TL;DR
- Can a business that arrives late still get cited by AI, or is the seat taken? A 2002 book on the web's link structure describes how links come to concentrate on a few pages, and two 2026 preprints show similar pieces in AI answers.
- In Linked (2002), physicist Albert-László Barabási reported that a web page with only one inbound link had less than a 10 percent chance of being indexed by any search engine, while pages with 21 to 100 inbound links were indexed close to 90 percent of the time. Those are his figures, a quarter century old.
- His explanation is two rules. Networks grow, and each newcomer links to what is already linked: a page with twice the links draws about twice the new ones. That is preferential attachment, the rich get richer.
- In a 2026 preprint from the University of Toronto (US commercial questions, small API models, associations only), ChatGPT wrote a brand its model already knew into its own searches 67 percent of the time, against 11 percent otherwise. In another, on scientific papers and not businesses, eleven models gave the top tenth of candidates 23.3 to 30.2 percent of their citations, against 15.6 percent if they had no preference.
- The book also names the way in for a late arrival: fitness, an advantage that does not depend on the links a node already has. By its account most networks are "fit get rich", not winner-takes-all. Reading AI citations that way is our interpretation, not a finding of either paper.
A question we get from Quebec business owners goes like this: a competitor who has been around longer keeps getting named by ChatGPT, so is the seat taken? Our post on why AI keeps citing the same sources covered an experiment that measured how concentrated AI citations are. This one is about how a concentration like that gets built, and what it leaves open. The mechanism was described 24 years ago, for the web's links, and it comes with a second half that most rich-get-richer summaries drop.
A word on the sources first. Linked: The New Science of Networks (Perseus, 2002) is a popular-science book written by one of the researchers behind the results it describes. Its figures are quoted here as the author prints them, without error bars, and its web numbers date from 1999 to 2002. We have not checked them against the underlying papers. The two 2026 papers are arXiv preprints. The link between the book and the papers is ours, and we flag it where it carries weight.
Why did a page with one inbound link barely exist in 2002?
In 2002 a web page with one inbound link was close to invisible, because search engines found pages by following links and, by Barabási's figures, almost all the links pointed at a few pages. His group mapped the University of Notre Dame's domain: of 325,000 pages, 82 percent had three or fewer inbound links, while about 42 had more than a thousand. In a later sample of 203 million pages, which the book's notes credit to Broder and colleagues, 90 percent had ten or fewer inbound links and about three pages were referenced by close to a million others (pp. 57-58).
A crawler that follows links inherits that skew. The book puts numbers on it:
| Inbound links pointing to a page | Chance of being indexed by any search engine, as the book reports it |
|---|---|
| 1 | Under 10% |
| 21 to 100 | Close to 90% |
His conclusion, one page later: "Your ability to find my Webpage is determined by one factor only: its position on the Web" (p. 175). A page was visible because of where it sat in the link graph, whatever it said.
We have no current version of that figure, and our own data cannot supply one. When we inspected every URL our managed sites submit to Google in September 2026, 77.6 percent of the 549 pages offered were indexed, but those are sitemap-submitted pages on seven Search Console properties we manage, not a sample of the web, and the count is not split by inbound links. Our reading, argued in the rest of this post, is that a similar skew now shows up one layer up, in which pages get cited in an AI answer.
What is preferential attachment?
Preferential attachment is the rule that a new link is more likely to go to a node that already has many links. In the book's words: "When choosing between two pages, one with twice as many links as the other, about twice as many people link to the more connected page" (p. 85).
The reasoning is ordinary. Nobody picks a news site at random. People link to the outlets they know, and the outlets they know are the ones everybody already links to (p. 85). Barabási's model pairs that preference with a second rule, growth: a network is assembled one node at a time, so the early nodes have had longer to collect links (p. 86). In his account either rule alone fails to produce hubs (p. 88). Together they produce a few very large ones and a long tail of nodes with one or two links.
What the model does not need is a hub that is better than the rest. In the plain version, a hub is ahead because it was early, and its lead feeds itself. The book borrows Hollywood's version of the trap: "You need to be known to get good roles, but you need good roles in order to be known" (p. 85).
Is the same loop running in AI answers?
Partly. Two 2026 preprints show the two halves of a rich-get-richer loop in AI answers, a preference for what the model already knows and a concentration of citations. Neither shows the loop closing, where a citation today makes one tomorrow more likely.
