GEO glossary

Large language model (LLM)

Large language models (LLMs) are AI systems trained on huge amounts of text to predict the next word in a sequence. That one skill lets them answer questions, summarise documents and write. ChatGPT, Gemini, Claude and the answers in Perplexity are all generated by LLMs.

Reviewed by Gareth Hoyle, Managing Director
Last reviewed 31 October 2026

How does an LLM work?

Most current LLMs are built on the transformer, an architecture described in the 2017 paper Attention Is All You Need by Ashish Vaswani and colleagues at Google. A model reads text as tokens, which are words or parts of words, and learns which tokens tend to follow which. OpenAI describes this first stage, pretraining, as predicting the next word in huge amounts of text.

Further training turns the model into an assistant that follows instructions. What it learns is fixed at a training cutoff date, so it knows nothing about later events unless it is given fresh information.

Where does an LLM get its answers?

An LLM answers from two places. The first is what it absorbed in training, stored as patterns rather than as a list of sources. The second is text supplied at the time of the question, such as search results fetched by ChatGPT search, Perplexity or Google AI Mode. That second route is called retrieval-augmented generation, and it is the one that produces links and citations.

What LLMs mean for search and AI visibility

Both routes matter to a brand. Training data shapes what a model says when it does not search: whether it knows the brand exists, what it does and who it competes with. Retrieval decides which pages are read and cited when it does search. OpenAI, for example, says sites that block its OAI-SearchBot crawler will not be shown in ChatGPT search answers.

Check robots.txt allows the AI search crawlers you want, such as OAI-SearchBot and PerplexityBot
State core facts (what you do, where, for whom, since when) in plain text
Earn mentions on sites that models and search tools already rely on
Track how assistants describe you with a fixed set of prompts

Example

Ask an assistant with search turned off which UK agencies specialise in GEO and it answers from training data that stops at its cutoff. Ask the same question with search on and it reads current pages first. A brand can appear in one answer and not the other, which is why AI visibility is worth measuring both ways.