Search is always changing. Normally we’d look at live algorithm updates to reflect this, but this month, the biggest sign comes from Google DeepMind’s research team. Their new paper, Autoregressive Ranking: Bridging the Gap Between Dual and Cross Encoders, features a very different way to decide what appears first in search results. This approach could eventually replace the system search engines have used for years.
This is early-stage research, but it still matters because it shows where ranking is going. This matters whether you’re optimising for Google search or AI-powered platforms from Google’s own AI Overviews to other big hitters like ChatGPT, Gemini and Perplexity.
What Google DeepMind actually built
Researchers from Google DeepMind, the University of Massachusetts Amherst and the University of Texas at Austin have proposed a system called Autoregressive Ranking.
Most search engines use a two-step process built around dual and cross encoders. Dual encoders turn queries and documents into vectors, enabling a huge index to be searched quickly, but they’re not especially precise at ranking. Cross encoders take a smaller shortlist and rank it more accurately by comparing the query and document together. This method is much more precise, but it’s too computationally expensive to use for the whole index.
Every major search engine has had to balance speed and exactness. ARR suggests removing the two-step process entirely. Instead, it uses one large language model to generate a ranked list of documents directly. The researchers train it with a method called SToICaL (Simple Token-Item Calibrated Loss), which teaches the model to weight higher-ranked documents more heavily as it generates its output.
In tests, ARR achieved accuracy close to cross encoders, but performed more like the faster dual-encoder approach. Importantly, the researchers showed it can scale to rank any number of documents without needing larger embeddings as the index grows. This is a real solution to a scalability problem that goes beyond small accuracy improvements.
Why this isn’t “just another algorithm story”
It’s easy to see this as just interesting theory and move on. To be fair, it’s still theoretical for now. There’s no sign that this is live in Google Search or any AI Overview yet. Still, there are a few reasons to pay attention:
- It’s Google DeepMind. When DeepMind publishes on ranking architecture, it’s a reasonable signal of the direction Google’s own research is heading, even years before anything reaches production.
- It blurs the line between traditional and AI search ranking. At its core, ARR is a substantial language model generating a ranked list. This is the same basic method that powers AI Overviews and answer engines when they decide what to cite. As ranking becomes more like generation, it makes less sense to treat classic Google Search and AI search as completely separate fields with distinct strategies.
- It supports what we’ve been telling clients for some time: treating “the algorithm” as a fixed thing you can optimise for isn’t a winning strategy. Search is moving from simple retrieval-and-rank systems to models that judge relevance more holistically. This means you need to consider topical authority as much as your keyword portfolio.
What this means in practice
We recommend not overreacting to any single research paper. Many promising ranking ideas never make it to production, or they change a lot before they do. Still, the overall trend matches what we’re seeing in Google search and AI search results: ranking systems are getting better at judging relevance more like a person would, not just by matching keywords or backlinks.
Focus on making sure your content shows real expertise and gives clear answers, not just keyword stuffing. Systems that work more like LLMs reward substance over tricks with structure. Make sure your technical SEO basics are solid. Even the best ranking models need clean, crawlable, and well-structured sites to do their job. Don’t treat “Google rankings” and “AI visibility” as two separate tasks. The systems are coming together, and your strategy ought to reflect that.
Keeping an eye on what’s next
We’ll keep following research like this and explain what really matters for brands that want to stay visible, whether in traditional results or AI-generated answers. If you’d like a second opinion on how ready your site is for the future of search, contact us, and we’ll walk you through it.