The problem is the downloaded pages of the papers themselves, which often completely lack proper bibliographic data (or even header/footer info other than a page number). Compare this with a page from a corporate tech report, where each page might have the title and document number somewhere.