Well, when "simply" extracting the core text of an article is a task where most solutions (rule-based, visual, traditional classifiers and LLMs) rarely score above 0.8 in precision on datasets with a variety of websites and / or multilingual pages, I would consider that not too bad.