First two points are great, but why then we see this: "Given the ratings I have, there is a 95% chance that the "real" fraction of positive ratings is at least

First two points are great, but why then we see this: "Given the ratings I have, there is a 95% chance that the "real" fraction of positive ratings is at least what?" What normal person thinks in terms of confidence intervals? The obvious answer is people want the product with the highest "real" rating. That is the rating the product would get if it had arbitrary many ratings. To get this you just find the mean of your posterior probability distribution. For just positive and negative reviews thats basically (positive+a)/(total+b) where a and b depend on your prior. His proposal would mean that a product with zero reviews would be rated below a product with 1 positive review. This may deal with spam and vote manipulation since things with less information are penalized more but that is a separate issue.