DeepMind tests a new AI search ranking model that works in one pass

Google DeepMind has published research on a new AI search ranking method that uses one model where systems today use two. The paper is called “Autoregressive Ranking: Bridging the Gap Between Dual and Cross Encoders”. It was written with the University of Massachusetts Amherst and the University of Texas at Austin. Search Engine Journal reported on it on 10 September.

Ranking today runs in two steps, ARR runs in one

Search ranking normally works in two stages. A dual encoder turns the query and every document into numbers and picks a shortlist fast. A cross encoder then reads the query and each shortlisted document together and puts them in order. The second stage is accurate but slow, so it never sees more than a handful of candidates.

The method in the paper, called Autoregressive Ranking or ARR, does both jobs at once. One language model reads the query and writes out document identifiers in the order it thinks they belong. The authors argue this scales better. A dual encoder has to grow as the list gets longer, while ARR is, in their words, “sufficient to rank an arbitrary number of documents”.

SToICaL training, tested on WordNet and ESCI shopping queries

The team also built a training recipe named SToICaL. It gives extra weight to documents that belong near the top. It also spreads the model probability across the tokens that form a valid document identifier, which plain language model training does poorly.

Tests ran on two public collections: WordNet and ESCI, a set of shopping queries. ARR matched cross encoders on quality without their cost. One shopping variant got worse at picking the single best result, even though the overall order improved.

Nothing in the paper is running inside Google Search

This is a research paper, not a product note. Google has not said that ARR sits in any live system, and DeepMind papers often stay in the lab. Read it as a signal about direction. It sits next to other moves that push models deeper into results, such as the switch to Gemini 3.8 Flash inside AI Mode.

For SEO teams, ARR changes nothing this month

The honest answer is that nothing needs to change today. But if one model ever writes the whole ranked list, the old idea of beating a rival page on a checklist of signals gets weaker. What matters more is whether a model can see that your page answers the question. That is the same pressure behind the finding that AI now writes 97% of People Also Ask answers. Put the answer high on the page, in plain words, and keep the facts visible instead of hiding them behind a click.