Recommended Free Tools
iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
No. A January 2026 paper gives a formal argument that a certain kind of language-model ranker can express more rankings than a dual encoder of the same size. That is a real and interesting result, but it is a theoretical claim plus experiments on two datasets. It does not show that Google Search has dropped its index, that autoregressive ranking runs in Google’s web search, or that results pages are going away. Google’s own public description of Search still combines ranking and retrieval systems, and retrieval includes scanning an index.
What the 2026 paper actually claims
The paper is “Autoregressive Ranking: Bridging the Gap Between Dual and Cross Encoders” by Benjamin Rozonoyer et al., arXiv identifier 2601.05588. It was submitted on January 9, 2026 and revised to version 4 on February 11, 2026. The abstract describes autoregressive ranking (ARR) as a method that generates document identifiers token by token and then ranks the candidates it produces with beam search.
The three model families being compared
Ranking systems that handle large collections usually sit on one of three designs, and the paper’s argument turns on how they differ:
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minute- Dual encoders (DEs) map a query and each document into separate vectors and score them by similarity. Because documents can be embedded ahead of time, this scales well, which is why it is common for first-stage retrieval.
- Cross encoders read the query and a document together. They tend to be more accurate for a given pair, but they must run once per candidate pair, which is expensive at corpus scale.
- Autoregressive ranking (ARR) lets a language model write out the identifier of a document one token at a time, with beam search keeping several high-scoring candidates alive. Ranking comes out of the generation process itself rather than from a precomputed vector comparison.
The capacity result
The paper’s main claim is that, within its formal model, a dual encoder needs an embedding dimension that grows linearly with corpus size to express an arbitrary ranking over that corpus. The authors prove that ARR can reach the same expressive capacity with a constant hidden dimension. Their summary sentence is: “In this paper, we first prove that the expressive capacity of ARR is strictly superior to DEs.” This is a statement about what the model class can represent under the paper’s assumptions. It is not a measurement of how well any deployed system ranks pages.
#1 Best Overall
The training loss and the experiments
To make ARR work in practice, the authors introduce SToICaL, described in the abstract as “a generalized rank-aware training loss for LLM finetuning.” The experiments use WordNet and ESCI. The abstract reports improved ranking metrics beyond top-1 retrieval, meaning the quality of the whole ranked list and not only whether the first answer is correct. Those datasets are small, controlled benchmarks. The abstract reports no web-scale deployment statistic, latency figure, or index-size comparison.
What the paper does not establish
- It does not describe Google’s systems. The paper does not state a Google affiliation, and nothing in its abstract links it to Search.
- It does not remove the need for an index. Its expressivity argument concerns what a model can represent. How a production system would store, refresh, and serve a web-scale corpus is outside the abstract.
- It does not measure freshness. Questions about how quickly new pages become rankable are not addressed in the abstract.
- Its formal assumptions limit the claim. A proof about capacity under stated assumptions does not say that a trained model reaches that capacity on real queries.
What Google has said publicly about how Search works
Google’s public explainer “How AI powers great search results,” by Pandu Nayak, Google Fellow and Vice President of Search, is dated February 3, 2022. It says: “Search runs on hundreds of algorithms and machine learning models, and we’re able to improve it when our systems — new and old — can play well together.” The post describes RankBrain as a ranking system, neural matching as helping retrieve relevant documents from the index, and BERT as contributing to both ranking and retrieval.
That post is more than four years old and is not a full technical specification. It is still the clearest official description available, and it describes an index-based system. Any claim that Google has publicly announced index-free Search as its current architecture is not supported by it.
The research line behind the headline
The 2026 paper is one step in a sequence of papers about replacing or reshaping the classic retrieve-then-rank pipeline. Each step is a research result, not a product announcement.
2021: Rethinking Search
“Rethinking Search: Making Domain Experts out of Dilettantes,” by Donald Metzler et al. and published in SIGIR Forum in 2021, proposed combining information retrieval with pretrained language models so that a system could answer an information need directly. The paper also named problems of language models at the time, including hallucination and the lack of document-grounded justification for answers. It was a proposal, not a product roadmap.
2022: Differentiable Search Index
“Transformer Memory as a Differentiable Search Index” by Yi Tay et al., presented at NeurIPS 2022, showed a single Transformer that stores information about a corpus in its parameters and maps query strings to document IDs. The paper demonstrated the approach in an experimental setting. It did not show deployment at web scale.
