The main problem is that the incentive for well-intentioned people to add detailed and accurate metadata is much lower than the incentive for SEO dudes to abuse the system if the metadata is used for anything of consequence. There's a reason why search engines that trusted website metadata went extinct.
That's the whole benefit of using LLMs for categorization: they work for you, not for the SEO guy... well, prompt injection tricks aside.
The point is that metadata lies. Intentionally, instead of just being coincidentally wrong. For example everybody who wants to spew LLM produced garbage in your face will go out of their way to attach metadata claiming the opposite. The value proposition of LLM categorization would be that the LLM looks at the same content as the eventual human (if, in fact, it does - which is a related but different problem)
That's the whole benefit of using LLMs for categorization: they work for you, not for the SEO guy... well, prompt injection tricks aside.