A language model generates text by selecting tokens in sequence. Its output depends on learned parameters and the context supplied for that request. These mechanics help explain why recommendations can vary and why fluent answers can contain errors.
Tokens and prediction
A tokenizer converts text into tokens. A token can represent a word, part of a word, punctuation, or another text sequence. The exact division depends on the tokenizer. The number of tokens in a brand name does not by itself establish how often that brand will be recommended.
At each generation step, the model computes values used to select a next token. Earlier text and learned parameters affect those values. A generated list of brands therefore differs from a stored list with fixed positions.
Temperature: why the same prompt gives different answers
Generation settings can affect token selection. Where temperature is supported, lower values generally concentrate selection on more probable tokens. Zero temperature does not guarantee identical outputs across implementations or repeated requests. Product instructions, model updates, retrieval, and other settings can also change answers.
Repeat prompts when measuring visibility. Record how often a brand appears across completed answers instead of treating a single response as a stable position. See Measuring AI Search Visibility.
Training data, weights, and the cutoff
Training adjusts model parameters using training examples. The information learned in those parameters is often called parametric knowledge or model memory. A generated brand description is not necessarily retrieved from a stored company record. Updating a website does not directly edit a deployed model’s parameters.
A published knowledge cutoff describes limits on a model’s training knowledge. It does not guarantee that the model knows every fact before that date. Information supplied in a prompt, retrieved from a source, or added through later model updates can affect an answer separately.OpenAI publishes each model's cutoff in its model documentation. Record the cutoff for the exact model used in a test. Other providers publish theirs in model cards.
What output tests can tell us
Training data can affect which names and associations a model generates. Public output tests alone cannot reveal how many mentions a specific brand had in training or the effect of any single source. The visibility research examines observable relationships without treating them as direct measurements of training data.
Factual errors
A model can generate a plausible statement that is false or unsupported. Fluency is therefore not a reliable check of factual accuracy. Source retrieval and verification can help, but a cited answer still needs to represent the source correctly.
For a business, check product names, features, pricing, service areas, and eligibility requirements. Publish current facts clearly and keep authoritative pages available for retrieval. Record errors by model and test condition so corrections can be evaluated.