Backlinks and AI visibility: what the link still buys, and what corroboration wins without it

By Simon Vasconcelos Lee Jul 25, 2026

For twenty-five years, the question "who deserves to be shown" had an accounting answer: links. Search engines could not read, so they counted votes. Generative engines read. A page that names you without linking to you teaches them who you are, what you do, and whether the rest of the web agrees. The consequence is not the death of the backlink: it is its demotion, from score to infrastructure. This study crosses the correlations published between 2024 and 2026, the engines' documented architectures, the academic literature and our own surveys to answer the question that GEO, generative engine optimization, has made concrete: is an unlinked corroboration enough to exist in AI answers, and what is left for the link?

The short answer fits in two sentences. Where the engine answers from memory or chooses its words, the corroborated mention is the currency, and the link plays no documented role. Where the engine retrieves candidate sources, the link still holds doors, real but unequal across engines, and loosening quarter after quarter.

What a backlink always measured

The link was never the thing itself. It was a proxy: a machine that could not understand a page needed a computable signal to estimate trust, and the hyperlink, a public vote, costly to obtain, left by a third party, was the best candidate available. PageRank industrialized that reasoning, and an entire economy grew on top of it: directories, exchanges, purchases, penalties, disavowals. Asking whether "backlinks still make authority" is asking whether that proxy still measures something today's machines cannot read directly.

The devaluation did not start with AI assistants: it started at the registry keeper's own desk. In March 2024, Google rewrote its documentation, links going from an "important factor" to "a factor" in determining relevancy, the word important dropping out of the text. A month later, an engineer on the search team told a conference audience: "we need very few links to rank pages". The May 2024 leak of internal documentation confirmed that link and site-authority attributes still exist in the machinery, without revealing their weights. And back in 2014, a patent granted to Google formalized "implied links": a reference to a target resource that is not an express link, in other words the mention, already counted as a vote. The idea that an unlinked citation carries authority predates generative engines by a decade: they merely made it operational at scale.

Where an AI answer is decided: three places, three currencies

To answer the backlink question properly, you have to decompose what it claims to influence. An AI answer is decided in three places, and the currency that circulates is not the same in the three. We mapped the full machinery in our study of retrieval pipelines; what follows focuses on the link's place in each.

The model's memory. A share of answers consults nothing: the model writes from what it learned in training. What a model knows about an entity is a measured function of co-occurrence frequency in its corpus: the reference study, published at ICML 2023 by Kandpal and co-authors, shows a large model's factual accuracy climbing from 25% to above 55% as the number of documents where the entity and the fact co-occur goes from a few dozen to ten thousand, and the relationship is causal, deleting those documents from the corpus makes accuracy collapse on the affected questions. No link graph enters that mechanism: what a model learns from a page is its text, not its HTML attributes. A linked mention and an unlinked mention weigh the same there. The link keeps a share of plumbing upstream, in collection: Common Crawl's crawl frontier, the substrate of most open corpora, is prioritized by centrality on the host graph, and the earliest corpora were literally link-curated, OpenAI's first being built from Reddit's outbound links. But modern pipelines filter on content quality, not on links. Put plainly: the link still helps a page get read; it adds nothing to what the page makes the model believe.

Retrieval. When the engine searches, machinery selects a finite set of candidate sources, and this is where the link keeps its most real grip, provided you look door by door. Google states in writing that its AI features are "rooted in our core Search ranking and quality systems" and that eligibility there is classic Search eligibility: indexed, snippet-eligible. Microsoft officially defines a site's credibility, for Bing and therefore for Copilot's answers, by which sites link to it. ChatGPT search opens with a binary door, allowing the OAI-SearchBot crawler, then leans on third-party search providers, historically backed by Bing's index. On the other side, Perplexity documents a self-built index of more than 200 billion URLs, hybrid lexical and semantic retrieval, rerankers, and publishes no definition of authority based on links anywhere; Claude's search rides an independent index. The picture is clean: the link weighs where the door is a classic ranking; it has no documented role on the doors the engines built for themselves.

Writing. Once sources are retrieved, the model chooses what survives into the answer, and that choice has been measured. Faced with conflicting evidence, models side with the majority of retrieved documents, with a spectacular dose-response relationship: in the reference study at ICLR 2024, a large model keeps its initial answer in 3.7% of cases when no evidence supports it, and in 99.8% of cases when all of it does. A second study, at ACL 2024, shows models weighing a document's relevance while largely ignoring authority cosmetics, institutional tone or outward signs of credibility. And the founding GEO experiment, published at KDD 2024 by a team from Princeton and Georgia Tech, established that adding statistics, source citations and quotations to a page raises its share of the generated answer by 28 to 41% in relative terms, when the page is already retrieved, while keyword stuffing pushes it down 9%. The same experiment measures a redistribution: the source ranked fifth in the retrieved set gains 115% visibility when optimized this way, the first loses 30%. A corrective came from NeurIPS 2025: those style gains do not generalize everywhere, three method-domain combinations out of 54 stay significant, and winning retrieval beats styling. But in this layer too, no engine documents any link signal: the final selection turns on content, agreement and position, never on the source's backlink profile.

