Web page transformer for structure information extraction

Topics: Agentic Commerce Optimization (ACO), Data Mining, Indexing, Knowledge Graph, LLMO / GEO, Probably in use, Shopping

The Google patent describes a transformer-based model, termed “WebFormer,” that jointly encodes both the structural layout (e.g., HTML DOM) and the textual content of web pages to extract specified fields (like title, price, date) as text spans. By introducing three token types—field tokens, structured-document-type tokens, and text tokens—and applying four distinct attention patterns (structure-to-structure, structure-to-text, text-to-structure, and local text-to-text), the model efficiently captures both local syntax and global layout information. Extracted spans are stored as contextual representations, enabling scalable, zero-shot extraction across diverse domains without per-domain templates.

This patent applies primarily to Google’s own AI and search systems, including the underlying infrastructure that powers Google Search, Google AI Mode, and potentially systems like Google’s AI Overviews. It is directly relevant to GEO (Generative Engine Optimization) because it reveals how Google’s AI reads and interprets the structural layout of web pages, not just the text, when deciding what information to extract and surface in AI-generated answers.