AI search did not appear from nowhere. Search has spent more than twenty years drifting in one direction: from matching strings to modeling meaning, and from listing pages to answering questions. Each pivot along the way rewarded the people who understood the machinery early, and that pattern is repeating right now. Readers who know the lineage make better decisions than readers who treat 2023 as year zero, because most of what works today was built as a response to something that broke earlier.
Directories, then links
The earliest way to find things on the web was a human-edited directory: Yahoo started as a categorized list of sites maintained by people. Crawler-based engines like AltaVista replaced the human editors with software that matched the words in your query against the words on a page. That worked until people learned to stuff pages with the words they wanted to be found for, and relevance collapsed.
Google's answer, published by Brin and Page in 1998, was PageRank: treat a link from one page to another as a vote, weight votes by the importance of the page casting them, and rank results by that accumulated authority on top of text matching."The Anatomy of a Large-Scale Hypertextual Web Search Engine", the original Stanford paper describing Google and PageRank. Links were harder to fake than on-page text, so results got better, and an entire economy formed around earning, buying, and manufacturing links. Twenty-five years later, links remain one of the strongest measurable signals of AI visibility; Links: The Signal That Refuses to Die has the data.
The lesson of the era: every ranking system gets reverse-engineered, and the practitioners who understood the actual mechanism, rather than the folklore around it, won the years that followed.
Answering before LLMs
Through the 2000s and 2010s, Google steadily reshaped the results page from a list into something closer to an answer. Universal search in 2007 mixed news, images, and video into the ten blue links. Featured snippets pulled a paragraph out of a ranking page and put it at the top of the results, answering the question before anyone clicked. The pattern that AI search now completes started here: Google composing the answer, your site supplying the raw material, and the click becoming optional.
The deeper shift came in 2012 with the Knowledge Graph, which Google introduced with the phrase "things, not strings": a database of more than 500 million entities and 3.5 billion facts about the relationships between them, so that a query about "taj mahal" could resolve to the monument, the musician, or the restaurant rather than to pages containing those letters.Google's 2012 announcement, "Introducing the Knowledge Graph: things, not strings", with the entity and fact counts as stated at launch. Being an unambiguous entity in that graph became a ranking asset, and it still is: entity work is one of the main memory-side levers in this book, covered in Entities: Becoming a Thing the Machine Knows.
Language understanding moved in stages: Hummingbird rebuilt the query engine around meaning in 2013, RankBrain added machine learning to ranking in 2015, and in 2019 Google shipped BERT, a neural language model applied to roughly one in ten English searches in the US at launch, to read queries closer to the way people mean them.Google's 2019 BERT announcement, which states the one-in-ten figure and the focus on longer, conversational queries. Voice assistants pushed the same direction from the demand side: a smart speaker cannot read you ten links, so it answers with one thing, a preview of the winner-take-all dynamics AI answers have now made general.
By 2019, then, everything AI search needed was already in place: entities to ground facts, neural models to read meaning, and a results page that increasingly answered instead of listing. What was missing was a model that could write.
The generative turn
ChatGPT launched in November 2022 and became the fastest-adopted consumer technology on record. For search, the important part was behavioral: hundreds of millions of people learned to ask a question in full sentences and receive one composed answer, with no results page anywhere in the experience.
The platforms then converged from both directions. Google bolted generation onto search: the Search Generative Experience experiment in 2023 became AI Overviews in 2024, an AI summary above the classic results, and then AI Mode, a full conversational search tab. OpenAI bolted search onto generation: ChatGPT search, launched in 2024, turns prompts into web queries against third-party search providers and partner content, and cites the pages it draws from.OpenAI's announcement, "Introducing ChatGPT search", describing the third-party search providers and publisher partnerships behind it. Perplexity built an answer engine from scratch on its own crawl, and Anthropic and Microsoft wired web search into Claude and Copilot. The full tour of who runs what, and which search index sits underneath each product, is in The AI Search Ecosystem.
Why the lineage matters
Two practical conclusions follow from this history.
First, AI search inherits its parts from the previous eras, which is why so much SEO knowledge transfers. The retrieval step of an AI answer runs on classic search indexes, so ranking still matters. The entity graphs of the 2010s feed what models know about brands. The link economy of the 2000s shaped the training data every model learned from. What Carries Over from SEO, and What Breaks maps the transfer in detail.
Second, at every pivot, the advantage went to practitioners who understood the new machinery while everyone else applied the old folklore to it: the people who read the PageRank paper while others swapped links, the people who built entity presence while others counted keyword density. The equivalent work today is understanding how a model's memory differs from its retrieval, and that is where this book goes next, starting with How an LLM Turns a Prompt into a Response.