PageRank and the anatomy of Google

Two Stanford graduate students presented a paper at the World Wide Web conference in Brisbane in April 1998 describing a search engine they had built, and the idea that made it work. PageRank treats a link from one page to another as a vote, weights the vote by the importance of the page casting it, and solves the resulting circular definition as the principal eigenvector of the web's link matrix. It was old mathematics, from the study of citation networks and Markov chains, applied to a graph of 24 million pages.

Google was incorporated five months later. It is on this timeline as a product and as a cause. The company it became funded a large share of the deep-learning era, Google Brain, DeepMind, the transformer and the TPU, out of the advertising revenue that PageRank made possible. The web that PageRank organised is also the training data: the text of every large language model is, in large part, the corpus Google was built to index.

The paper's authors worried in its final section that advertising-funded search engines would be biased towards advertisers. That concern became one of the central arguments of the 2020s, about who trains the models and to what end.