This is dead. If you need something similar, better look for RDFa instead.
And even before thinking about that, you can actually put the dimensions in a description, which some do (like Ikea) and Google is definitely able to pick up on that, no RDFa was ever needed. As far as I can tell, LLMs can work that out just fine as well.
The problem with the metadata discussion is that if they are actually useful, there is no reason that they are not useful to humans as well, so instead of trying to make the human work for the machine it is much better to make the machine understand humans.
I personally think semweb-related technologies could play a significant and productive role in synthetic data generation, but that's a whole different conversation that is beyond the current era of LLMs.
If you show me an LLM that can take in serialized RDF and perform reasoning on it, I will be surprised. Once the LLM takes in the RDF serialization, it is a dead end for that knowledge: by principle, you can't rely on anything the LLM does with it.
In a world of LLMs, it makes much more sense to put the semweb technologies alongside the training step instead. You create ontologies and generate text from triples, then feed the generated text as training for the model. This is good because you can tweak and transform the triples-to-text mechanism in all sorts of ways (you can tune the data while retaining its meaning).
It doesn't make much sense to do it now, but if (or when?) training data becomes scarce, converting triples to text might be a viable approach for synthetic data, much more stable than having models themselves generate text.
It's easy, you just ask the LLM to convert your question to a SPARQL query, then you sanity-check it and run it on your dataset. The RDF input step is just so that the LLM knows what your schema looks like in the first place.
Can you make me a quick demonstration using a publicly available model and the dbpedia SPARQL endpoint?
And there is the fact that people/orgs lie/maintain poorly and thus you cannot trust any open tag/schema anymore than the data it is supposed to describe. You end up in a conundrum where your data may be sketchy but on top of that the schema/tags may be even worse, at the end of the day if you got to trust one thing it's got to be the data that is readily visible by everyone because it's more likely to be accurate and that's actually what's relevant.
To top all of that, tagging your stuff with RDFas just makes it easier for google and others to parse your stuff and then exploit the data/information without ever sending anyone to your site. If you are Wikipedia, it's mostly fine, but almost anyone else would benefit from receiving the traffic for a chance at value conversion/transaction.
All those metadata things are really idealized academic endeavor; they may make sense for a few noncommercial stuffs but the reality is that most website on the web need to find a way to pay the bills, making it easier for others to exploit your work is largely self-defeating.
And yes, LLMs didn't even need that to vacuum most of the web and still derive meaning, so at the end of the day it's mostly a waste of time...