The whole idea of the model is to fit it on some reference so you can project it onto new data you haven't classified yet. So your first few weeks, you might ge

The whole idea of the model is to fit it on some reference so you can project it onto new data you haven't classified yet. So your first few weeks, you might get some spam sure. But eventually if you are optimizing the system by e.g. only ranking technical blogs with no seo bullshit, then technical blogs with no seo bullshit is what would fit best with that model among the new data and what you see on your results. That can't be gamed without knowing the sort of model fit you are generating and this fit is defined by the user.