A Web Growing Faster Than Its Indexes
In 1993, fewer than 100 websites inhabited the planet. Within five years, the web had become a far larger and faster-moving place: estimates cited in the 1998 PageRank paper put it at more than 150 million pages, with a doubling life of less than a year. The National Science Foundation records the earlier scale; the PageRank paper’s historical record captures the acceleration that followed.
That growth changed the basic task of finding information. A small web could be approached as a collection to catalogue, and in 1994 two Stanford students started Yahoo! as a manually constructed table of contents for websites. Early search tools were also beginning to crawl across the internet, while services including Lycos and WebCrawler offered new ways into the expanding collection. The NSF’s account of the period places those tools alongside Yahoo!’s directory, each designed for a web whose shape could still seem graspable.
The numbers soon made that premise feel quaint. Every new page added another possible answer to a query, and every query raised the same practical question: when many pages matched the words, which result deserved to appear first? A directory could sort sites into categories, yet a category with a growing mass of pages still left the user facing a choice. Search needed an ordering principle that could work at the scale the web was reaching.
The problem carried unusual stakes because the first result offered more than convenience. It directed attention. As the web multiplied, an index increasingly had to make a judgment about importance before a person could make one for themselves. The early tools had opened routes into the network; the next challenge was to decide how those routes should be ranked.
Science Had Already Invented the Key Move
That ranking problem had already occupied researchers studying a different kind of network: the chain of influence that runs through scholarly work. In 1953, Leo Katz introduced a “status index” for a graph, defining status as the weighted sum of all paths between two nodes and giving shorter paths greater weight. Katz’s formulation supplied a powerful premise for any crowded system: position could emerge from relationships rather than from a simple count. A node gained standing through the paths that led to it, and the structure of the network became part of the judgment.
Scientific citation offered a living example of that premise. Eugene Garfield proposed a citation index for scientific literature in 1955, creating a way to follow the references connecting one piece of research to another. Garfield’s proposal turned the bibliography into something more useful than a record of reading: it could map a field’s internal lines of attention. By 1972, Garfield had stated the essential interpretive step, proposing that citations of a research article could be understood as endorsements of it. That idea made every reference a small act of judgment, and a large collection of references into evidence about which work commanded regard.
Gabriel Pinski and Francis Narin pushed the idea further in 1976 by assessing the influence of scientific journals through their citation relationships. Their method weighted a reference according to the influence of its source, so a citation from a more influential journal carried greater force in the calculation. Their work captured the move that would matter most for ranking a vast network: authority flows through connections, and the authority of the source changes the value of what it points toward.
That produces a recursive model of status. To judge the importance of one node, the system considers the status of the nodes connected to it; to judge those nodes, it follows their own connections in turn. A later analysis of PageRank describes this as measuring a node’s status recursively through both the number of connected nodes and their respective status. In that account, a link from a highly ranked site is worth more than one from a site with low rank. Science had therefore supplied a transferable way to rank an expanding world of documents: let references reveal standing, then let standing shape the weight of every new reference.
When a Link Became a Referral
In that account, a link from a highly ranked site carries greater weight than one from a site with little standing. Larry Page and Sergey Brin brought that recursive logic into the Stanford research environment where Page worked as a graduate student supported by the National Science Foundation’s Digital Library Initiative, while Brin held an NSF Graduate Student Fellowship. The institutional setting mattered because it gave the pair room to treat the web as a connected body of documents whose relationships could be calculated, rather than a collection of pages to be sorted one by one. The National Science Foundation traces both researchers’ support to that wider effort.
Brin supplied a clean organizing thought: information on the web could form a hierarchy through “link popularity.” PageRank’s account of its development records the idea in those terms, and Page and Brin gave it a demanding consequence. A page could earn a quality score from the normalized quality of the pages that linked to it. Every link now carried a judgment, yet the value of that judgment depended on the standing of the page making it.
That made the web’s everyday connective tissue do double duty. A hyperlink took a reader somewhere else, and PageRank treated it as a referral. The model held that a high-quality page should receive referrals from other high-quality pages, a principle documented by the Engineering and Technology History Wiki. Standing could therefore travel through the network: a respected page passed some of its weight onward, and the recipient’s score reflected the quality of the sources willing to point toward it.
