Large language model (LLM)

How they generate text

A language model predicts the next word, or token, over and over, based on the context so far. For a search query, relevant web sources are fed in as well through retrieval-augmented generation: the model "reads" the sources, writes an answer and can link back to them. One point deserves emphasis: these models do not understand in the human sense, they produce statistically likely text. They can be factually wrong — the phenomenon known as hallucination — which makes citing sources all the more important on consequential topics.

Pre-training, fine-tuning and human feedback

• Pre-training — the model learns to predict the next token across enormous web corpora: Common Crawl, Wikipedia, GitHub, academic papers. This is where its knowledge of the world settles.
• Fine-tuning — targeted further training for particular domains such as code, medicine or law.
• Reinforcement learning from human feedback — human raters judge answers, and the model learns to be more helpful, more courteous and safer.

What this means for GEO

The implication matters: language models "know" domains that appeared in pre-training or fine-tuning better than unfamiliar ones. A brand cited frequently in authoritative sources — Wikipedia, Wikidata, major publishers — is treated intuitively as reliable. Brand building has therefore become a GEO lever in its own right. A domain with a clear Wikidata entity, Schema.org markup and consistent author bylines is far likelier to be judged a citable source than a technically similar domain without those signals.