Google researchers have now shown, on paper, that a single language model can rank an unlimited number of documents without a separate index. The big shift: the ranked list stops being a surface you read and becomes a hidden step on the way to an answer.
One model replaces the two-stage pipeline
Today’s search stacks pair a fast dual encoder for retrieval with a slower cross encoder for reranking. The new research from Google’s team collapses both steps into one autoregressive model. In a January 2026 paper, the researchers show that a dual encoder’s embedding dimension must grow linearly with the number of documents to express every possible ranking, while a model with a fixed hidden dimension can rank an arbitrary number.
They also introduce a training loss called SToICaL that pushes the model to care about the whole ranked list, not just the top result. This is a research foundation, not a deployment: the team fine-tuned Mistral-7B, used in-context reranking rather than open-corpus retrieval, and ran tests on WordNet’s roughly 82,000 concepts and a 310-example shopping set.
Why page two could disappear
The mechanism is beam search. A generative model produces document identifiers one token at a time and keeps only a fixed number of the most probable partial sequences. That fixed width is the entire set of results the system computes. Position eleven is not a bad result; it was never generated.
That changes visibility. Below the beam there is nothing to measure, and from the outside a page that ranked badly and one that was never produced look identical. Practitioners should track whether a page is in the generated set, how consistently, and for which query shapes.
The revenue angle
Alphabet’s most recent quarterly report put Google Search and other revenue at $63.27 billion, up 17%. This paper does not touch the ad auction, but it attacks the cost side. Removing a separate nearest-neighbor index and avoiding per-document scoring lowers the structural cost of generative results, and weakens the last reason to keep ten blue links as the default surface.
Duane Forrester, founder of CitationIQ, has argued that ad formats announced at Google Marketing Live, with creative generated from a query’s wording, are an early piece of this shift. The ranking proof completes the mechanism: organic document IDs and ad text could eventually be generated in one decoding pass.
What to do now
- Audit pages that are not distinct by meaning; near-duplicates fight for the same generated address.
- Start logging inclusion in AI-generated answer sets now, so you have a baseline before visibility flips to binary.
- Re-read paid strategy against query-time creative, because that side has already started shipping.
Google has spent five years moving in this direction. The paper supplies the theory, not the product. But the practical question has shifted from where do we rank to whether we are in the set the model generates.
Source: Search Engine Journal
