Google DeepMind is testing an AI model that could replace search ranking as we know it

Google DeepMind has published research on a system called Autoregressive Ranking, built to replace the two-stage architecture that has powered search ranking for years. The paper is a collaboration between DeepMind, University of Massachusetts Amherst, and The University of Texas at Austin, and it proposes a single LLM that ranks documents directly.

How ranking works today

Most search ranking systems run in two stages.

• A Dual Encoder converts queries and documents into vectors and quickly retrieves a shortlist of likely matches. It is fast and inexpensive to run at scale.
• A Cross Encoder then reviews that shortlist and ranks it with far more precision, but it is too computationally expensive to run across an entire index.

The split exists because no single system has managed to be both fast enough for retrieval and precise enough for ranking. Autoregressive Ranking, or ARR, is DeepMind’s attempt to close that gap with one model.

How the new model learns to rank

The researchers built a training method called SToICaL, short for Simple Token-Item Calibrated Loss, to teach the LLM which documents deserve to rank higher.

Training works two ways. Higher-relevance documents are weighted more heavily during learning, and the model is guided toward token choices that lead to those higher-ranked results. The outcome is a model that learns relevance directly and actively suppresses irrelevant documents, rather than treating ranking as a side effect of retrieval.

What the testing showed

DeepMind tested ARR against WordNet and ESCI Shopping Queries datasets, and compared it directly with standard Dual Encoders and Cross Encoders on WordNet.

The results were strong in most areas. SToICaL improved ARR’s ranking beyond simple top-1 retrieval, and it sharply reduced cases where irrelevant documents outranked relevant ones. On WordNet, ARR performed close to the expensive Cross Encoder and clearly ahead of the Dual Encoder.

One gap surfaced in the shopping-search test: one version of the model got better at overall ranking quality but weaker at surfacing the single most relevant result first. The researchers flag this as an open area for further work.

Why this could matter more than it looks

The theoretical case is the more interesting part. Dual Encoders hit a structural ceiling: to represent every possible ordering of documents, their vector size has to grow with the number of documents being ranked. DeepMind’s researchers prove that ARR does not carry that constraint. A model with a constant hidden dimension can, in theory, rank an arbitrary number of documents.

That is not a claim that Google is about to swap this into live search. It is a research paper, tested on academic datasets, with a still-open weakness on top-1 accuracy in at least one scenario. But it signals where Google’s own research is pointed: toward a ranking system built on a language model’s reasoning rather than a fixed two-stage pipeline.

What this signals for search strategy

Search has already moved toward language-model reasoning in how results and AI-generated answers get assembled. This research suggests the same shift may eventually reach the ranking layer itself, the part of the stack that decides which pages are relevant in the first place. If ranking increasingly runs through an LLM’s judgement rather than a strict vector match, relevance signals such as structure, clarity, and how well a page answers the underlying intent behind a query carry more weight than keyword-matching ever did.

That is the same shift already playing out in how Google’s AI Overviews select and cite sources, and in how models like ChatGPT and Perplexity decide what to surface. Teams that have already restructured content for AI-driven relevance through Google AI Overview optimisation are working with the same underlying principle DeepMind is now formalising in research: ranking is moving from matching to reasoning.

It is early. But it is worth watching, because it shows Google is actively researching what comes after the ranking systems every SEO strategy has been built around.