The transfer had a built-in discipline. Each page distributed its endorsement ability uniformly across every page it referenced, so a link sat inside a finite allocation of attention rather than serving as an unlimited stamp of approval. The PageRank milestone account describes this rule as the uniform distribution of a page’s endorsement ability among its outgoing references. A page with many outbound links spread its influence more widely; a page with fewer shared its available weight among fewer destinations.
PageRank gave a familiar web action a formal social meaning. Links became evidence of referral, referrals became weighted signals, and each signal drew force from the network behind it. That chain of reasoning turned the web’s own structure into a ranking system, setting up the mathematical question of where its accumulated weight would finally settle.
A Simple Question With a Mathematical Answer
That question had a surprisingly human answer: imagine someone moving through the web by clicking one link after another, with each click chosen at random. PageRank turns that imagined journey into a probability distribution, estimating the likelihood that the person will arrive at any particular page. A page’s score therefore expresses a share of possible attention, as a value between 0 and 1, rather than a simple tally of how many links point its way. PageRank’s formal description gave the web a way to convert its tangled paths into a set of comparable chances.
Those chances change with every link because the starting point matters. A link from a highly ranked site carries greater weight than a link from a low-ranked one, since the first site already sits closer to the places a random visitor is likely to reach. The calculation is recursive: each page’s standing depends on the standing of the pages connected to it, and their standing depends in turn on the wider network. An analysis of the model describes this as a measure of status that accounts for both the number of connected nodes and their respective status.
That recursion gave PageRank its practical force. Search had a vast field of pages that might match the same words, while the link structure supplied an additional signal about which pages had accumulated attention from influential parts of the network. Google summarizes the approach as counting both the number and quality of links to estimate a site’s importance. The explanation captures the essential shift: importance could emerge from the web’s relationships, then be expressed as a mathematical ranking.
PageRank entered a field already alive with related experiments. In 1996, Robin Li designed RankDex, described as the first search engine to use page-ranking and site-scoring algorithms, and he later used its technology in founding Baidu in China. Larry Page cited Li’s work in some U.S. patents for PageRank, a small sign of how quickly ideas about network-based search were circulating. Then, in 1998, Jon Kleinberg published HITS, another influential approach to reading authority through links, in the same year PageRank entered public view. The race to interpret the web’s structure was already underway, and PageRank’s particular answer was ready for a far larger test.
BackRub Tested the Theory Against 24 Million Pages
That ferment gave Page and Brin a clear reason to test whether their ranking method could survive contact with the web at scale. Their prototype was called BackRub, and its equipment came through the Digital Library Initiative alongside industrial contributions, tying a research idea to the physical work of collecting, storing and processing a rapidly expanding network of pages. The National Science Foundation’s account places the project inside the Stanford research effort that supplied the machinery for that test.
BackRub’s task was demanding in a very practical way: the system had to work across a body of material large enough for the web’s patterns to become visible. Page and Brin tested their ranking approach on an initial set of 24 million pages, according to the NSF. That test set turned the web’s link structure into something the prototype could calculate across, page by page, while its full-text and hyperlink database covered at least the same 24 million-page scale. A historical account of the 1998 prototype describes that combined database as the working foundation beneath the experiment.
The size of the test mattered because it made the question of scale immediate. A method that could rank a small collection had little to prove; a system operating across millions of documents had to remain useful as its hardware and data demands grew together. BackRub gave Page and Brin a place to find out. By late 1997, the NSF reports, the approach had proved sound, expandable and popular, three qualities that moved it beyond a promising research result and toward an operating search system.
That shift changed the project’s practical horizon. The work now needed room for a growing hardware facility, and by the end of the Early DLI Age in 1998, Page and Brin had obtained funding to move that facility from Stanford’s campus into a friend’s garage. The NSF’s history records the move as the next step after BackRub’s successful run at scale, with the experiment ready to take a public form.
The Paper That Became Google
That public form arrived on January 29, 1998, when Larry Page, Sergey Brin, Rajeev Motwani and Terry Winograd published “The PageRank Citation Ranking: Bringing Order to the Web.” The title made a practical promise at a moment when the web’s sheer volume had made order a valuable service, and the paper put the ranking method into a record that others could examine. The published paper gave PageRank a name and a clear public identity, carrying the Stanford project beyond its prototype stage.
