Large language models, explained for a business owner: what they are, how they answer, and why they get things wrong

What a large language model is

A large language model learns from text: books, articles, web pages and much else, gathered before it was trained. From all of it, it learns how words and ideas tend to follow one another, well enough to continue any piece of text in a way that reads as if a person wrote it. It works in tokens, pieces of words, and writes an answer one token at a time, each chosen from what came before. Anthropic's glossary defines these terms for its own models, and the ideas are the same across the field.

Two things follow for a business. The model does not look anything up by itself; it writes from what it learned. And what it learned stops at a date, its knowledge cutoff, so anything that changed after that date is news to it.

Two places an answer comes from

An assistant is more than its model. When web search is on, it can look things up: it runs searches, fetches a handful of pages, and writes its answer from what those pages say, often with the pages listed as sources. Anthropic's documentation of Claude's web search tool describes this for Claude, and the general pattern has a name, retrieval-augmented generation.

Retrieval is the route a business can reach quickly, because it reads pages as they are now. It is also the reason access matters: pages have to be fetched before they can be read, and each assistant fetches through crawlers its operator names. AI search engines at a glance lists them.

Why the same question gets different answers

Ask the same question twice and the answer can change. The model chooses each next token with some room for variation, which Anthropic's glossary describes under temperature; a search can find different pages on a different run; and the earlier messages in a conversation shape what comes next, because the model reads the whole exchange in its context window each time.

So a single answer is a sample, not a verdict. Measuring whether an assistant names a business means asking the same questions several times, on each assistant, in fresh conversations, and reading the answers together. How to measure AI visibility sets out the method.

Why it gets facts wrong

• It filled a gap. When nothing it read states a fact, a model can write a plausible one anyway. That is what people call a hallucination: fluent text with nothing behind it.
• It learned an old version. Training data from before a change still describes the business as it was.
• Its sources disagree. If your site, your map listing and a directory give three phone numbers, an answer has to pick one.
• It found the wrong business. Two businesses with similar names in the same town can be merged into one.

Each has a different fix, and when AI gets facts about your business wrong goes through them. What they share is that the remedy is in the sources, not in the chat: typing a correction into an assistant changes that one conversation and nothing else.

What this means for your site

An answer can only use what the assistant could read or had learned. For a business, that turns into a short list of things to get right, and our audit of 9,483 small business websites shows how often each is missing.

• Let the crawlers in. 7.2% of the audited sites served our crawler a challenge page instead of the site.
• Put the facts in text. 2.1% kept their real content in a PDF, and 3.0% delivered it only through scripts.
• Be one business. 43.1% were inconsistent about who and where they are across their own three main pages.
• Answer the questions buyers ask. 61.5% had no question-and-answer content of any kind.
• Date what changes, so a reader can tell which version is current.

Why ChatGPT does not know your company expands the list into six causes, with a fix for each.

What a language model does not do

• It does not know your business unless it read about it, in training or in the pages it fetched.
• It does not keep your corrections. A fact typed into one chat does not reach the next person's answer.
• It does not take submissions. We know of no documented way for a business to submit itself to an assistant's answers.
• It does not rank pages the way a search engine's results page does, in any way its operators publish. What they publish is how their crawlers fetch.