News & analysis
Google Proves It Can Rank Without an Index
Google researchers published a January 2026 paper proving a single language model can rank unlimited documents without a separate search index, building on research the company has published since 2021.
By Roomer Editorial · · 4 min read
What did Google's researchers actually prove?
The paper shows that a dual encoder, the retrieval method search engines use today, needs its embedding size to grow with the number of documents it ranks. An autoregressive model with a fixed hidden size does not have that limit. It can rank an arbitrary number of documents without growing.
The researchers also introduced a training method called SToICaL, designed to make the model care about the full ranked list rather than only the top result. Search Engine Journal's Roger Montti covered the mechanics of the underlying encoders; this paper supplies the theory behind replacing them.
How is this different from search today?
Current search runs two stages: a cheap dual encoder pulls a broad set of candidates, then a slower cross encoder reranks them. Google's proposal collapses both into one generative model.
In that model, a document is not a URL. It is a docID, a short token sequence the model generates the way it generates words. In the paper's shopping test, each product's identifier was derived from its title, compressed into three numbers, so similar products share leading digits. The model ranks by generating the most probable identifier through beam search. Nothing is looked up. The address is produced.
This traces back to 2021's "Rethinking Search" paper, which argued a search system should answer directly from a corpus rather than hand back a list of links. The 2022 Differentiable Search Index showed a single Transformer could map queries straight to document identifiers. The new paper is the theoretical proof connecting those two threads.
What are the limits of this research right now?
The experiments are small. For fine-tuning, the researchers picked Mistral-7B, a capable open model but not Google's most advanced one. To test ranking, they fed candidate identifiers directly into the prompt and let the model choose among them, which reranks a known set rather than retrieving from an open corpus. One test ran on WordNet, a database of roughly 82,000 English noun concepts. The shopping test covered just 310 product examples. In one setup, the model got better at ordering the lower results but lost some accuracy at picking the single best one.
Google's own follow-up research, DSI++, found that continually adding new documents to this kind of model caused forgetting of documents already learned. Mitigations exist, but freshness becomes a training problem rather than a crawl-and-insert problem.
Nobody has replaced a working web index with this approach. The paper proves capacity and demonstrates a training method, not a shipped product.
Why would this change the search results page?
Beam search works by keeping only a fixed number of the most probable partial results at each generation step, then extending those and discarding the rest. The finished list can never be longer than the number kept along the way.
A dual encoder with a nearest-neighbor index produces a top-1,000 list almost for free, because scoring is cheap. A beam of ten produces ten results and stops. Position eleven is not ranked poorly. It is never generated.
Position eleven is not ranked poorly. It is never generated.
That makes page two not a design choice but something that stops existing in this architecture. From outside the system, there is no way to tell "ranked low" from "never produced," because both look the same to the person asking the question.
The paper's own results show the improvement landing in the part of the list users are least likely to see. On the WordNet test, the new training approach lifted recall for the second through fifth positions from roughly fifty percent into the mid-nineties, even as the visible surface shrinks toward fewer results overall.
Alphabet reported Google Search and other revenue at $63.27 billion in Q2, up 17%, following 19% growth the prior quarter. The paper does not touch the ad auction directly. Where it matters is cost: removing the separate nearest-neighbor index is a structural cost reduction for running generative results at scale, which weakens one of the remaining reasons to keep ten blue links as the default surface.
Roomer Editorial's take
This is a five-year-old Google bet getting its proof, not a live product. Treat it as a signal about direction, not a deadline.
The operator takeaway: ranking position is already becoming a worse way to think about visibility than "am I in the set that gets generated at all." A travel creator posting Lisbon hotel reels doesn't control Google's architecture, but control over the underlying page does the same job either way — a page that is distinct, specific, and clearly yours is easier for any ranking system, generative or not, to identify as worth including.
That's the argument for owning the page itself rather than renting a profile. On Roomer, that means a site built from your own videos, with bookable pages and commissions that stay on your site rather than a platform's. If search results compress toward a handful of generated answers, the pages that survive are the ones with a clear, singular identity to point to. Thin or duplicated content has the most to lose under this kind of system, proof or no proof.
Action: audit your own travel content for duplication against your own back catalog and competitors before worrying about where a generative ranker might place you. Distinctness is the one variable this research says will matter more, not less.
Source: Search Engine Journal
