TheMadScientist - 10:14 pm on Jan 13, 2013 (gmt 0)
Well, it does include quite a few 'on-page metrics' when you take into account they could easily use color, template, design, contrast, etc. and the 'more reliable way' to do it, (like I tried to say earlier but probably didn't very well, because this is tough to explain), is to look at behavior in the results and determine how to 'gauge that behavior' within them more accurately based on external sources, so, spelled out...
You write an algo to refine the results shown for a person based on their behavior in the results. Then you look at the user behavior within the results where you have access to the external data for visitors and use the external sources + in result behavior to see if pages/sites exhibiting positive signals via those external sources move up or down. (IOW: Did you 'get it right' based on 'external confirmation' from data you have access to.)
You could probably even 'do it the other way around' and look at the external sources for 'positive behavior signals', then look at the behavior of those visitors in your results and figure out how to write an algo that applies 'Internet wide on a per-visitor behavior basis' to further refine the results for a person by making sure you move the sites exhibiting positive signals up for people exhibiting similar behavior in the results to those you can 'see after they left' via external sources.
What you would not do, because it doesn't apply Internet wide is actually write the data from the external sources into the algo ... You would use them for 'verification' or 'direction', but not directly, because if you use them directly you 'miss' too much, but if you use them for 'direction' or 'verification' and find a better way to refine your results, then it's an overall win that scales to the entire index easily.
You is of course generic in this post.