interesting: IE preview wins among browsers. Firefox is close second, and both are ahead of and V8 and even Google's Traceur transpiler.
Writing tests for 'let' support, for instance, is harder than just testing that all the Array.prototype methods are now on TypedArrays, so those methods are worth twice that of let support, even though let is considerably more complicated to implement and game changing in its consequences.
And that's why we give different weight to different features. Smallest ones (Math methods, octal & binary literals, etc.) have a weight of 0.25. Medium ones (Map, Set, rest parameters, etc.) have a weight of 0.5. Finally, landmark features (arrow functions, classes, generators) have a weight of 1.
The final % score is calculated according to all this.