This looks promising and useful, but like most ranking factors can be gamed pretty easily. Imagine a low quality journalism site that caters to both partisan e

This looks promising and useful, but like most ranking factors can be gamed pretty easily. Imagine a low quality journalism site that caters to both partisan extremes by the simple expedient of giving each audience different intra-site recommendations. Such a site would simply be relying on the existing filter bubbles to segregate the audiences from each other (and to the extent that those bubbles leak, a "let's you and him fight" strategy would probably work to induce more outrage), and incorporating audience diversity data gives them an extra boost. You would need the 'partisan diversity' measure to indicate whether readership is spread along the entire partisan spectrum to avoid being fooled by a bimodal distribution. Just measuring variance won't do it.