Google DeepMind researchers have published a paper proposing to replace the traditional two-stage search ranking system with a single large language model. If adopted in production, this would mark one of the most fundamental shifts to how search results are ordered in years, with direct implications for how brands get found, cited, and recommended by AI-powered search.
How does search ranking actually work today?
According to Search Engine Journal’s analysis of the paper, today’s search ranking pipeline runs in two sequential stages. The first is a Dual Encoder (DE), a system that converts both the search query and candidate documents into vectors (lists of numbers that represent meaning in mathematical space), then quickly identifies which documents are likely relevant. Dual Encoders are fast and cheap to run at scale, but they are limited in how precisely they can distinguish one document from another.
The second stage is a Cross Encoder (CE), a more powerful model that reads the query alongside each candidate document to produce a refined ranking. Cross Encoders are more accurate but far more computationally expensive, so they only process the shortlist the Dual Encoder has already narrowed down, not the full index.
This two-stage pipeline is the current backbone of large-scale search ranking, including at Google. Search Engine Journal notes it is still in use today.
What is Autoregressive Ranking, and how does it replace both stages?
The paper, titled “Autoregressive Ranking: Bridging the Gap Between Dual and Cross Encoders,” proposes replacing both stages with a single LLM. The model is autoregressive (meaning it generates output one token at a time, as language models do) and produces a ranked list of documents in one pass rather than running a two-step pipeline.
The research was co-authored by teams from Google DeepMind, the University of Massachusetts Amherst, and the University of Texas at Austin.
To train the model, the researchers developed SToICaL (Simple Token-Item Calibrated Loss), a training technique that teaches the LLM which documents should rank higher and which should rank lower. It works in two ways:
- Documents that should rank higher receive more weight during training; those that should rank lower receive less.
- Known ranking order from training data is used to steer the model toward token sequences that lead to higher-ranked document identifiers.
The result, according to the paper, is a model that learns to surface relevant documents and suppress irrelevant ones in a single generation step.
How does Autoregressive Ranking perform in tests?
The researchers tested ARR (Autoregressive Ranking) on two datasets: WordNet and ESCI Shopping Queries, a product-search dataset. Results were largely positive but not uniform across all conditions.
In the WordNet experiments, ARR performed on par with the Cross Encoder and significantly better than the Dual Encoder alone. The SToICaL training method “drastically reduced” errors where irrelevant documents ranked above relevant ones, as stated in the paper.
In the shopping-search tests, the rank-aware training “significantly improved ranking metrics beyond top-1 retrieval,” meaning the overall ordering of the results list got better. However, one variant of the approach became worse at placing the single most relevant result at the very top. The researchers flag this as an area requiring further work.
The paper also establishes a theoretical finding: for a Dual Encoder to represent every possible ranking of k documents, its embedding dimension (the size of the numerical vector it uses to represent each document) must grow linearly with k. ARR, with a constant hidden dimension, does not face this constraint in theory, meaning it could rank an arbitrarily large set of documents without the model needing to expand proportionally. The researchers are explicit that this is a theoretical result and does not guarantee the same behavior in a real-world production system at scale.
What this means for AI-search visibility
Google may be building toward a future where one AI model understands a search query and decides which content best answers it, in a single step, rather than splitting those tasks between two separate systems. For brands, that changes what “being relevant” actually requires.
Here is a concrete way to think about it. Today’s ranking system works like a two-round hiring process. A fast screener (Dual Encoder) reviews every application and passes a shortlist to a slower, more careful interviewer (Cross Encoder). ARR proposes replacing both with one expert who reads everything and ranks it in a single pass. That expert makes different judgment calls than a two-step process does.
From a Hingewise perspective, three points stand out that the paper itself does not address directly.
First, an LLM-based ranker can potentially weigh a broader set of contextual signals than a vector-similarity system. Current Dual Encoders essentially ask: “how close is this document to the query in mathematical space?” An LLM ranker can consider semantic coherence, entity relationships, and whether the content clearly addresses the question, not just whether the right words appear. Brands that have invested in well-structured, answer-focused content may hold an advantage under this kind of model compared to those optimizing primarily for keyword repetition.
Second, the scaling argument has direct relevance for AI-search specifically. AI Overviews, AI Mode, and answer engines like Perplexity all rely on RAG (retrieval-augmented generation, a process where the AI pulls relevant documents before generating an answer). If ARR-style ranking becomes the retrieval layer for these systems, it could change which documents get selected as source material for AI-generated answers, not just which links appear on a results page. Being cited by AI search depends on being retrieved first, and the retrieval mechanism is exactly what ARR targets.
Third, the shopping-search weakness is a real caveat, not a footnote. If the most relevant result gets pushed down under certain training conditions, brands whose primary goal is to be the top AI recommendation in a product or service category face a specific risk. ARR is not a deployed system, but this result signals that LLM-based ranking is not yet a straightforward upgrade in high-stakes commercial contexts.
Hingewise view: ARR is the clearest signal yet that Google’s research direction treats search ranking as a language understanding problem rather than a vector matching problem. The gap between “keyword-optimized page” and “content an LLM would confidently rank first” is likely to widen as this research moves closer to production. Brands still treating SEO as a keyword matching exercise are optimizing for an architecture that Google is actively researching how to retire.
What to watch as this research develops
- Whether Google patents or engineering blog posts begin referencing autoregressive retrieval methods, which would signal movement toward production.
- Follow-up papers from the same team addressing the top-1 degradation issue in shopping-search tests.
- How AI Overviews and AI Mode evolve in the types of documents they surface as citations, which may reflect underlying retrieval changes.
- Independent replication of ARR results by researchers outside Google DeepMind.
- Any updates to Google’s Search Quality Rater Guidelines that shift how “relevance” is described, which often previews algorithmic direction.
The two-stage ranking system has been the foundation of large-scale search for years. This paper does not end it. But it maps out a credible research path toward what could replace it, and the direction is consistent: toward models that understand content the way a careful reader would, not a fast keyword scanner.
Lam Nguyen · Hingewise
Sources:
Search Engine Journal, “Google DeepMind Develops New AI Search Ranking Model”
This article was drafted with AI assistance from the sources cited above and checked against them before publishing. The "What this means for AI-search visibility" section is Hingewise's editorial analysis, built on the frameworks and positions we use in our GEO work. How we work: Editorial policy.