Retrieval-Augmented Generation (RAG)

RAG allows current web information to affect an AI answer without waiting for the next model training cycle. Updating the pages, profiles, and reviews that retrieval finds can change the evidence available for future answers, though the result remains dependent on the query and sources selected.

A simplified pipeline has three stages. Retrieval finds candidate documents, often through a search query the assistant composes. Selection narrows them to sources that fit the context window. Generation writes an answer from the selected context. A page must first be retrieved and then provide information relevant enough to remain in the selected context.

Selection may use only a few pages from a larger result set. Those pages shape the available evidence before the model generates or explains a recommendation.

Several controls affect this process: the search query, document index, embedding or keyword match, reranking method, context limit, and instructions given to the model. Consumer assistants do not disclose every control. Measurement should therefore record the visible conditions and sources without claiming to know the complete retrieval pipeline.

Answers can change even when the business has not. Retrieval depends on phrasing, location, timing, and the assistant's search query. A competitor's new comparison page or a newly prominent forum thread can change which sources are fetched.

To improve a page's chance of retrieval, make its subject, location, date, and claims explicit in visible text. Use descriptive titles and headings, provide crawlable facts, and link the page from relevant established pages. These steps improve clarity but do not guarantee selection for any one answer.

Reviewing retrieved sources shows which pages supported an answer such as "best roofer in Phoenix." Routeless research records those sources by trade and market.