Definition
An LLM (large language model) is an AI model trained on a very large amount of text so that it can understand, summarise, translate and generate language. ChatGPT, Gemini, Claude and the models behind Perplexity and Google AI Overviews are all LLMs or systems built on them. The model keeps predicting the most likely next piece of text, and does that so well that coherent answers come out.
An LLM has two kinds of knowledge. The first sits in the model itself: everything it saw during training, up to a certain date. That knowledge is broad but frozen and can be outdated or inaccurate. The second kind comes from outside: pages the system fetches live through search and hands to the model as context. That second mechanism is called retrieval-augmented generation and is the reason an AI assistant can cite current sources.
For SEO and GEO, that distinction matters a lot. You only influence the knowledge in the model indirectly: by being mentioned consistently and often across the web, so that future model versions know your brand. You influence the retrieved context directly: by publishing pages that are reachable for AI crawlers and that answer a question so directly and factually that the system picks them as a source.
An LLM also has limitations you need to know. It can hallucinate: produce convincing but wrong information, such as a wrong price or a non-existent product. It gives a different answer to the same question each time, because it works with a degree of randomness. And it does not weigh sources the way Google does with links, but by how well a passage fits the question.