The retrieval tax, where AI agents repeatedly fetch and process web pages within the context window, inflates token bills by up to 48 times compared to answering from memory, due to re-sending the entire growing context with each search hop. A deliberately simple experiment on a known question, "What was Y2K?", showed a web search loop consuming 28,700 tokens over three hops, versus 600 tokens from memory. Utilizing an owned, pre-processed web index, which cleans and stores documents ahead of time, reduced token consumption to 6,900 tokens in a single call, achieving a 76% cost saving by eliminating repeated fetching and parsing.
Agent operational costs are heavily burdened by inefficient retrieval, creating a significant demand for optimized data pipelines and pre-indexed web services.