The issue isn't sample size (nor many of the other things people are pointing out).
First off we're ultimately talking about ordinal data here and even the Evan Miller post in another comment misses that ordinal data is fundamentally tricky. Everyone is talking about the p("most loved") but that makes assumptions that "most loved" has a consistent meaning.
The problem with ordinal data is that the only thing that's known about it is an ordering of the values, but the distance between values is undefined. That is the difference between a 5 star and a 4 star review is not necessarily that same a between a 2 and 1 star review. This means that you can't meaningfully average these.
However there is an even bigger problem, which is what I think parent is feeling here, and that is selection bias. It's very hard to compare the average rating of a film such as Star Wars and a film like Cannibal Holocaust. The first is a general audience film and the later is an extremely niche subset of gore horror. If you show Cannibal Holocaust to the general population it will receive wildly lower rating than Star Wars. However if you were to do a survey of all people who have a DVD of Star Wars and a DVD of Cannibal Holocaust, I wouldn't be at all surprised if Cannibal Holocaust was more loved. There are plenty of people that have Star Wars on their shelf and feel 'meh' about it, but almost no one who owns Cannibal Holocaust on DVD that doesn't feel like it is an essential horror classic.
The annoying truth is, there's really no statistical way to solve this. There are some solutions to better modeling ordinal data (you could for example, transform this problem into an ordinal regression problem). But ultimately "loved" is not really a measurable thing, so we're always looking for imperfect proxies.