GlossaryAI search
What is a large language model (LLM)?
Also called: LLM, language model, foundation model
Definition
A large language model (LLM) is an AI model trained on very large amounts of text to predict the next piece of text. That one skill lets it answer questions, summarize, write and reason in plain language.
Large language model (LLM), explained
Under the hood, an LLM reads text as tokens, small chunks of words, and learns statistical patterns across billions of examples. Given a prompt, it predicts the most likely next token, then the next, and so on. Scale is what makes this surprising: with enough data and parameters, prediction starts to look like understanding. ChatGPT, Claude, Gemini and the models behind Perplexity are all built on LLMs.
Two sources of knowledge matter for founders. The first is training data: whatever text the model saw before its knowledge cutoff. That's where a model's general sense of your category, and maybe your brand, comes from. It's fixed until the next model version. The second is retrieval: text the system fetches at answer time, usually through a web search, and puts in front of the model. That's how assistants cite fresh pages and current prices.
LLMs are good at summarizing, comparing and rewriting. They're bad at knowing when they don't know. A model can state a wrong price or an invented feature with total confidence, a failure usually called hallucination. Retrieval reduces this but doesn't remove it.
For your marketing, the practical takeaway is that LLMs reward the same things good readers reward: clear statements, consistent facts, and pages that answer one question well. A model deciding what to say about you has your pages, directory listings, reviews and forum threads to work from. If those agree with each other, the model is more likely to describe you correctly.
For your product, LLMs are also a building block. Many startups now ship features on top of model APIs, which brings its own vocabulary: tokens, context windows, embeddings and retrieval. The entries linked below cover each one.
Why it matters for founders
Buyers increasingly ask an LLM-based assistant before they search. Knowing how these models learn and retrieve tells you which levers you actually control: crawlable pages, consistent facts and clear answers, not tricks.
Example
A buyer asks an assistant to compare three invoicing tools. The model uses what it learned in training for general context, runs a web search for current pricing, and cites the vendors' pricing pages in its answer.
Common mistakes
- Assuming the model knows your current pricing because it knows your name.
- Treating an assistant's answer about you as fixed. Retrieval can change it week to week.
- Trying to game models with hidden text. They read what crawlers can read, and so do spam systems.
Related terms
- Retrieval-augmented generation (RAG)Retrieval-augmented generation (RAG) is a technique where a system first retrieves relevant documents, then passes them to a language model so it can answer from that material instead of memory alone.
- Tokens (AI)Tokens are the small chunks of text a language model reads and writes, often parts of words. Models count input, output, limits and API pricing in tokens, not words or characters.
- Knowledge cutoffA knowledge cutoff is the point in time after which a language model has no training data. Anything that happened or launched later is unknown to the model unless it's retrieved at answer time.
- AI hallucinationAn AI hallucination is a plausible-sounding statement from a language model that is false, such as an invented feature, price, quote or source, delivered with the same confidence as a correct answer.
- GroundingGrounding is giving an AI model real source content, such as search results or documents, at the moment it answers, so the answer is based on that content instead of only what the model memorized in training.