What the published correlations say

Large-scale observational data tells the same story as the architecture. The widest study to date, published in May 2025 across 75,000 brands, correlated a battery of signals with each brand's presence in AI Overviews answers:

Signal measured (75,000 brands, May 2025)	Spearman correlation with answer mentions
Branded web mentions	0.664
Link anchors carrying the brand name	0.527
Branded search volume	0.392
Domain authority score (the study's proprietary metric)	0.326
Referring domains	0.295
Backlinks	0.218

The mention weighs three times the backlink, and the hierarchy held when the same team replicated the measure in December 2025 across three surfaces, ChatGPT, AI Mode and AI Overviews: branded web mentions correlate at 0.656 to 0.709 there, link metrics are described as very weak, and the most correlated factor everywhere is the YouTube mention, around 0.73, a direct echo of what our own surveys show about videos in AI answers. An independent study from January 2025, on 10,000 questions put to an OpenAI model through its API, had found the same asymmetry from another angle: ranking on Google's first page correlates at 0.65 with the model's mentions, backlinks barely at all, around 0.10, domain authority scarcely better, around 0.25.

Two precautions are due, and they are part of the result. First: all of this is correlational, and brand size confounds everything, a large brand having more mentions, more links and more AI visibility all at once. The only published causal evidence in the field, the KDD 2024 experiment cited above, concerns the page's content, neither links nor external mentions: nobody has yet randomized backlinks against mentions. Second: a November 2025 study, run at page level rather than brand level across 216,524 pages, finds the reverse, referring domains as the best predictor of ChatGPT citations, 8.4 citations on average above 350,000 referring domains against 1.6 to 1.8 below 2,500. The contradiction is instructive rather than embarrassing: at page level, a massive link profile travels with the fame of a big brand, mentioned everywhere besides; at brand level, where mentions and links compete directly, the mention predicts better. The link accompanies authority; the mention precedes it.

One last precaution: treat every number in this section as a dated photograph. Citation mixes move fast, one weekly tracking of 230,000 prompts watched Reddit go from roughly 60% of ChatGPT's answers in early August 2025 to roughly 10% six weeks later. What is measured one quarter gets remeasured the next.

The ranking door is loosening, engine by engine

The link retained one solid indirect channel: rankings. Ranking well, hence being well linked, put you in the source set of engines that ground themselves in classic search. That channel still exists, and it is narrowing in plain sight.

On AI Overviews, the surface most anchored to rankings, the share of citations coming from the organic top 10 went from about 76% in July 2025 to 37% on January 2026 data, measured across 863,000 keywords, with more than a third of citations now coming from pages ranked nowhere in the top 100. Rank still pays there at answer level, 90% of AI Overviews cite at least one top-10 URL and position 1 is cited in 43% of cases across 362,000 queries from October 2025, but the table has widened. Among standalone assistants, anchoring was never the rule: across 15,000 long-tail queries measured in August 2025, only 12% of the URLs cited by ChatGPT, Gemini, Copilot and Perplexity ranked in Google's top 10 for the query, and close to 80% ranked nowhere in the top 100. An academic study from December 2025 across 55,936 queries puts a number on the decoupling: 37% of the domains cited by generative engines never appear in classic engines' results at all.

The conclusion of this series is not that ranking no longer serves: it is that ranking is becoming again what it is, one of the doors, the best documented, the most contested, and less and less the only one.

What our own surveys show

Our trackers put the questions of three markets to the same four engines every week and archive the answers with their sources. Three observations from our data bear directly on the link question.

The memory layer is alive in production, and its size depends on the engine. On a July 20, 2026 measure at identical perimeter, 54 answers per engine across four shared trackers, the share of answers written without any web search is 0% on Perplexity, about 2% on Gemini, 28% on ChatGPT and 44% on Claude. In other words, depending on which engine matters for your market, between none and nearly half of the answers play out on model memory alone, the layer where co-occurrence is the currency and the link has no documented role. How you split your effort between corpus and rankings is not doctrine: it is a reading of your own measurements, engine by engine.

Being read and being credited are two distinct events. Across the 197 Claude answers in our archive as of July 20, the engine retrieved 784 pages and credited only 342; 197 domains out of 394 read were never cited once. Retrieval, where the link can help, does not guarantee credit, which is decided at writing time.

