yeah, LLMs sound cool for extraction, but I'd recommend you - go hybrid. Use XPath for the dead-certain stuff like prices, SKUs, IDs, whatever and only throw semantic extraction at the messy, inconsistent bits. Seriously cuts your latency and doesn't torch your budget. The thing that actually nukes production systems is silent failures, which is why we built output density monitoring - just watches when your extracted fields dip below normal and fires alerts before layout shifts poison the whole downstream pipeline. On the residential proxy side, yeah the heat's on, but it's all about timing: blast your campaigns too hard and you'll get made quick, but if you play it cool with naturalistic patterns and sticky per-target rotation you're still golden - the real squeeze isn't whether it works anymore, it's volume and discipline. Sites going full JS-on-demand is actually the move that kills traditional scraping dead, but it's not ending the game, just reshuffling the playbook