James Dooley speaks with Benjamin Tannenbaum from AISO about the personalisation of large language models and why identical prompts in ChatGPT, Gemini and other AI search systems can produce different answers. Benjamin Tannenbaum explains the difference between true personalisation and simple probabilistic variability, showing how response volatility often comes from token selection rather than user history. The conversation breaks down location signals, memory settings, login state, and how query fan out becomes weighted by personal attributes. They also explore why personalisation can reduce randomness by narrowing candidate sets, how share of voice replaces fixed rankings in AI visibility tracking, and why optimising for LLMs increases probability rather than guaranteeing deterministic results.