What "Ranking Without an Index" Means for Your Business (2026)

There is a sentence buried in the search-industry news this month that sounds like a math-department footnote and is actually a warning shot for every business that depends on being found online. A Search Engine Journal column by Duane Forrester walks through a January 2026 research paper and lands on a blunt conclusion: AI can now rank documents without a traditional search index, and when it does, the familiar results page — ten blue links, a page two, a slow climb from position eleven to position three — stops existing.¹

I want to be honest with you up front, because that is the voice we use here. This is a research paper, not a product that rewrote your Google rankings last night. Nobody in Fort Wayne lost their Map Pack spot because of a math proof. But the direction it points is real, and it rhymes with changes you can already feel — the zero-click answers, the AI Overviews, the ChatGPT recommendations that name a competitor instead of you. If you understand the mechanism now, you make calmer, cheaper decisions for the next two years. If you don’t, you’ll keep paying for rank reports that may soon measure nothing.

Key Takeaways

• A January 2026 paper proved a single language model can rank an unlimited number of documents without a separate search index — by generating results, not retrieving them.
• The practical consequence a Fort Wayne owner should care about: AI visibility becomes binary. You are either inside the generated set of results or you are invisible — there is no “page two” to climb from.
• Two near-identical pages no longer take adjacent spots; they compete for the same slot, and one loses outright. Distinctness becomes the thing you optimize.
• This is a research trajectory, not a live Google product. Don’t panic-buy anything because of it.
• The durable moves are the ones we already recommend: be genuinely distinct, earn real citations, keep your structured data clean, and stop grading yourself on a rank-tracker number.

What does “ranking without an index” actually mean?

Traditional search works in two steps. A crawler builds an index — a giant, pre-sorted filing cabinet of the web — and then, when you search, the engine retrieves matching documents from that cabinet and sorts them into a list. Position one, position two, all the way down. Your SEO work has always been, at bottom, a fight to move up inside that retrieved list.

Generative retrieval throws out the filing cabinet. Instead of looking documents up, the model generates the identity of the documents that should appear, token by token, the same way it generates a sentence. This isn’t a brand-new idea — Forrester traces it back through a line of research, including a 2021 Google Research paper on “rethinking search” that argued a model could answer directly from a corpus,⁴ and the 2022 Differentiable Search Index paper, which showed a single Transformer could map a query straight to a document identifier with no external index at all.³ What changed in 2026 is that the math caught up to the ambition.

If you already read our take on how your website’s job is changing in an AI-search world, this is the engine under that hood. The website stops being a destination the engine sends traffic to and becomes a source the engine reads from — and, increasingly, regenerates.

What is SToICaL, and what did the researchers actually prove?

The January 2026 paper is titled Autoregressive Ranking: Bridging the Gap Between Dual and Cross Encoders, and its contribution is a training method the authors call SToICaL — a “Simple Token-Item Calibrated Loss,” described as a generalized rank-aware training loss for fine-tuning a language model.² In plain English: it teaches the model to care about the whole ranked list, not just whether it got the single top answer right.

What makes this notable is less the clever name and more the honesty of the experiment. The researchers did not use a frontier mega-model; they fine-tuned Mistral-7B, a comparatively small open model. They tested it on two datasets — a WordNet task covering roughly 82,000 noun concepts, and a shopping-query dataset — and measured how well the rank-aware training pulled the right items into the top handful of results.²

Two cautions worth stating plainly. First, a result on WordNet and a shopping dataset is a controlled lab finding, not proof that live commercial search behaves this way tomorrow. Second, the recall jump the paper reports is an in-experiment measurement, not a promise about Google’s production system. Treat it as a strong signal that the approach works, not as a forecast of your rankings.

Why does Forrester say “page two is gone”?

This is the part that matters for a business owner, and it hinges on one word: beam.