Rank #4
- New
- Mint Condition
- Dispatch same day for order received before 12 noon
- Guaranteed packaging
- No quibbles returns
2023: Generative retrieval at scale
A Google Research EMNLP 2023 study by Ronak Pradeep et al., “Understanding Generative Retrieval at Scale,” tested generative retrieval on the MS MARCO passage ranking task, which contains 8.8 million passages, and considered models up to 11 billion parameters. The authors reported that, with existing techniques, scaling model parameters beyond a point could hurt retrieval effectiveness, and they called for more fundamental improvements. These figures describe the size of that experiment. They do not mean that the whole live web is represented in model weights.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
2026: Autoregressive ranking
The ARR paper adds a formal expressivity result and a rank-aware fine-tuning loss, tested on WordNet and ESCI. It is the most direct theoretical argument so far for why generative ranking might scale differently from dual encoders, but it is still a paper.
September 2026: Retrieve-for-Train
A Google Research blog post by Pengcheng Jiang and Judith Yue Li, dated September 15, 2026, describes Retrieve-for-Train, a framework for generating complementary result slates for specialized retrieval tasks. It reports tests in fashion and music retrieval settings and efficiency figures for a diffusion retriever in those experiments. It is a separate method with different use cases from ARR, and the post does not say it replaces Search’s index or results page.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Where “the results page does not survive it” comes from
The headline’s second half is a forecast from Duane Forrester’s Search Engine Journal article, which carries the same title. Forrester argues that the beam width of a generative ranker could become a visibility boundary, and he discusses future SEO measurement, click effects, and changing ad surfaces. The article itself says that no web index has been replaced and labels its forward-looking sections as informed speculation. It also discloses that Forrester founded CitationIQ, a platform for AI data and visibility measurement. Treat those predictions as one analyst’s view, not as findings of the paper or as measured outcomes.
Comparing the two architectures
The table below compares the conventional index-based design, as described in Google’s explainer, with autoregressive and generative ranking, as described in the papers above. Where the sources do not establish a value, the cell says so.
| Axis | Index-based retrieve-then-rank (Google’s public description) | Autoregressive or generative ranking (as published) |
|---|---|---|
| Candidate generation vs. ranking | Retrieval systems, including neural matching, find candidates; ranking systems order them (Google explainer, 2022) | The model generates document IDs and ranks them with beam search (ARR abstract, 2026) |
| External index used | Yes; the explainer says neural matching scans the index | Not stated for web search; the ARR abstract does not address serving infrastructure |
| Corpus scale | Web scale in deployed Search | Experiments on MS MARCO with 8.8 million passages (Google Research EMNLP 2023 study); ARR experiments on WordNet and ESCI only; web-scale behavior not stated |
| Freshness and updates | Not detailed in the 2022 explainer | Not stated in the ARR abstract or the generative retrieval study |
| Computational cost and latency | Not quantified in the 2022 explainer | Not stated for ARR at production scale |
| Ranked-list quality beyond the first result | Not quantified in the 2022 explainer | ARR abstract reports improved ranking metrics beyond top-1 retrieval on WordNet and ESCI |
| Evidence maturity | Deployed in Google Search | Published experiments; no deployment described in the sources reviewed |
How to check a claim like this yourself
- Open the arXiv abstract page for 2601.05588 and confirm the version history: version 1 on January 9, 2026 and version 4 on February 11, 2026. Later revisions can change the claims.
- Look for the phrase “production,” “deployed,” or a named product in the paper. If none appears, the paper is describing a method, not a service.
- Read the experimental section to see which datasets and metrics were used. Two benchmarks are not evidence about the open web.
- Check Google’s official Search documentation and announcements for any mention of autoregressive or index-free ranking. Absence of such an announcement is not proof, but a claim of deployment should cite one.
- Separate three questions: whether a model can represent a ranking, whether it can be trained to produce good rankings, and whether it can serve billions of pages quickly and up to date. The 2026 paper speaks clearly to the first and only partly to the second.
Readers who want to follow the topic can also track each new paper’s stated scope. The distance between a formal capacity result and a live search product is where most headlines about “replacing the index” go wrong.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
