In more detail
An LLM learns patterns in language from a training set of text, much of it collected from the public web. Once trained, it generates answers one word (strictly, one "token") at a time, choosing what is most likely to come next given the question and anything else in its context.
That gives an LLM two different ways to "know" about your business:
- Training data. What the model absorbed when it was built. This is fixed until the next model is trained, can be months or years out of date, and you can't edit it directly. Crawlers such as GPTBot and ClaudeBot collect content that may be used for training.
- Live retrieval. Many AI products now search the web at the moment someone asks a question and pass the pages they find to the model. This is called retrieval-augmented generation, and it is why AI answers can cite current pages and link to sources.
LLMs can be confidently wrong. They may mix up businesses with similar names, repeat outdated prices or invent details that sound plausible. The more consistent and easy to find the facts about you are across the web, the less room there is for those mistakes.
Why it matters when hiring an agency
Agencies selling "AI SEO" or generative engine optimisation are, in practice, trying to influence what LLM-powered tools say about you. Ask whether their work targets live retrieval (which you can influence through accessible, well-structured, well-cited pages) or training data (which changes slowly and can't be controlled). Be wary of anyone who claims they can get you "into ChatGPT's training data" on a schedule, or who promises a fixed position in AI answers. No one outside the AI companies controls either.