New AI Reads Websites as Well as Human-Coded Tools — But Cheaper

Meet PACE: an agentic framework that learns how individual publishers structure their pages, then converts those insights into reusable extraction configurations. During training, LLMs analyze page structure and aggregate patterns; at inference, those learned configurations drive a fixed deterministic extractor—so no additional LLM calls are needed, keeping it scalable. Tested across article-body, metadata, and multimodal extraction, PACE outperformed scalable non-manual baselines and approached the quality of manually engineered publisher-specific parsers, with stronger extraction of article text, metadata, images, and tables. It shows that agentic configuration learning can automate publisher-specific extraction for LLM-ready page representations that go far beyond plain article text.