Run free audit
AI Engines and Surfaces

Large Language Model (LLM)

A large language model (LLM) is a machine learning model trained on huge amounts of text to predict the next token in a sequence, which lets it generate fluent natural-language responses and power products such as ChatGPT, Perplexity, Gemini and Copilot.

Also known as:LLM, foundation model, generative language model

LLMs are statistical models built on the transformer architecture. During training they read enormous text corpora and learn the probability of each possible next token given the tokens that came before. At inference time they sample from that distribution one token at a time, which is why their output looks like coherent language even though the underlying operation is sequential prediction rather than understanding in any human sense.

Two properties matter for visibility work. First, an LLM only knows about a brand if information about that brand was in its training data or is retrieved at query time. A brand that has almost no presence in either is effectively invisible to the model. Second, the model does not pull information up neatly by source; it blends what it has learned, which is why details can drift or become inaccurate. Both properties make grounding, retrieval and high-quality external coverage critical for accurate brand mentions.

Modern AI products usually combine an LLM with retrieval, search and tool-use layers so that answers can include fresh, sourced information rather than relying purely on what the model remembers. This is the foundation of retrieval-augmented generation and is the reason a brand can become visible in AI answers fairly quickly through new web content, even if it never appears in the underlying training set.

Key points

  • LLMs predict the next token in a sequence given prior context.
  • They generate language but do not retrieve information cleanly from a known source.
  • Brands absent from training data and from retrievable sources are invisible to LLMs.
  • Most AI products combine an LLM with retrieval to ground answers in fresh sources.

Frequently asked questions

What does LLM stand for?

LLM stands for large language model. It is a type of AI model trained on huge text datasets to generate language by predicting one token at a time.

Do LLMs know everything on the web?

No. An LLM only knows the content that was in its training data plus anything retrieved at query time. Sites that block AI crawlers or have very little web presence may be unknown to the model.

Related terms

Retrieval-Augmented Generation (RAG)
Retrieval-augmented generation (RAG) is an AI architecture that first retrieves relevant documents from an external source and then feeds them to a language model so the model can ground its answer in those documents rather than relying only on what it memorized during training.
Grounding
Grounding is the practice of constraining an AI models answer to specific retrieved sources so that the response is supported by evidence from those sources rather than generated freely from the models internal knowledge.
Hallucination
A hallucination is an AI-generated statement that is presented as factual but is actually invented, distorted or otherwise unsupported by reliable sources, and it is one of the central risks of using language models in answers about brands.
Vector Embedding
A vector embedding is a numerical representation of a piece of text (or other data) as a list of numbers in a high-dimensional space, designed so that texts with similar meaning end up close together and can be compared by a fast distance calculation.
Answer Engine
An answer engine is a search or chat product that returns a direct synthesized answer to a question instead of (or in addition to) a list of links, with ChatGPT, Perplexity, Google AI Overviews, Gemini and Copilot being the most widely used examples.
See how AI engines describe your brand.

Free audit. Score across ChatGPT, Perplexity, Gemini and Google AI Overviews.

Run a free audit
Back to the dictionary