That identity soon belonged to a company. PageRank formed the basis of the search engine behind Google’s founding in 1998, and by the end of that year Page and Brin had incorporated Google, Inc. the IEEE Engineering and Technology History Wiki records. The garage facility gave the growing operation physical room; the paper gave its central idea a durable form. Together, they turned a research effort into a search engine with a method for deciding what rose to the top.
The method’s commercial consequence followed from that decision. PageRank is widely credited as the key factor that enabled Google to surpass established competitors including AltaVista, Excite and Infoseek, according to the IEEE’s account of the milestone. Search results had become a judgment about relative importance, made at web scale through the network of links between pages, and Google built its position around that judgment.
As Google rose, the underlying approach traveled further. Virtually every search engine adopted a variant of PageRank, the same history notes, which meant the breakthrough extended well beyond one company’s results page. A paper that set out to bring order to the web had supplied a model for how search engines could assign attention across it. That model would soon carry value far beyond the act of searching.
A Ranking System Became a Publishing System
That model carried value far beyond the act of search itself, because PageRank turned the web’s links into a standing system for the pages they connected. In the language of software studies, it functioned as an evaluative metric, a calculative procedure designed to assess importance, quality, relevance and performance, drawing on techniques associated with graph theory and matrix algebra. That framing helps explain why the algorithm mattered beyond the search box: it gave the sprawling act of publishing online a machinery for sorting whose signals counted most.
That machinery worked through a rule readers could feel even when they never saw the calculation. A page accumulated standing through links, and links from highly ranked pages carried greater weight than links from low-ranked ones. The model measured status recursively, taking account of both the number of connected pages and the status each of those pages already held. The result was an architecture of attention in which pages backed by many high-quality referring pages rose toward greater visibility.
Visibility then fed directly into the practical stakes of publishing. Every outgoing link became a share of a page’s endorsement ability, distributed uniformly among the pages it referenced. The PageRank account describes that distribution as a core part of the algorithm, and it made editorial choices consequential at network scale. A link could send a reader somewhere, while also contributing to the recipient’s calculated place in the wider web. Publishers, institutions and site owners therefore operated inside a system where being cited by already prominent sources could materially shape discoverability.
The invention also created a more literal form of institutional value. Stanford University held the PageRank patent, rather than Google, and received 1.8 million Google shares in exchange for its use; in 2005, Stanford sold those shares for $336 million. The patent record captures the unusual path from a ranking method to a university asset with a defined financial return. The patent itself was granted in 2001, tying a formal legal claim to the method that had become central to Google’s search engine.
By September 24, 2019, all patents associated with PageRank had expired. Their expiration closed one chapter of ownership, while the broader logic remained embedded in the way the web had learned to treat links as evidence of standing. That logic had already made every citation a small editorial act, and it would soon surface again in systems built to answer questions rather than simply rank pages.
AI Search Is Relearning the Value of a Citation
In systems built to answer questions, that editorial act now happens inside the answer itself. A May 2026 Muck Rack analysis examined more than 25 million links cited by ChatGPT, Claude and Gemini across 17 sectors, and found that third-party platforms supplied 84 percent of those citations. Journalism accounted for 27 percent, a share that held steady across three editions of the study. The pattern gives outside publishers a central role in the material AI systems present as support, placing the value of independent mention directly inside the new interface for finding information.
That role grows sharper when an answer needs to choose among many possible sources. Evertune studied nearly 400 million citations across six models and found that 63 percent led to list pages; among those list pages, 71 to 86 percent were numbered rankings. Peec AI found a related advantage across 5.7 million data points and eight models: brands at the top of lists were 13 to 16 percent likelier to appear in an AI answer. A ranked list packages judgment in a form that travels easily, giving a model a ready-made ordering when it gathers evidence for a response.
That creates a familiar contest in a different setting. The question is increasingly whether a company, publication or expert appears in the sources that models choose to cite, and whether it appears high enough in those sources to be carried into the answer. In 2026, platforms emerged to track how language models mention brands, while Muck Rack found that 76 percent of PR professionals already use generative AI. The work of earning attention now reaches from search results into the evidence beneath an AI-generated sentence.
The strongest route into that evidence may be material that others can use. A peer-reviewed study by Princeton University and Georgia Tech tested nine content-optimisation techniques across 10,000 queries, and original statistics produced the largest gain: visibility in AI answers rose by 41 percent. That finding carries the old proposition forward with unusual clarity. Attention flows toward sources that give other people something worth citing, and every new citation extends their standing into the next answer.
Comments (0)