The thing to know here is the bayes factor. That’s the true positive rate divided by the false positive rate. In this context, it’s the percent of successful articles that have a property (like using the number 10) divided by the number of unsuccessful articles that have that property. This removes any advantage a property gets from being more common.

Right.

The result for headlines of 65 chars - shared 50,000 more times than 60 chars or 70 chars - seems too incredible to occur at random and suggests instead that a popular news source has implemented a 65 chars policy.

[Edited to note: Yep. YouTube is dominant as the popular publisher in this review, and truncates headlines at 66 chars - that's what this article observes]

The charts use median engagement, which helps normalize against frequency.

That said, a boxplot with 25th and 75th percentiles would likely indicate there is a heavy skew, as tends to be the case with social media data.