Key Takeaways
- Think of it as a writer, not a filing cabinet. A large language model produces text by predicting what comes next. It does not hold a record for your business that it looks up.
- Separate the model from the assistant. ChatGPT, Claude and Gemini are products built around a model. The product can search the web. The model on its own cannot.
- Expect the wording to change between asks. The model picks from several likely next words, so the same question can bring back a different answer. Ask more than once before you conclude anything.
- Put your facts in plain sentences. Text is the only thing a language model reads. What you do, who for, where and at what price should each be a sentence on a page.
- Say the same thing in the same words everywhere. A model continues a sentence with what it has read most. One steady description across your site and profiles gives it one thing to repeat.
You keep reading that ChatGPT, Claude and Gemini are "large language models". Nobody stops to say what that means, and you have a practice to run.
What is a large language model? It is a computer system that has read a very large amount of text and writes by predicting which word comes next. That one fact explains most of what an assistant says about your business, and most of what it leaves out.
Below is the short answer, how the writing works, three things a language model is not, how the model differs from the assistant you type into, and what it changes for a coach or consultant.
What is a large language model?
A large language model is a computer system built from a very large amount of written text that produces new text by predicting, one piece at a time, what should come next. People shorten the name to LLM. ChatGPT, Claude and Gemini are each built on one.
Anthropic's glossary of terms describes these models as "trained on vast amounts of text data" and able to "generate human-like text, answer questions, summarize information, and more". It is a short description, and each half matters. The first half says where the ability comes from, and the second says what comes out.
"Large" means two things. One is the amount of text the model was built from. The other is the size of the model itself. Google's machine learning glossary defines a large language model as, at a minimum, "a language model having a very high number of parameters". A parameter is one of the numbers inside the model that was adjusted while it was being built.
Those numbers are all there is. There are no pages in there, and no list of businesses. That is the part worth holding on to.
How it writes an answer, one word at a time
It writes one piece at a time. Given the words so far, the model works out which words are likely to come next, picks one, adds it to the end, and does the same thing again until the answer is finished.
This is how the models are built in the first place. Anthropic's glossary says they are trained "to predict the next word, given the previous context of text in the document". The pieces are called tokens, and a token is a word or a part of a word. You will see the term in articles. You do not need it for anything in this post.
The figure shows the idea twice. The examples are made up, and the bars show more and less likely, not measured amounts.
An ordinary sentence
A good coach listens before giving
-
advice
Read most often after those words
-
feedback
Read fairly often
-
orders
Read now and then
A sentence that ends in a business
A well-regarded executive coach in Denver is
-
A name it read often beside those words
The likely pick
-
A name it read a few times
Comes up some of the time
-
A name it never read
Not a candidate at all
It does not look a name up. It continues the sentence with what it has read most.
The sentences in the figure are invented for the example, and nobody outside the companies that build these systems can see the real weights. The shape is the point. A business the model never read about is not ranked low. It is missing from the choices.
The model also does not always take the most likely word. It chooses among the likely ones, which is why you can ask the same question twice and get two differently worded answers. Anthropic's glossary notes that even with the randomness setting turned all the way down, "the results will not be fully deterministic".
What a large language model is not
A large language model is not a database, not a directory and not a search engine. Each of those stores things and fetches them. A language model stores no entries and fetches nothing. It writes.
Three consequences follow, and each one shows up in what assistants say about real businesses:
- It can be fluent and wrong. Predicting a likely sentence is not the same as checking a fact. A model can state a price or a credential that sounds right and is not.
- It stops at the day it was built. On its own, a model has read nothing written after that point. A service you launched last month is not in it.
- It cannot show its working. A model cannot tell you which page taught it a thing, because it kept patterns and not pages.
The table sets the model beside the tools you already know, using one buyer question.
| Tool | What it holds | How it answers "who is a good executive coach in Denver?" | How your business gets in |
|---|---|---|---|
| A directory | One entry per business | Shows every entry filed under that category and city | You or the directory add an entry |
| A search engine | A copy of pages it has visited | Lists pages that match the words, in an order it chooses | It visits your page and keeps a copy |
| A large language model on its own | Numbers that capture patterns in the text it was built from | Writes a sentence, continuing with the names it read most often beside those words | It read about you, more than once, before it was built |
| An AI assistant with web search | The same model, plus pages it fetches while you wait | Runs searches, reads a few pages, then writes a sentence using them | It read about you before, or it finds a page that states the answer plainly |
The last two rows are the ones that matter to you, and the difference between them is the subject of the next section.