The first is by Mahe Chen and Nick Koudas of the University of Toronto (arXiv 2610.15033, version 1, 10 October 2026). They collected 90 days of answers from ChatGPT, Gemini and Perplexity to commercial questions in two markets, US home internet (651 questions) and proposal software for architecture, engineering and construction firms (55 questions), 468,037 citations in all. They then replayed the questions through the ChatGPT and Gemini APIs to see the searches those engines ran (Perplexity could not be replayed). Separately, with search switched off, they asked each engine's model which brands and websites it would recommend. Call that list the model's prior: what it already knows before it looks anything up.
The prior showed up at every step they could observe. When a brand was in the prior for a question, ChatGPT wrote that brand into its own search queries 67 percent of the time, against 11 percent otherwise. Websites in the prior had 1.60 times the odds of being cited at equal search rank. And 47 percent of ChatGPT's citations went to websites its model names without searching, against 25 percent when the comparison used the prior for a different question. Retrieval gates all of it: every one of the 22,315 pages ChatGPT cited in the replay was a page its own search had returned.
Put those together and the pattern points the way the book's does, as an association and not as a demonstrated cause. What the model already knows goes with what it searches for, and what it searches for bounds what it can cite. If the authors' hypothesis holds, the known name gets the next citation partly because it is the known name.
The second preprint is the one covered in our post on citation concentration (Alemohammad et al., arXiv 2608.19230). With every candidate equally visible and stripped of status signals, eleven models from three vendors still gave the top tenth of candidates 23.3 to 30.2 percent of their citations, against 15.6 percent under indifferent selection. Its subject is scientific papers on a single topic, not businesses.
Here is where the comparison is weakest. The book's law needs a loop: a link today makes a link tomorrow more likely. For AI answers that would mean being cited today feeds what the next model knows about you. Neither paper measures that. Chen and Koudas call their chain from prior to retrieval a hypothesis and report associations, not effects. The other experiment never lets a citation change what gets shown later. And a model's memory is not a count of links. So this is an analogy with evidence on each side of it, not a demonstration that AI citations follow Barabási's model.
Can a late arrival still get cited?
Probably, on our reading. In Barabási's 2002 model of the web, a late arrival overtakes older nodes through fitness, and the 2026 preprints point to two things that may play that part for AI citations, though neither tested his model. Pure preferential attachment predicts that the oldest node always wins. The book's counterexample is Google: launched in 1997 against search engines that had held the market for years, it became the biggest node in under three years (p. 94).
The fix his group proposed is to give every node a fitness, an intrinsic ability to compete for links, and to make attachment depend on fitness multiplied by links. Then, "between two nodes with the same number of links, the fitter one acquires links more quickly" (p. 96). At equal fitness the older node still has the advantage. A fitter late arrival passes less fit incumbents whatever its age (p. 97).
Fitness in the book is whatever makes a node worth linking to, "content for Websites" in his phrase (p. 95). Two results in the 2026 preprints look like candidates for that role in AI citations. Calling them fitness is our reading.
- A page that answers the question that was asked. Comparing pages of the same website, Chen and Koudas found that only match to the questions held up across their five robustness tests. A clearly better match between a page's best paragraph and the questions (one standard deviation, the paper's unit) went with 15.5 percent more citations, and the same improvement in title match with 14.1 percent. The title effect showed for Gemini and Perplexity, not for ChatGPT, which rewrites the question before searching. FAQ blocks, schema.org markup, summary boxes and author boxes looked effective across websites (7.5 to 20.9 percent more citations) and showed no significant effect within a website. The rewrites most GEO pitches sell had their own test, covered in GEO tricks, measured. The citation-concentration paper is compatible on one narrower point: paraphrasing left the models' preferences unchanged, and changing the meaning moved them. It does not show which changes of meaning help.
- For a business the model does not know, a cited third-party page that names it. In the niche software market, ChatGPT's model named the tracked vendors for 2 to 16 percent of questions with search off, against 61 to 72 percent for the large internet providers. Gemini's model named the same vendors for 60 to 96 percent of questions, so "unknown" describes one engine's model here. In that market, a cited third-party page that named a vendor went with 39.7 times the odds that ChatGPT's answer named it, and 17.7 times for Gemini, against 4.2 and 6.3 times for the large providers.
