The metric which would be a tad closer to ideal would be some average of u/t where u := the number of users who upvoted a given comment, t := total number of users who saw that comment. Where "saw" can mean either "the comment was already there when the user loaded the comments page", or, if you want to get fancy, "the comment was within the browser's viewport long enough to be read" (which some clever JavaScript can tell you.)