TL;DR
- Google DeepMind, UMass Amherst and UT Austin researchers proposed Autoregressive Ranking (ARR), where one fine-tuned LLM generates ranked document IDs token by token.
- ARR would replace the usual two-stage pipeline of a fast dual encoder for retrieval and a costly cross encoder for reranking.
- On WordNet ARR performed similarly to a cross encoder and far better than a dual encoder; on ESCI shopping queries one variant got worse at ranking the top result first.
- Search Engine Journal reported the paper on 10 September 2026; nothing in it says Google Search uses ARR or gives a deployment timeline.
Researchers from Google DeepMind, the University of Massachusetts Amherst and the University of Texas at Austin have published a paper proposing that a single large language model could replace the two-stage retrieval and ranking pipeline most search systems use today. The method, called Autoregressive Ranking (ARR), has the model generate a ranked list of document identifiers directly. Search Engine Journal reported the paper on 10 September 2026. It is research, not a product announcement, and nothing in it says Google Search uses the approach.
Still, the paper is worth understanding, because it describes a different way for a machine to decide which pages come first. This post explains the current two-stage design, what ARR changes, what the experiments showed and did not show, and what a site owner can sensibly take from it today.
How most search ranking works today: two stages
The paper, titled "Autoregressive Ranking: Bridging the Gap Between Dual and Cross Encoders", starts from the standard architecture. Search systems usually split the job in two because doing it in one step is too expensive.
Stage one: the dual encoder
A dual encoder (DE) turns the query into a vector and turns every document into a vector, separately. Because document vectors can be computed in advance and stored, finding the documents whose vectors sit closest to the query vector is fast and cheap, even across a very large index. The trade-off is precision. The query and the document never "see" each other during encoding, so fine distinctions between two similar pages are easy to miss.
Stage two: the cross encoder
A cross encoder (CE) reads the query and a candidate document together and scores how well they match. That joint reading is far more accurate, and far more expensive, because it has to run once for every query-document pair. So systems use the dual encoder to shortlist candidates and the cross encoder to put the shortlist in order.
For SEO, this split has been a useful mental model for years: first you have to be retrieved, then you have to win the reranking. ARR proposes collapsing both into one model.
What Autoregressive Ranking changes
In ARR, a language model is fine-tuned so that, given a query, it generates document identifiers (docIDs) token by token, in ranked order. Instead of comparing vectors or scoring pairs, the model's own next-token probabilities decide which document comes first, second and so on. Each docID can be several tokens long, and the set of valid docIDs is organised as a prefix tree so the model can be steered toward identifiers that actually exist.
The obvious risk with generating identifiers is that a model invents one that does not correspond to any document, or ranks an irrelevant document highly because it is a likely token sequence. The paper's training method is aimed squarely at that problem.
The training method: SToICaL
The researchers call their loss function SToICaL, short for Simple Token-Item Calibrated Loss. As Search Engine Journal summarised it, it teaches ranking in two ways. Documents that should rank higher get more weight in training and documents that should rank lower get less. And the ground-truth ranking is used to push probability toward the tokens that lead to higher-ranked documents. The paper describes this as item-level reweighting combined with prefix-tree marginalisation, spreading probability over valid docID tokens according to their relevance.
| Approach | How it ranks | Trade-off described in the paper |
|---|---|---|
| Dual encoder (DE) | Separate query and document vectors, nearest vectors retrieved | Fast and cheap, but less precise; embedding size must grow with the number of documents to express any ordering |
| Cross encoder (CE) | Reads query and document together and scores each pair | Precise, but too expensive for large-scale retrieval, so used only on a shortlist |
| Autoregressive Ranking (ARR) | One fine-tuned LLM generates ranked docIDs token by token | Performed close to the cross encoder on WordNet; one variant got worse at ranking the single best result first on shopping queries |
What the experiments showed
The team tested on two datasets: WordNet, a lexical database, and ESCI, a set of shopping queries with graded product relevance. According to the coverage, the findings were mixed in a way worth stating precisely.
- SToICaL improved ARR's ranking, and the researchers wrote that rank-aware training "significantly improves ranking metrics beyond top-1 retrieval".
- On WordNet, the rank-aware methods sharply reduced the error of placing irrelevant documents above relevant ones.
- In a separate WordNet comparison, ARR performed similarly to the cross encoder and significantly better than the dual encoder.
