LLM (Large Language Model)

Large language models (the GPT-5.x family from OpenAI behind ChatGPT Search, Claude 5.x from Anthropic behind Claude, Gemini 3.x from Google behind Gemini, DeepSeek V4 behind DeepSeek, Muse Spark from Meta Superintelligence Labs, which reportedly powers Meta AI, Mistral, and others) are transformer-based neural networks trained on hundreds of billions to trillions of tokens of text. They learn statistical patterns of language that let them generate coherent text in response to a prompt.

For an AEO / GEO programme, LLMs are the runtime behind every answer engine. Understanding how they work shapes how you write for them:

• They predict the next token given context. The "context" includes the user's question and any retrieved documents (the RAG step). Documents the retriever pulls in have a much larger effect on the answer than documents the model "memorised" during training.
• They favour confident, specific, well-attributed sentences. A page with quotable claims is more likely to be reproduced in the generated answer than a page of hedged generality.
• They're prone to [hallucination](/glossary/hallucination/). When grounding is missing or weak, LLMs invent plausible-sounding content. Engines partly mitigate by aggressive citation requirements; publishers help by being the high-confidence source the model grounds against.

The training corpus question (what LLMs "know" from training) is mostly a distraction for publishers in 2026. Major LLMs publish little about training data, opt-out signals are honoured inconsistently, and the grounded retrieval step dominates AI-search behaviour. Optimise for being retrieved and cited; training-corpus inclusion will follow as a byproduct.