I think it's as simple as taking every search result and lumping each into one of the three categories (pos, neg, indiff).
There would be lists of positive words and negative words to tally for each link.
More good words than bad words on a particular link, and that result is counted in the 'Positive' pile.
I doubt they even use any unique keywords for 'Indifferent', as it would most easily be implemented as a category for results that are roughly balanced in total + and - keywords.
If my guess is correct then it's basically guilt-by-association, so to speak.
Ahh, the complexity of free, online, automated opinion-makers.
24 hour TV news should add this to their repertoire:
' Well I don't know, Let's check online! ... What does the internet think? '
http://www.whatdoestheinternetthink.net/news
OK, that a definitive conclusion is drawn at all there^ bothers me ... how accurate are these results again?
"This, of course, produces questionable results which should not be taken very seriously. However, the more results (hits) returned, the more reliable these results can become."
"The results are provided 'as is' and should not be considered reliable, nor do they reflect the opinion of whatdoestheinternetthink.net, its creators or Microsoft. Furthermore, results may vary greatly on a daily, or even hourly, basis. The results are merely a reflection of a majority in search term results reported by said search-engine."
Which leads me to believe that any robocalypse in the near future would be schizophrenic at best. If I am permitted to anthropomorphize hypothetical events, I should do so more often.