Credit travels through entity knowledge, not through links. In a Perplexity survey from July 17, 2026, a passage from a hosting company's blog is rephrased almost word for word in the answer, and the credit goes to a brand the passage never names, the engine itself writing "(implied)" next to the attribution. The engine connected content to an entity by inference, with no hyperlink involved anywhere. That is the corroboration mechanism caught in the act, in both directions: what others say about you can credit you without linking to you, and what you publish can serve without crediting you.

What an unlinked corroboration accomplishes, and where

A corroboration is a page where a third party retells your story: your name, your facts, in their own words. It is the earned half of your presence, the one you do not write yourself. What the three layers do with it is now clear.

In model memory, it is the unit of learning: every page where your name co-occurs with your facts raises the probability that the model knows them, link or no link. At writing time, it is the unit of consensus: models side with the camp supported by the majority of retrieved documents, and every source repeating your facts grows that camp. And in retrieval, it has a property the backlink never had: it does not need you to rank, it needs its host to be read. A corroboration published on a source an engine already retrieves inherits its host's door. The engines overwhelmingly cite third-party pages rather than brand sites anyway: across a series of 25 million citations recorded through 2025 and 2026 on ChatGPT, Claude and Gemini, 84% point to earned media, coverage by others, rather than to the brand's own properties.

What makes a corroboration effective is documented too: precise facts, dated, liftable as they stand, the same wording from one source to the next. The KDD 2024 experiment measures it on the page, statistics and source citations grow the answer share, and the ICLR 2024 measure shows it across pages, the concordant majority wins. A stable story, repeated by third parties in consistent terms, plays on both boards. That is exactly the canon's role in our method: a reference version of your facts, that your surfaces carry and your corroborations repeat.

Honesty requires drawing the boundary: no published study yet randomizes mention against backlink, and the relative weights of the two remain unmeasured as of July 2026. What the literature establishes is the mechanism: the corroboration acts directly on two of the three layers and inherits the third, where the link acts, indirectly, on only one.

What the link still buys

Nothing above makes the backlink useless; all of it repositions it. It still buys, concretely: discovery, a crawler follows links and crawl frontiers are still prioritized by graph structure; candidacy, on the doors held by classic rankings, Google's AI surfaces by their own documentation, Bing and whatever grounds itself there, the share of ChatGPT search still leaning on it; and human traffic, which remains the oldest reason to want a link. On those doors, ranking work keeps its value, and abandoning it would be reading the transition backwards.

What the link no longer buys is the verdict. It does not choose the answer's words, it does not weigh in the model's memory, it is not documented in any engine's final selection, and the doors where it counts grow fewer as pipelines go proprietary. The practical reversal fits in one sentence: the backlink stops being the objective and becomes again what it originally was, the verifiable by-product of successful coverage. A quality corroboration rarely comes alone; when it carries a link, so much the better, it will serve at the doors that still count votes. But it is the corroboration doing the work, on all layers at once.

The war nerve has moved

Will backlinks remain the war nerve of online visibility? No. The war nerve, in the era of generative engines, is corroborated proof: the same name, the same facts, repeated across pages each engine actually reads. The link remains a useful toll at some doors, and the natural attestation of a corroboration well earned. Here is what this study makes actionable.

Measure your mix before splitting the effort. The share of answers played out in memory runs from 0% to 44% depending on the engine in our surveys, and ranking anchorage varies sixfold across surfaces. Your own series, engine by engine, says where rankings still pay and where only the corpus speaks: that is the instrument, Epovest Tracking, before the doctrine.
Put down a corroborable story. Precise facts, dated, phrased to be lifted as they stand: the only variable whose causal effect on answer share is demonstrated. A stable canon gives third parties something exact to repeat.
Prioritize corroborations by your engines' doors. A mention is worth its host: your measurements show which sources each engine cites on your questions, and your Atlas holds the map and the registry. Aiming at the sources that carry your answers beats aiming at the sites that sell links.
Keep rankings where they hold a door, without reading them as a proxy. Classic search work stays profitable on the surfaces that ground themselves in it; what must go is the idea that a rank predicts a presence in answers, and we documented why that inference no longer holds.
Date your actions, read their effects. A corroboration published, a link earned, a page rebuilt get recorded in your Logbook, and the next series says what moved. That is the gesture that turns a bet into a measurement.

For a quarter of a century, a domain's authority was bought in votes. The engines writing today's answers know how to read: they count agreements. The war is no longer won by accumulating links; it is won by giving the web coherent reasons to say the same thing about you, and by verifying it, answer by answer.