For comparison with Google Prediction API I used classifiers embedded in tool Weka (namely J48, SMO, RandomForest, etc.). The results are interesting. Google Prediction API behaves more or less the same as other classifiers, despite the fact the whole process of machine learning was much simpler, since there is no need to select the propper ML method or customize the parameters of selected method. My train set consisted of ~3000 randomly chosen web pages, which of 1000 were negative (non-czech) examples.

Results: Google Prediction API/Wekka clasifiers accuracy: ~98%

More details can be found in the presentation in the download section, although it is written in Czech language.