Understanding it
An LLM learns by reading billions of sentences and guessing the next word each time. It builds a statistical representation of language and of the knowledge it carries. Most rely on the Transformer architecture, published by Google researchers in 2017.
Its knowledge stops at a training cutoff, the date of its training data. For a recent or specific question, assistants run a web search, read a few pages and write an answer that cites them. This second mode is the one that directly concerns websites.
For a brand, there are therefore 2 ways in: being present in what the model has learned, which depends on its reputation across the web, and being readable at the moment of the live search, which depends on its site. Practitioners report that LLMs pick up pages passage by passage: a section that answers a question on its own is more likely to be reused.
A concrete example
Illustrative example: a question about a new offer
An online bank launches a new offer for freelancers. Asked without web search, an AI assistant does not know about it: it postdates the training data.
With search enabled, the assistant consults pages and cites those that describe the offer clearly: conditions, pricing, target audience. The bank’s product page, structured as questions and answers, is picked up. A vague press release would not have been.
The common mistake
Assuming an LLM “knows” everything about your brand
An LLM can mix things up, oversimplify or make things up, especially about a small brand. What it says depends on the sources available. Publish clear, dated, consistent information on your site and elsewhere, and regularly check what assistants say about you.