AI search glossary
What is an LLM? Large language models explained for site owners
A large language model (LLM) is a machine learning model, usually built on the Transformer architecture with billions of parameters, that predicts the next token in a sequence of text. That one skill lets it answer questions, summarize, translate and write.
Also called: LLM, LLMs, language model, foundation model
Why it matters for AI search
Every AI answer engine runs on an LLM. What the model says about your brand comes from two places: what it absorbed in training, and what it retrieves from the web at answer time. You cannot edit the first directly, but you can make the second easy: crawlable pages that state who you are, what you sell and how you compare, in plain sentences the model can quote.
How a large language model works
Google's introduction defines a language model as one that estimates the probability of a token, or a sequence of tokens, appearing within a longer sequence. A token is the smallest unit the model works with: a word, part of a word or a single character. For English, Google says one token is roughly three quarters of a word.
A large language model does the same job at a much bigger scale. Modern LLMs use the Transformer architecture, which relies on self-attention: for each token, the model weighs how much every other token in the input affects its meaning. That lets it read the whole context at once rather than one word at a time, which is why LLMs handle long questions far better than older models.
Training happens in two broad steps. The model first learns general language patterns from a huge text corpus. It is then tuned to follow instructions and hold a conversation. The result is a model that can answer, summarize, translate and write.
What it means for AI search
An AI answer engine is an LLM plus a way to find sources. When a product like ChatGPT search or Google AI Mode answers a question, it often searches the web first and gives the LLM the results to write from. That pattern is retrieval-augmented generation.
So there are two routes for your brand into an answer:
- Training data. What the model learned before its knowledge cutoff. Slow to change and hard to influence directly.
- Live retrieval. Pages the engine fetches at answer time. This is where crawlable, specific content pays off, and where answers can be grounded in sources.
When neither route has clear information, the model fills the gap, which is how LLM hallucinations about businesses happen.
How to make your site easy for LLMs to use
- Let the AI crawlers you want reach your pages. Check with the AI crawler checker.
- State the basics in text: what you sell, who it is for, price, where you operate.
- Keep facts consistent across your site and profiles so retrieved sources agree.
- Ask the main engines about your category and note what they get wrong.
Related ideas
LLM SEO covers the practice of shaping how these models describe a brand. For how specific engines use LLMs, see the ChatGPT, Claude and Gemini engine pages.
Questions
- What is an LLM in simple terms?
- It is a program trained on a very large amount of text to guess what comes next. Because it has seen so much language, those guesses add up to fluent answers, summaries and drafts. ChatGPT, Claude and Gemini are products built on LLMs.
- What makes a language model large?
- Mostly the number of parameters (the learned weights) and the amount of training data. Google's crash course notes that Transformer models can have hundreds of billions or even trillions of parameters, and that larger models tend to perform better.
- Do LLMs know facts about my company?
- Only what appeared in their training data, which stops at a cutoff date, plus whatever they retrieve live when a product such as ChatGPT search or Perplexity searches the web. If neither source has clear information about you, the model may omit you or guess.
- Why do LLMs get things wrong?
- They predict likely text rather than look up a fact table, so they can produce confident mistakes, known as hallucinations. Google's course lists hallucination and bias among the known problems with LLMs.