When a test break, the developer looking at it must be able to easily understand what was the correct behaviour that was tested. For example, if all the dev knows is that the output file differs in such and such way from the "correct" output, he shouldn't have to scratch his head in order to understand whether that difference matters or not (often, while the HTML output of a script changes slightly, the actual behavioural intent of the script is still fuilfilled... maybe the difference is just a different way of formatting the output...)