Google filed a patent in 2023 describing how a large language model could construct a structured understanding of a business, brand, or product from its website and public sources, even when none of that material was ever marked up for machines to read. The filing, surfaced and analyzed by Search Engine Land, outlines a four-step process: identify the entity from a domain, interpret the content across pages, extract attribute relationships, and supplement with third-party signals including Maps data, job listings, and user reviews.
The system does not extract verbatim text. According to the patent language, the output is “an interpretation of the extracted content rather than a verbatim duplication.” The LLM synthesizes an entity summary, closer to a brand characterization than a page index, organized into a hierarchical graph that connects services, audiences, reputation signals, and differentiators into a structured model.
This is not a ranking patent. It is a representation patent. The distinction matters for search teams. Google has long indexed what a page says. This describes a system for deciding what a business is, built from evidence spread across the entire web.
The practical pressure this creates is different from traditional on-page optimization. A service page that ranks well for a keyword may simultaneously teach the model that your business is a commodity provider with no differentiated positioning, if the surrounding signals (reviews, press, job descriptions, social content) reinforce that reading. The system, as the patent describes it, synthesizes across sources rather than treating any single page as authoritative.
GEO, generative engine optimization, the practice of optimizing for LLM-driven answers, has mostly been treated as a citation problem: get mentioned by the right sources, get quoted in AI Overviews. This patent frames it differently. Before AI Overviews can cite an entity, the underlying model needs a coherent characterization of that entity. An incoherent or thin characterization may mean the brand is simply omitted from a generated comparison or recommendation, not outranked.
The filing applies equally to local businesses, ecommerce products, and enterprise organizations. For a local service firm, the characterization might draw as much from Google Business Profile reviews as from the website. For a B2B software vendor, job postings and analyst mentions may register alongside product pages. The patent explicitly names maps data, business directories, and job listings as supplementary inputs.
One important caveat: Google publishes patents by the thousand, and most never reach a shipped product. The patent does not describe an active ranking signal or a confirmed change to AI Overviews citations. What it does describe, in concrete engineering language, is where Google is steering its entity-understanding research.
Search teams running GEO audits in the next ninety days should add an entity audit step: ask what an LLM would characterize your brand as if it read your homepage, your top 20 reviews, your LinkedIn page, and your three most recent press mentions together. The gaps in that characterization are the gaps in your AI search visibility.
Reporting by Rich Sanger for Search Engine Land, published June 22, 2026.
