How it works
During training, the model learns to predict how a text continues, across billions of sentences. It picks up a general knowledge of language, common facts and simple reasoning. When it answers, it generates the reply piece by piece, based on the question and the context it's given.
Two limits matter for an e-commerce site:
- Its knowledge stops at its training date. It doesn't know today's prices, current stock or the latest collection.
- Without the right information at hand, it can produce an answer that sounds right but is wrong. That's called a hallucination.
That's why the useful content is passed to it at question time: product pages, FAQ, return policy. This method is called RAG (retrieval-augmented generation).
Why it matters in e-commerce
LLMs let shoppers ask in their own words, like "a firm mattress for someone with back pain", instead of guessing the right keywords. They also power ChatGPT, Gemini and the other assistants consumers use to look for products.
So brands deal with LLMs in two places: on their own site, where they make search and advice conversational, and outside it, where they decide which products get mentioned.
Further reading: Claude for Commerce: what Anthropic shipped, and what you still have to build
Real-life example
A shopper asks: "Does this backpack count as carry-on luggage?" An LLM on its own answers from general knowledge and may make up the dimensions. The same LLM, connected to the brand's product page and FAQ, reads the backpack's real dimensions and answers with the brand's information.
Common confusion
Many people assume an LLM knows a brand's catalog because it can name well-known products. It only knows what it saw during training, which is often incomplete or out of date. Prices, stock and new arrivals have to be passed to it when the question comes in.