When LLMs Get Personalized

LLM-based search can vary depending on user context, but that doesn't mean answers scatter into complete disarray. The author argues that even with personalization, the model's shared priors, overlapping retrieval areas, and the same decoding constraints remain at work, so answers still retain a stable common structure.

The key point is not token-level variation but semantic-level difference. Answers to the same question may differ across users in examples, emphasis, order, and regional details, but on the parts that actually matter, they're likely to converge into a form where a shared core is held in common and only the periphery varies as a variable margin.

The piece first clears up a few misconceptions. It pushes back on the claims that personalization makes every answer completely different, that a probabilistic system can't have modelable patterns, and that personalization makes optimization itself impossible, explaining that even natural language answers have sufficiently repeatable architecture and probabilistic order. It concludes that both extremes are lacking—viewing LLM search as a return to the pre-SEO world, or, conversely, treating it as an entirely new game altogether.