Large Language Model (LLM)

Two properties of LLMs matter commercially. They are non-deterministic — the same question can produce different answers on different runs, which is why credible AI-visibility measurement records many answers rather than one. And they are evidence-driven when grounded — with web search enabled, the model reads real pages before answering, so what those pages say about your business decides what the assistant says.

Non-determinism is not a defect to be engineered away; it is how these systems sample from many plausible continuations. For measurement it has a specific consequence: a business named in one answer out of one has told you almost nothing, while a business named in eleven of eighteen answers has told you something reliable. Any AI visibility claim without a sample size behind it is describing a single roll of the dice.

It also helps to be clear about what the model is and is not doing when it recommends a business. It is not consulting a ranked database or applying criteria. It is producing the most plausible text given the retrieved evidence and the question, which means the businesses that appear are the ones the evidence describes most clearly and most favourably — not necessarily the best ones. Clarity of evidence is a competitive variable, and it is the one businesses can actually change.

The distinction between the model and the product matters for anyone testing this. ChatGPT the consumer app is an LLM plus search, plus retrieval logic, plus system instructions, plus safety behaviour. Calling the underlying model through an API with default settings measures none of that scaffolding, and produces conclusions about a system no customer uses.

An ungrounded LLM answering from memory is a different, weaker signal, limited by its knowledge cutoff. Our methodology uses grounded assistants exclusively, because that is what a real buyer's app does.