Navigational search is the task of finding one specific page among hundreds of millions. Current systems either only answer queries they have seen before or rank pages by similarity, often missing the exact target.
Octen-WebGR stores web pages inside a language model's parameters. Each page gets a two-token code; the model generates that code from a query, and a table maps the code to the real URL.
On new queries for pages with many past clicks, the model ranked the correct page first 30.9% of the time, compared to 17.2% for a dense retriever. This shows generative retrieval generalizes better to new phrasing, improving search for familiar pages.