A human-targeted web structure that contains some lies is still useful for humans because humans can filter out those lies with somewhat satisfactory efficiency.

A machine-targeted web structure that contains some lies is not useful for machines because they can't filter out those lies yet. It might become useful when they can (but that might be a hard-AI problem), but it's simply not usable until that point.

What's the point of having the marks at all if they are not actually correct?

If you're answering the question, I think it would be good to answer it directly.

We as humans have a lot of intuitive tools for knowing whether a source of data is trustworthy. AI could possibly approach this ability given enough training... we'd need to do something like add a "trust" score to every node in the graph.