Jeff Dean, who left his post as Google’s chief scientist after 27 years to start an independent research venture, told Y Combinator partner Diana Hu that the AI model a team picks now matters less than what surrounds it. For search and content teams building on large language models, the practical question is no longer which model to license but whether a retrieval and tooling setup can feed that model the right information at the right moment. Dean’s answer, given in a recorded Y Combinator interview and reported by Search Engine Journal, describes a shift that overlaps closely with concepts SEO practitioners already track under a different name.
Hu put the shift to Dean directly: real progress once came from building bigger models on more training data, and now comes more from the tools, retrieval and memory layered around a model, a combination she suggested is folding into a single practice often called context engineering. Dean agreed and went further, saying a model is one piece within a larger system built to solve a specific problem. What matters more, he said, is whether that model can call the right tools, retrieve relevant information and draw on a record of what worked on similar problems before.
Dean also addressed multi-agent orchestration, the coordination of several AI agents as they call tools and retrieve information to work through a task. He drew a contrast between a model’s training data, an undifferentiated mass compressed across trillions of parameters, and the narrow set of information assembled for one specific task, which the model can use far more directly. Orchestration, in his framing, means identifying which tools apply, breaking a problem into a sequence of tool calls, and testing multiple approaches to find which one produces a working result.
Asked for concrete advice on getting better at this work, Dean pointed away from retraining. Adjusting a model’s parameters directly is difficult for anyone outside the lab that built it, he said. The more available lever is writing clear guidelines and reusable skills that tell a model how to handle a specific class of problem, then refining those guidelines each time the system fails on a real case. Repeated enough times, he said, that process produces a setup that keeps getting better on its own.
Dean was describing how AI companies build products, not how search results get ranked. The mechanism he described, a system’s usefulness depending on what gets pulled into its context and how clearly that material is structured for the model to use, is this publication’s own connection, not Dean’s claim. He did not mention search visibility, citations or publisher content in the interview as reported. GEO, generative engine optimization, is the practice of optimizing for LLM-driven answers. AEO, answer engine optimization, optimizes for direct-answer surfaces. Both disciplines exist to influence exactly the retrieval and structuring behavior Dean just described, from outside the system rather than from inside it.
The omission is worth noting. Dean explained how a well-built system decides what to retrieve and how to use it once retrieved. He did not say how that system chooses which external sources qualify for retrieval in the first place, and that gap is precisely the mechanic a content publisher cannot see or control directly.
Search and content teams should read this as confirmation rather than news. The work of structuring pages into clear, self-contained, retrievable units matters more as the systems reading them get built exactly the way Dean described. Auditing whether key pages function as standalone answers, independent of surrounding navigation and design, is a reasonable benchmark to run in the next ninety days.
Search Engine Journal reported this account of Jeff Dean’s Y Combinator interview on August 7, 2026.
