Google DeepMind has published interesting search-ranking research. It has not announced a Google Search ranking update. Those two things are being blurred together in the coverage, and they lead to very different decisions for anyone responsible for a website.
The paper proposes Autoregressive Ranking, or ARR: an LLM-based way to retrieve and rank documents. It asks whether one generative model could do work usually split between a fast retrieval system and a slower reranker. That is a meaningful technical question. It is not evidence that this architecture is live on Google, or that rankings have just changed because of it.
Still, research from a Google DeepMind team is worth reading for the direction it points. ARR is another sign that search may become better at handling the whole relationship between a question and a document, rather than relying so heavily on a compressed similarity score. For SEO, AEO and GEO, that makes the basics more demanding rather than less important.
What the paper is actually proposing
Most large search systems have to solve two different problems at once. First, find a manageable set of possible answers from an enormous index. Then look at those candidates more carefully and decide their order.
The first job is commonly done by a dual encoder. It turns the query and each document into numerical representations, then finds likely matches quickly. That speed is useful, but the shortcut has a limitation: the query and document have largely been represented separately. A cross encoder can make a more detailed judgment because it considers the query and a candidate document together. It is also expensive, which is why it is generally used only after the shortlist has been made.
ARR is a proposal to bridge that gap. Instead of using a separate retrieval stage and reranking stage, a causal language model generates the identifiers of relevant documents token by token. The paper argues that this can keep more of the expressive power associated with detailed query-document interaction without having to score every document individually as a cross encoder would.
The researchers also introduce a rank-aware training method called SToICaL. Ordinary next-token prediction does not inherently care whether the first useful result comes before the fifth. SToICaL is designed to put more probability on the document identifiers that correspond to higher relevance. In experiments on WordNet and the ESCI shopping dataset, the approach reduced invalid document outputs and improved ranking measures beyond simply finding the top result.
That is promising research. It is also a long way from a public confirmation that Google Search uses ARR. The paper describes experiments and theory; it does not announce a production deployment or a Google ranking-system change. Treating it as one would be the same category error as treating a new benchmark paper as a change to every Google results page.
Why SEO people should still pay attention
The useful implication is not “optimise for document IDs” or any other new trick. It is that the gap between a page that mentions a topic and a page that genuinely resolves a specific need may matter more over time.
A conventional retrieval system needs to be very economical at the first stage. It cannot afford to read the web as a person would. A more capable generative ranking approach has the potential to model nuance better: the constraint in the question, the comparison a reader is actually making, the meaning of a qualifier such as “for a small clinic” or “without moving platforms”, and whether the page answers it rather than merely containing the right terms.
That does not make classic SEO obsolete. Google has said that its AI Mode is rooted in its core quality and ranking systems. Crawling, indexing, page experience, canonicals, internal links and useful content remain the conditions that make a page available to be considered in the first place. A sophisticated ranker cannot reward a page it cannot crawl or a claim it cannot verify.
What it does put pressure on is thin topical coverage. A page assembled around one phrase, with vague promises above the fold and the useful explanation nowhere in sight, may satisfy a crude word match while failing a more complete relevance judgment. The content should make the subject, the audience, the conditions and the answer plain. The evidence should sit close to the claim.
It also strengthens the case for building pages around real questions, not a spreadsheet’s isolated keyword. Google says AI Mode can fan one complex question out into several related searches across subtopics and data sources. In practical terms, a page can be eligible through more than one route. The page that explains the decision properly has more ways to be useful than the page that repeats one head term.
What ARR could mean for AEO
Answer engine optimisation is often used as shorthand for making a source easy to retrieve and cite in a direct answer. ARR does not create a separate AEO rulebook, and no one should pretend that this paper does.
It does reinforce the discipline behind good answer content. Put the answer in the page. State who or what it applies to. Explain any important exception. Show the source, method, date or first-hand experience that lets a reader judge the claim. Then make the next question easy to find with useful headings and links.
That is not writing for a machine at the expense of a reader. It is what a good human answer looks like as well. A page that makes a reader hunt for the actual conclusion is hard for a search system to use responsibly. A clear claim without support is just as weak. The useful unit is a complete, attributable answer.