One more result belongs here, because the book is about links. In the Toronto data, pages with more referring websites were cited more often (22.7 percent more per standard deviation, within a website), mostly on Gemini (33.7 percent, against 3.7 percent for ChatGPT). The links were counted after the citations, so the direction is not established.
The scope on those numbers is narrow. They are associations, measured on English questions about US markets, through the small API models behind ChatGPT and Gemini and not the consumer apps. The cited pages also kept changing: between two two-week windows ten weeks apart, ChatGPT's citation counts shared only 14 percent of their within-website variation. Nothing in the paper was measured on French-language Quebec questions.
Even so, the niche market as ChatGPT's model sees it is the one that resembles a Quebec trade or professional service. When we tested what models know about Quebec roofing with search off in July 2026, 7.1 percent of responses (5 of 70) named a local business we could verify, and 69 percent named no concrete business at all. The gaps were filled by national manufacturer brands and lead-generation marketplaces. On that evidence the local seat in a market like that one is mostly unclaimed, which is a different problem from a seat a competitor holds. It was one trade, twelve French prompts and nine models: a probe, not a study.
Is AI search winner-takes-all?
Nothing in the two 2026 preprints looks like winner-takes-all in AI citations, and the 2002 book treats that outcome as the exception. Barabási sorts competitive networks into two kinds (pp. 102-103).
- Fit get rich. The fittest node becomes the biggest hub, but the book adds that "The winner's lead is never significant" (p. 103): the largest hub is closely followed by a smaller one, and a hierarchy of hubs runs below them. He puts the Web, the Internet and Hollywood in this group (p. 106), and says of Google that its rise "is not an indication of winner-takes-all behavior; it only tells us that the fit get rich" (p. 103).
- Winner takes all. One node takes nearly every link and the hierarchy collapses into a star, with "no room for a potential challenger" (p. 103). His one real-world example is the operating-system market of his day (p. 105).
AI citations in the 2026 data look like the first kind. The top tenth of candidates takes under a third of the citations, which is concentrated and far from all of them. In the Toronto data, 13 percent of the variation in citation counts sat between websites and 87 percent between pages of the same website, and the cited pages kept turning over. Neither paper tests which of Barabási's two kinds applies, so that classification is ours. One result cuts the other way: in the concentration paper, the models' preference tightened onto fewer targets as machine-written papers joined the pool. Treat "fit get rich" as a description of the 2026 data, not a forecast.
For a business deciding whether AI visibility is worth the effort, the practical point is that the job is not to unseat the incumbent. In the book's fit-get-rich networks the largest hub is closely followed by others, so nodes rise without the leader collapsing. If AI citations behave that way, which is our reading and not a measured result, getting cited, and then cited more often, does not require the leader to lose. Nobody can promise you the top seat.
What does this not tell you?
This post does not show that AI citations follow Barabási's model, and neither preprint measured French-language Quebec questions or Google's AI Overviews.
- That the web still works the way the book describes. Linked is a 2002 account. Later research has examined how universal its scale-free pattern is. We have not reviewed that work, and the argument here needs only heavy concentration and the two kinds of competition, not an exact curve.
- That AI citations obey the book's model. The feedback from citation to memory is not measured in either paper.
- Anything about French-language Quebec queries, or about Google's AI Overviews and AI Mode. Neither was in the two preprints' data. Our own roofing probe used French prompts but ran with search off, so it measured memory, not citations. What Google itself documents about those surfaces is in how AI chooses what to cite.
- How large any effect will be in your market. The 2026 figures are US and English snapshots of engines that change continuously.
What we do with it
Here is what changes in how we scope GEO work. None of it guarantees a result.
- Check whether there is a hub at all. Ask the models, with search off, who they name for the questions your buyers ask. If nobody local comes back, you are closer to the niche case, where a position came from evidence the engine could retrieve.
- Work on the two things that were measured. One page per real question, with a title in the question's own terms and a paragraph that answers it in full. Then find the third-party pages the engines already cite for those questions and earn a mention on them.
- Measure again on a schedule. The cited pages turn over within weeks, so one audit score cannot track them. Our AI visibility funnel lays out what to measure at each stage, and the baseline count is part of our SEO audits.
The 2002 claim was that a page almost nothing points to was close to invisible to a search engine. The 2026 data suggests something similar for a business the model does not already know: in the niche market the Toronto authors studied, it was far more likely to be named when a cited third-party page named it.