The model and the assistant are two different things
The model is the part that writes. The assistant is the product built around it: a chat window, instructions from the company that made it, the conversation so far and, often, a web search. When people say "ChatGPT told me", they mean the assistant.
That difference decides where an answer about your business comes from. Without a search, the model can only continue from what it read while it was being built. With a search, the assistant fetches pages and places them in front of the model as part of the text so far.
Anthropic's glossary has a name for that space. It calls it the context window, a "working memory" for the model, and says it is different from the large body of text the model was trained on. A page that lands in the working memory can be quoted in the answer even if the model never read it before. We cover the two sources in what retrieval-augmented generation is, in plain English.
Google describes its own version of the lookup. Its page on AI features and your website says AI Overviews and AI Mode may use a "query fan-out" technique, which means running several related searches to build one response. The searching is done by the product. The writing is done by the model. For how that differs from a list of links, see what AI search is and how it differs from Googling.
What this changes for your business
It changes three habits: write your facts as sentences, describe the business the same way everywhere, and check more than once. None of them needs a developer, and none needs you to learn another term.
Write your facts as sentences. Text is the only thing a language model reads. The same Google page advises "making sure that important content is available in textual form". Who you work with, what you charge and where you are based should each be a sentence on a page, not a detail inside an image or a video.
Describe the business the same way everywhere. A model continues with what it has read most. If your site says "executive coach", your LinkedIn says "leadership advisor" and a directory says "business consultant", it has read three thin descriptions and no strong one. This is our reasoning from how the models work, not something we have measured.
Check more than once. One answer is one draw from several likely ones. Ask the question a few times, in the words a buyer would use, with web search on and with it off if the assistant lets you choose. The method is in how to check if AI recommends you.
Here is what we cannot tell you. Nobody outside the companies that build these models can open one and see what it holds about your business, and that includes us. Anyone who offers to "add you to ChatGPT" is describing something no outside firm can do. What can be done is to read the answers, and to improve the pages an assistant reads.
The free check asks ChatGPT 25 real buyer questions from your niche at $0 and emails a plain-English readout: your score, who got named in your place, and the one fix to start with. It runs as soon as you ask for it, and the report lands within the hour between 7am and 5pm Central, or by 10am the next morning outside those hours.
Questions we hear the most
What is a large language model?
A large language model is a computer system built from a very large amount of written text that writes by predicting what comes next, one piece at a time. ChatGPT, Claude and Gemini are each built on one. People shorten the name to LLM.
Is ChatGPT a large language model?
ChatGPT is a product built around a large language model. The model does the writing. The product adds the chat window, the company's instructions and, when it is switched on, a web search. The same is true of Claude and Gemini.
Does a large language model know my business?
Only if it read about your business while it was being built, or reads about it during the conversation. It keeps no record for you that anyone can open, so the practical test is to ask it and read what comes back.
Why does the same question get a different answer each time?
Because the model chooses among several likely next words instead of always taking the top one. A different early word leads to a different sentence, and sometimes to different businesses being named. That is normal, and it is why one ask proves little.
Does a large language model search the internet?
Not on its own. A large language model only continues text. The assistant built around it can run a web search and place the pages it finds in front of the model, which then writes an answer using them.
Do I need to understand large language models to show up in AI answers?
No. You need three habits that follow from how they work: state your facts in plain sentences, describe the business the same way everywhere, and check what assistants say more than once. The rest is detail.
What does the AnswerHalo free check show me?
It shows whether ChatGPT names you. The free check asks ChatGPT 25 real buyer questions from your niche at $0 and emails your score, who got named in your place and the one fix to start with. It runs as soon as you ask for it.
What does the $500 audit add to the free check?
The audit runs 125 checks across ChatGPT, Claude, Gemini, Perplexity and Google AI Overviews, keeps every answer so you can read it, and lands by this time tomorrow. The free check covers ChatGPT only, with 25 questions.
Your buyers are already asking. Find out what AI tells them.
The free check asks ChatGPT 25 real buyer questions about your niche and sends you the report within the hour, at $0. If it shows you are already getting named everywhere, we will say so in plain words.
Get my free check →