Structured data still has a place where Google supports it, especially when it accurately describes visible content. It is not a substitute for the content itself, and ARR gives no reason to expect a new technical file or a magic schema property to unlock AI answers. Our earlier guide to what makes a page legible to AI search covers the less glamorous work that actually makes a difference.
What it could mean for GEO
GEO, or generative engine optimisation, is the broader problem of being represented accurately when systems such as Gemini, ChatGPT, Perplexity and Google’s AI features assemble an answer from sources. The word is useful if it reminds us that the answer is not always a ten-blue-links result. It becomes unhelpful when it is sold as a secret discipline separate from evidence, reputation and technical accessibility.
An improved ranker may make source selection more precise, but it will not turn missing information into trustworthy information. A business still needs consistent facts on its own site and across the places people check. It needs proof for claims that matter, current service details, pages that explain what it does in concrete terms, and a reputation that does not contradict the story told on its website.
That is particularly important because an AI mention may be the start of a journey, not the end of it. People can meet a brand in an answer, then look for confirmation in reviews, results pages, maps, social posts and the company site. The click is not always the outcome any more, but a useful result still has to help somebody choose and act.
What to do now
- Keep the technical foundation clean. Make important pages crawlable, indexable and canonical, then link to them from relevant pages in plain language. Those are not old-fashioned tasks. They are how a source becomes available to any retrieval system.
- Write the answer before the sales pitch. For each priority page, ask what question it settles, what conditions change the answer and what evidence lets a reader trust it. If the page cannot do that, more AI vocabulary will not repair it.
- Make important facts consistent and specific. Services, locations, credentials, prices, policies, availability and contact details should agree across the web properties a customer might use to check you.
- Maintain the evidence around the page. First-hand work, case studies with defensible facts, primary sources and clearly dated explanations give a ranker and a reader something better to work with than broad marketing claims.
- Measure presence and outcomes separately. An AI citation, an AI impression, a ranking, a click and a lead are different events. Google’s own reporting has made that distinction more important, not less. Read our note on the Search Console AI impressions report before treating visibility as traffic.
What not to do
Do not rewrite a site around the assumption that ARR has launched. Do not start manufacturing LLM-friendly filler, inventing a new markup strategy, or cutting back on the technical and editorial work that earns inclusion now. The paper is a useful glimpse of where ranking research is going; it is not a set of implementation instructions for publishers.
And do not let a new acronym split the job into three budgets that never speak to one another. SEO makes a page discoverable and relevant. AEO makes a useful answer easy to locate and support. GEO makes sure the brand information an AI may assemble is accurate across the wider web. The overlap is substantial. The common denominator is a clear, evidence-backed source that is technically available to be found.
Frequently asked questions
Is Google using Autoregressive Ranking in Search?
Google has not publicly said that it is. The paper describes research by Google DeepMind-affiliated authors and university collaborators. It does not announce a Google Search deployment or a ranking update.
Does ARR mean traditional SEO no longer matters?
No. The research is about how a system may retrieve and order documents. Pages still need to be crawlable, indexable, relevant and useful to be candidates. Google has also said its AI Mode is rooted in core quality and ranking systems.
Should I create a separate GEO strategy?
Use the term if it helps organise the work, but do not treat it as a collection of secret AI-only tactics. Start with accurate, well-supported brand information, useful pages, strong technical foundations and a realistic way to measure the outcome. Those improvements help across ordinary search and generative answers.
The sources and the sensible conclusion
The primary source is the paper, “Autoregressive Ranking: Bridging the Gap Between Dual and Cross Encoders”, by researchers affiliated with Google DeepMind, the University of Massachusetts Amherst and the University of Texas at Austin. Search Engine Journal’s coverage offers a useful plain-English walkthrough of the architecture. TechWyse’s analysis makes the deployment caveat explicit. Google’s own explanation of AI Mode explains that it uses query fan-out and remains rooted in core quality and ranking systems.
The sensible conclusion is modest. ARR is not a reason to announce an algorithm update or scramble a content plan. It is a reason to stop treating relevance as a keyword-counting exercise. Search systems are moving toward a richer judgment of whether a source actually helps with the question in front of it. The websites that hold up best under that kind of scrutiny are usually the ones that were useful before the model arrived.
