What are LLMs, and why should a business owner care?
By Bhavesh Barot
- Updated
- Reading time
- 8 min read
- Guide number
- No. 021

On this page
- 01What does "large language model" mean?
- 02How does an LLM write an answer?
- 03Where does an LLM's knowledge come from?
- 04Why do LLMs make mistakes?
- 05What does this mean for how buyers find you?
- 06Which LLMs power which tools?
- 07Can you check what LLMs say about your business?
- 08How does Seen in Search help?
- 09Questions people ask about LLMs
LLMs (large language models) are the AI systems behind tools like ChatGPT, Google's Gemini, Anthropic's Claude and Meta's Llama. They're computer programs trained on huge amounts of text to predict which words should come next, which turns out to be enough to write answers, summaries, emails and code that read as if a person wrote them.
You don't need to know how to build one. But if your buyers ask AI tools for recommendations, it helps to know where the answers come from, why they change, and why they sometimes get your business wrong. This guide covers that in plain terms.
What does "large language model" mean?
Each word does some work.
- Language model. A program that has learned the patterns of language well enough to predict the next piece of text. Your phone's keyboard suggestions are a tiny language model.
- Large. Modern LLMs have billions of internal settings, called parameters, adjusted during training on a vast amount of text from the web, books and other sources. Size, plus the amount of training text, is what lets them handle almost any topic.
Most of today's LLMs use a design called the transformer, introduced by Google researchers in a 2017 paper, "Attention Is All You Need". It lets the model weigh how each word in a passage relates to every other word, which is what makes long, coherent answers possible.
How does an LLM write an answer?
In simple steps:
- Your question is split into tokens. Tokens are pieces of words. "Landscaping" might be one or two tokens.
- The model predicts the next token. Based on everything it learned, it scores which tokens are likely to come next and picks one, with a little randomness.
- It repeats. The new token is added and the model predicts the next, and so on, until the answer is done.
That's all it does: predict the next piece of text, over and over. There's no database of facts it looks up, unless the tool around it adds one. Knowledge, style and reasoning all come out of patterns learned in training.
The randomness in step 2 is why the same question can get different answers. It's set on purpose, so answers read naturally instead of identically.
Where does an LLM's knowledge come from?
There are two sources, and the difference matters for a business.
| Source | What it is | Can it know about your business? |
|---|---|---|
| Training data | The text the model learned from, up to a cutoff date | Only if your business was well covered on the web before that date, and only as the web described you then |
| Live search (retrieval) | Pages the tool fetches at the time of the question | Yes, if your pages and mentions can be found and read today |
Training data is frozen. If your business opened after the cutoff, or changed its name, services or address, the model's built-in knowledge may be missing or outdated.
Live search is how AI tools stay current. ChatGPT search, Google's AI Overviews and AI Mode, Perplexity and Copilot all fetch pages while answering and usually link to them. Developers call this retrieval-augmented generation, or RAG. For local and current questions, this is the part a business can influence. Our guide to AI search engines covers how each tool does it.
Why do LLMs make mistakes?
Because they predict likely text, not true text. Most of the time the likely answer is also the right one. Sometimes it isn't.
- Hallucination. The model writes something plausible that isn't true, such as a phone number with the right area code but wrong digits, or a service you don't offer. OpenAI says ChatGPT can produce incorrect information that sounds convincing.
- Outdated facts. Training data has a cutoff; your hours, address or owner may have changed.
- Mixing up similar names. Two businesses called "Summit Plumbing" in different states can blur together.
- Repeating bad sources. With live search, the model summarizes what it finds. If a directory lists your old address, the answer may repeat it.
You can reduce these errors for your own business by making your facts clear and consistent wherever they appear, so both the training data and the live search find the same right answer.
See what ChatGPT says when buyers ask about the services you sell. One free report per business. Use your business email. No card.
What does this mean for how buyers find you?
When a buyer asks an AI tool "who's a good accountant for a small restaurant in Phoenix", an LLM writes the reply. If the tool searches, the answer depends on which pages its search step found and what those pages said. If it doesn't search, the answer depends on what the model learned in training.
Either way, three things help:
- Clear pages. Pages that state your services, area, process and policies in plain sentences are easy for a model to summarize correctly.
- Consistent facts. The same name, address, phone and services on your site, your Google Business Profile and every directory.
- Mentions on other sites. Reviews, trade directories and local news give the model more sources that agree about you.
And one thing to avoid: tricks aimed at the model, such as hidden text or pages stuffed with keywords. They break Google's spam policies, and in the GEO research paper, keyword stuffing didn't improve visibility in AI answers. Our post on generative engine optimization covers what did.
Which LLMs power which tools?
| Tool | Model family behind it | Made by |
|---|---|---|
| ChatGPT | GPT models | OpenAI |
| Google AI Overviews, AI Mode, Gemini app | Gemini models | |
| Claude | Claude models | Anthropic |
| Microsoft Copilot | OpenAI's GPT models and Microsoft's own | Microsoft |
| Meta AI | Llama models | Meta |
| Perplexity | Several models, including its own and others' | Perplexity AI |
The model matters less to a business than the tool around it: whether it searches, which index it uses and how it cites sources. Two tools using similar models can still name different businesses for the same question. Providers also update their models often, sometimes several times a year, so an answer you saw in spring may come from a different model by fall. That's one more reason to check on a schedule instead of relying on one old result.
Can you check what LLMs say about your business?
Yes. Ask the AI tools your buyers use the questions your buyers ask, in their words, with your area. Use a private window. Note whether the tool searched (look for source links), whether you were named, what it said and which sources it cited. Ask again on another day, since answers vary. Repeat monthly with the same questions so changes mean something.
How does Seen in Search help?
A Seen in Search report does this with a fixed method. It reads your homepage, writes buyer questions based on what the homepage says you sell, puts them to AI assistants and scores how visible you are in the answers, in code, with fixed, published weights. The score itself is computed in code, not by an AI model. The free report checks ChatGPT. The full paid report covers ChatGPT, Google AI Overviews and Perplexity. Our methodology page shows every step. Reports are private when they finish.
Questions people ask about LLMs
What does LLM stand for?
Is ChatGPT an LLM?
How do LLMs learn?
Do LLMs search the internet?
Why do LLMs give different answers to the same question?
What is an LLM hallucination?
Can an LLM know about my small business?
What is the difference between AI and an LLM?
Are LLMs safe for business use?
How will LLMs change how customers find businesses?
One free report per business, private unless you publish it. Use your business email. No card. Questions about the report? Contact us.


