> I doubt they are so sophisticated.
Most decisions are based on machine learning that uses both manually crafted and automatically generated features, so it is not usually possible to find out what the decision is made on. But count of supporting and contradictory statements seems like a reasonable feature to use.
As to how far someone would go in optimizing the system – one high-end web designer went public complaining about his work for google some years ago. He designed a good-looking blue line on the google main page. Google engineers tested his against another hue and found out that the other hue gets 5% improvement in something, so they changed the hue to the more plain-looking one despite the protests of this artist.