An AI model trained on vast text to predict and generate language, powering tools like ChatGPT.
A large language model is a type of AI model trained on very large amounts of text to predict the next piece of text given what came before. From that simple objective, at scale, models learn to answer questions, summarise, translate, write code, and hold a conversation. LLMs power tools such as ChatGPT, Claude, Gemini, and the AI features increasingly built into search.
An LLM does not look facts up or reason like a person. It generates plausible text based on patterns learned during training, which is why it can be fluent and useful but also confidently wrong.
The model breaks text into tokens and learns statistical relationships between them across billions of examples. At use time it predicts the most likely next tokens one at a time, guided by your prompt. Its knowledge is frozen at training time unless it is given extra information at runtime.
LLMs now shape how people find information: AI assistants and AI-powered search summarise answers directly and cite a handful of sources. That changes the game for visibility and for building products.
The main pitfall is trusting an LLM as a database of facts. Because it generates rather than retrieves, it can hallucinate, invent citations, or state outdated information confidently. Serious applications ground the model in verified data rather than relying on its memory.
A Large Language Model (LLM) is an AI system trained on vast amounts of text to predict and generate language, and it is the technology behind tools like ChatGPT, Claude and Gemini. It matters because LLMs have rapidly become both a way people search and create, and a building block developers embed in products — from support agents to content tools to code assistants. For marketing and search specifically, LLMs power the AI answer engines reshaping how discovery works, which is why understanding their strengths and limits is now core knowledge, not a niche interest.
An LLM is built on the transformer architecture and works by predicting the most likely next token (a word or word-piece) given the preceding context, having learned statistical patterns of language from its training data. That simple mechanism, at enormous scale, produces surprisingly capable text generation, reasoning and translation. But it also explains the limitations: the model has no built-in notion of truth — it generates plausible continuations, which is why it can "hallucinate" confident falsehoods — its knowledge is frozen at its training cut-off unless connected to live data, and its output quality depends heavily on the prompt and context it is given.
Because LLMs generate plausible rather than guaranteed-correct text, using them well means designing around their limits. Grounding the model in trusted data (retrieval-augmented generation) dramatically reduces hallucination and keeps answers current; clear, well-structured prompts and provided context improve output; and for anything user-facing, guardrails and evaluation are essential to catch unsafe or wrong responses. In production, LLM features need reliability engineering — cost control (tokens add up), latency management, fallbacks, and monitoring — not just a raw API call. The pattern is to treat the LLM as a powerful but fallible component, engineered into a reliable system rather than trusted blindly.
Part of our defined terms knowledge graph — browse every entry in this branch.
Grounding an LLM's answers in retrieved documents so it responds from real, current data.
Optimising content to be cited and surfaced by AI answer engines, not just ranked in blue links.
A proposed plain-text file that gives AI models a curated, easy-to-read map of your site's content.
A longer, more specific search query with lower volume but clearer intent and less competition.
The ratio of a customer’s lifetime value to the cost of acquiring them.
Common questions
Straight answers on how this fits your marketing and build.
Still have questions? Talk to a specialist