Language model

A language model has read more text than any person could in a lifetime, and has learned the patterns in language. When you write something to it, it works out which words most likely belong in a good answer. That makes it good at writing, summarising and explaining. It does not look things up in a database of facts, unless someone has connected it to one.

Why it can be wrong with great confidence

The model is built to give an answer that sounds right. If it does not know, it often fills the gap with something that seems likely. This is often called hallucination. Opening hours, prices and names are typical places it happens, because they are small details that vary from business to business.

That is also why an AI chat on a website should get access to the business's own texts, for example with RAG, instead of answering from memory.

See what the models think of you

Ask ChatGPT, Claude or Gemini about your business: what you do, where you are based and what it costs. Write down what is wrong. Then check that the correct information is clearly stated on your website. AI tools that search the web can only pick up what is actually there.