When a model generates results, it uses beam search — it keeps a fixed number of the most probable candidate sequences at each step and discards the rest. Forrester’s argument is that this fixed set is the results page. As he puts it, “the beam width is the complete set of results the system computes,” and “Position eleven is not a bad result. It was never generated.”¹

Sit with that. In classic search, position eleven exists — it’s just on page two, and a determined owner (or a determined agency) can grind it up to page one over months. In a generative world, there is no page two. The model computed a short list and stopped. If your page wasn’t in the beam, it wasn’t ranked low; it was never a candidate. Visibility stops being a slider you can nudge and becomes a light switch: on or off.

That reframes a lot of the anxiety we hear from Fort Wayne owners about AI Overviews eating their traffic. The problem was never only that fewer people click. It’s that the shape of the result is collapsing from a long list you can fight your way up into a tiny generated set you’re either in or out of.

Does rank tracking still measure anything?

Forrester’s sharpest line is that “rank tracking below the beam measures nothing, because there is nothing below the beam to measure.”¹ If he’s right, the monthly rank report — the one plenty of agencies still charge for — is measuring the depth of a well that’s being filled in.

Here’s where we owe you full disclosure, because Forrester does the same in his piece: he founded CitationIQ, a tool that measures AI visibility rather than keyword rank. A person who sells the replacement for rank tracking has an obvious interest in declaring rank tracking dead. That doesn’t make him wrong — the mechanism he’s describing is real research — but it means you should treat “rank tracking is obsolete” as his argument, not a settled fact, and weigh it accordingly.

Our honest, in-the-field read: keyword rank still tells you something useful today, because classic search still drives real calls for Fort Wayne businesses right now. But it’s a shrinking slice of the picture, and if you’re paying for a rank report and nothing else, you’re already measuring the wrong thing. This is the same budget shift we walked through in the move from SEO to GEO demand: the dollars don’t disappear, they move toward being chosen by AI rather than ranked by an index.

Why do near-duplicate pages now lose outright?

There’s a second consequence that’s easy to miss and expensive to ignore. Forrester frames it this way: “two near-duplicate product pages are no longer competing for adjacent positions. They are competing for the same address, and one of them loses outright.”¹

In old-school SEO, a little redundancy was survivable — two similar service pages might take spots four and five and both catch some clicks. In a generative system that outputs one canonical answer per idea, two near-identical pages collapse into one slot, and only one of them gets generated. The other doesn’t rank lower. It evaporates.

This is why we keep telling local businesses to stop spinning up five thin, nearly identical pages — one for “HVAC repair Fort Wayne,” one for “AC repair Fort Wayne,” one for “furnace repair Fort Wayne” — that all say the same thing. In the index era that was merely inefficient. In the generative era it’s self-sabotage: your pages compete against each other for a single slot. Distinctness — genuinely different, genuinely useful pages — becomes the whole game. That’s also why earning real citations from places AI already trusts is quietly becoming more valuable than cranking out more pages.

What should a Fort Wayne business actually do now?

Let me translate the abstraction into Northeast Indiana terms, because that’s the part that pays your bills. If you run an HVAC company in Allen County, a dental office in Auburn, or a restaurant downtown, here’s the honest guidance.

First, what not to do: don’t panic, and don’t let anyone sell you an “index-free-ranking optimization package.” There isn’t one. This is a research trajectory that may shape how AI search works over the next couple of years, not a switch Google flipped this quarter. Your Google Business Profile, your reviews, and your current local SEO still matter this week.

What to actually do is unglamorous and durable, and it’s the same list we give in The Fort Wayne Small Business Guide to AI: be genuinely distinct instead of publishing near-duplicate pages; earn real mentions and citations from sources AI already reads; keep your entity data and structured data clean so a model can identify you unambiguously; and stop grading your marketing on a rank-tracker number that may soon measure nothing. The businesses that win the “beam” won’t be the ones who gamed a list — there won’t be a list to game. They’ll be the ones a model can confidently identify as the distinct, well-cited answer for a specific Fort Wayne need.