I think the issue is that the percentages may be measured in lines of code.

If one language is more verbose, counting lines of code will throw off conclusions about how much functionality is implemented using each language.