Every time Google decides which page shows up first, two separate systems are quietly doing the work behind the scenes. One casts a wide net fast. The other narrows it down with much more care — and DeepMind just published research asking whether you need both.
TL;DR: Google DeepMind published research on a single AI model that could replace the two-stage system search engines use to rank pages. It's early-stage research tested on academic datasets, not something live in Google Search — but it hints at where ranking technology might head next.
🧠 How Search Ranking Actually Works Right Now
Right now, ranking a page happens in two steps. Step one uses something called a dual encoder. It turns your search and every page in the index into number-based representations, then quickly shortlists a batch of documents that look relevant.
Step two hands that shortlist to a cross encoder. This one is slower, but it actually reads through each candidate page to rank it properly. Running it on the whole index would be too expensive, so it only touches the smaller list the dual encoder already found.
One Model For Both Jobs
DeepMind, working with researchers from UMass Amherst and UT Austin, built something called Autoregressive Ranking, or ARR. Instead of two separate systems, one large language model handles both jobs — casting the wide net and doing the careful ranking. They trained it with a method called SToICaL, which teaches the model to push relevant documents up and irrelevant ones down.
In testing, ARR ranked about as well as the slower, more precise cross encoder — while working more like the fast dual encoder. That's the appeal: cross-encoder-level accuracy without needing a separate, expensive second stage. The researchers also proved something theoretical worth noting — ARR can rank an arbitrary number of documents without needing more computing power as the index grows, something the dual encoder can't claim.
It didn't win everywhere. On one e-commerce dataset, one version of ARR actually got worse at picking the single best result, even as its overall ranking improved. None of this is live in Google Search — the tests ran on two research datasets, not real search traffic, and DeepMind hasn't published cost or speed numbers at the scale Google actually runs at.
How Ranking Works Today
Dual encoder quickly shortlists likely pages from the whole index
Cross encoder then carefully re-ranks just that shortlist
Two separate systems, each doing one job well
What ARR Proposes
One LLM handles shortlisting and ranking together
Matched cross-encoder accuracy in early tests
Still a research paper — not deployed anywhere
"Two systems doing one job is expensive. One system doing two jobs is still unproven. That's the bet DeepMind is making with ARR."
🛠️ What This Means For Your SEO Right Now
Keep Optimizing For Relevance, Not Tricks
Whether ranking runs through two systems or one, the underlying signal is the same: does your content actually answer the query? ARR's whole selling point is reading relevance more accurately, not differently. Writing genuinely useful pages, backed by real experience, still wins under either architecture.
Watch How This Research Moves From Paper To Product
Big ranking shifts — RankBrain, BERT — all showed up first as research before they touched live search. If ARR, or something like it, ever ships, it likely arrives quietly inside an update rather than as an announcement. Bookmark this one and check back when Google actually talks about ranking architecture changes, not before.
Keep Pages Fast and Clean Regardless
A more capable ranking model still has to read your page to rank it. Clear structure, fast load times, and content that's easy to parse help any system — LLM-based or not — understand what you're offering. None of that changes no matter which architecture ends up winning.
Hit reply
Would you trust one AI model to both find and rank pages in a single step — or does splitting the job into two feel like the safer design?
One thing I'm thinking about: SEO people love treating research papers like they're ranking factor updates. This one isn't — it's DeepMind testing an idea on a couple of academic datasets, nowhere near a live index with billions of pages and real user behavior.
But the direction is worth noticing anyway, because every major ranking shift Google has made, from RankBrain to BERT, showed up first in a paper that looked exactly like this one.
Till next time,
