We have published our second research analysis, describing the communication protocol between the visitor’s browser and the language model that powers our AI concierge. To our knowledge, there is currently no public equivalent of a guidance-only browser–LLM protocol. Unlike traditional approaches, the action vocabulary for text guidance is closed and contains no instructions for actions like click, fill, or navigate. Within this protocol, the assistant only highlights and explains, and the visitor performs every click themselves.
To achieve this safely and efficiently, our SDK does not send the page’s entire DOM to the model. Instead, it transmits only the rendered visible text alongside an indexed list of interactive elements. The index-to-element mapping remains secure because it never leaves the visitor’s browser. On a real production storefront, this approach costs 2,866 tokens, whereas sending a simplified DOM costs 335,106 tokens-representing a 99.1% reduction in data overhead.
For a complete breakdown of the architecture, security properties, and token efficiency data, read the full analysis at Guiding, Not Clicking: A Browser–LLM Guidance Protocol.