Large Language Model (LLM)

Two properties of LLMs affect AI-visibility measurement. They are non-deterministic, so the same question can produce different answers across runs. Record multiple answers instead of treating one as representative. When grounded with web search, the model reads current pages before answering, so those sources influence what it says about a business.

Non-determinism occurs because these systems sample from multiple plausible continuations. A business named in one answer out of one provides little evidence about a broader pattern. A business named in eleven of eighteen answers provides a more stable estimate. Include the sample size with every AI-visibility result.

When an LLM recommends a business, it produces text from the question and retrieved evidence. It may not be consulting a fixed ranked database or applying a transparent set of criteria. The businesses it names are those the available evidence supports clearly and favorably, which does not prove they are the best options. Businesses can improve the clarity and accuracy of that evidence.

Distinguish the underlying model from the consumer product when testing. ChatGPT combines an LLM with search, retrieval logic, system instructions, and safety behavior. An API call with default settings may omit those product features and therefore measure a different experience.

Record the product, model when disclosed, search setting, location, date, and conversation state so another test can reproduce the conditions.

An ungrounded LLM answering from training memory provides a different signal limited by its knowledge cutoff. The Routeless methodology uses grounded assistants to match the consumer products being studied.