Listen: How Brave's AI Search Works

Brave Search has a feature called Brave Ask, which uses artificial intelligence to drive web searches, filter results, and present answers. Recently, leaked system instructions revealed exactly how this pipeline works under the hood. The system operates in two halves: retrieval and generation. First, the retrieval layer turns a user query into a body of web evidence. In single search mode, this happens in under five seconds. In research mode, the AI iterates, running multiple searches to analyze hundreds of pages and extract thousands of individual statements. The generation half then takes over. To prevent the AI from making things up, Brave switches off its internal knowledge. The model is strictly instructed to rely only on the retrieved web statements. Anything irrelevant is discarded. While composing the answer, the model actively builds the page by inserting custom blocks for videos, images, discussions, or product links. Crucially, the search queries behind these blocks are written by the AI itself, based on the text it just drafted. This means the citations you see on the page are actually the results of fresh, real-time searches. To keep costs down and speed up response times, the instructions demand extreme conciseness. Redundancy is banned, prose is preferred over lists, and formatting rules are strict. Finally, Brave instructs the AI to remain neutral, present multiple viewpoints, and preserve real-world nuances rather than oversimplifying. It is a highly coordinated dance where search retrieves the facts, and artificial intelligence glues them together.