- On the ESCI shopping data, one version of the method became worse at putting the most relevant result first, even as it improved the ordering of the list overall.
That last point matters. For many commercial searches, the first result takes a large share of attention, so a system that improves the list but slips on the top slot is not straightforwardly better. The researchers flag it as an area for further work.
The theoretical claim, and its limits
The paper's central theoretical argument concerns capacity. The researchers show that for a dual encoder to be able to express any ordering of k documents, its embedding dimension has to grow linearly with k. An ARR model, they prove, can in principle rank an arbitrary number of documents with a constant hidden dimension, given a mild condition on the embedding matrix for docID tokens.
Search Engine Journal was careful to note that this is a theoretical result. Proving that a model has enough expressive room to represent any ranking is not the same as showing it will rank well on a real web-scale index with billions of pages, constant updates and adversarial content. The experiments used two research datasets, not live search traffic.
What the source did not say
The coverage does not say Google plans to deploy ARR, that any Google product uses it, or on what timeline anything might change. It gives no latency or cost figures for running a generative ranker at search scale, which is the practical barrier any such system would face. It does not describe how freshness, spam resistance or index updates would work when documents are identified by generated tokens. And it does not report tests on web search data. Readers should treat claims that "Google is replacing its ranking system" as unsupported by this paper.
Why it still matters for SEO and AI visibility
Search Engine Journal's own takeaway cut both ways: the paper shows dual and cross encoders still underpin ranking today, and it also shows research moving toward systems where one model makes the whole decision. Three practical points follow from that, none of which depend on ARR shipping.
Retrieval and ranking are both model decisions
Whether ranking stays two-stage or becomes one generative step, the decision is made by models reading your content in relation to a query. Pages that state clearly what they are about, answer specific questions in plain language and use consistent entity names give any of these architectures less to guess. That was good practice for cross encoders and it remains good practice for generative rankers.
Shopping queries behave differently
The ESCI result is a reminder that product search has its own difficulty. Graded relevance (exact match, substitute, complement, irrelevant) is harder to order than a yes-or-no match. For e-commerce sites, clean product data, distinct titles and attributes that separate close variants help any ranking model tell products apart.
Generative systems already rank sources
AI answer engines already choose which pages to cite, and that choice is a ranking problem. The ARR paper is not describing AI Overviews or AI Mode, but it belongs to the same family of ideas: a language model deciding order. Work on AI SEO and GEO is about making content easy for those systems to select and quote.
What to do now
- Change nothing because of this paper alone. It is research on two datasets. Rankings will not shift because it was published.
- Keep content unambiguous. One main topic per page, the answer near the top, entity names used consistently.
- For product catalogues, invest in attribute quality. The shopping result shows why close variants are where ranking models struggle.
- Be sceptical of vendors selling "ARR optimisation". There is no deployed system to optimise for.
What this means for Thai marketers
The paper does not mention Thai or any other language beyond its English research datasets, so there is no basis for claims about how ARR would handle Thai queries. What Thai businesses can take from it is the direction of travel. Search is moving toward language models making more of the ranking decision, and language models depend on text they can read and match with confidence. Thai sites that publish machine-translated pages, mix Thai and English product names inconsistently, or bury answers under long introductions give any model a harder job. A clear SEO Thailand content structure helps with today's ranking systems and with whatever comes next.
Frequently asked questions
Is Google using Autoregressive Ranking in Search?
There is no evidence it is. The paper is research from Google DeepMind and two universities, and the coverage does not say it is deployed in any Google product or give a timeline.
What is the difference between a dual encoder and a cross encoder?
A dual encoder converts queries and documents into vectors separately for fast retrieval, while a cross encoder reads each query and document together for more accurate but more expensive scoring. Search systems typically use the first to shortlist and the second to order the shortlist.
Did ARR beat existing ranking methods?
Partly. On WordNet it performed similarly to a cross encoder and significantly better than a dual encoder, but on shopping queries one variant got worse at ranking the single most relevant result first.
Should I change my SEO strategy because of this paper?
No, not on the strength of a research paper alone. The durable lesson is to keep pages clear and specific, which helps every ranking architecture, including the AI systems that already choose citations.
Research papers are a useful look at where search could go, as long as they are read for what they prove and nothing more. For help building content that works with today's ranking and AI citation systems, speak to Relevant Audience.
