What is a large language model (LLM)? Definition and how it works

A large language model (LLM) is an AI model trained on vast amounts of text to predict the next piece of text from everything that came before it, which lets it understand questions and write fluent answers. LLMs are the engines behind ChatGPT, Perplexity, Google AI Overviews and Google AI Mode.

For marketers and SEOs, the LLM is the new writer between your brand and your customer. Instead of a list of ten links, the user reads one answer, and the model decides which brands appear in it, in what order and in what tone.

How a large language model works

You do not need the math to work with LLMs, but four ideas explain most of their behavior:

• Tokens. Text is split into tokens, which are whole words or pieces of words. The model reads and writes tokens, not pages.
• Training. During pre-training, the model reads a huge corpus of public web pages, books, code and other text, and adjusts billions of internal numbers (its parameters) to get better at predicting the next token. Later stages, including feedback from human reviewers, teach it to follow instructions and answer helpfully.
• Generation. When you ask a question, the model writes the answer one token at a time, each chosen from a range of likely options. That is why the same question can produce different answers.
• Architecture. Most current LLMs use the transformer architecture, introduced by Google researchers in 2017, which lets the model weigh how every word in a passage relates to every other.

“Large” refers to both the number of parameters and the amount of training text. Scale is what turned next-token prediction into models that can summarize, compare and recommend.

What an LLM knows and what it looks up

An LLM answering a question in an AI search engine draws on two kinds of knowledge:

1. What it learned in training. This memory is frozen at the model’s knowledge cutoff. It shapes the model’s default associations: which brands it links to a category, and how it describes them.
2. What it retrieves at answer time. Many engines search the web first and pass the results to the model, a technique called retrieval-augmented generation. This is where cited sources come from.

The difference matters for your strategy. Training data changes slowly and you cannot see it. Retrieval happens every time a question is asked and depends on which pages are available, crawlable and relevant today.

Why LLMs matter for brand visibility

In classic search, your page ranks or it does not. In an LLM answer, several things can happen to your brand at once: it can be recommended, listed among alternatives, compared unfavorably, mentioned without a link, linked without being named, or left out. The model also writes the description of your product in its own words, which can drift from how you describe it yourself, and sometimes into outright hallucinations.

Because generation is probabilistic, there is no fixed ranking to check once. Your visibility is a rate: how often, and how well, the model names you across many runs of the questions your customers ask.

How to make your brand easier for LLMs to describe

• Say plainly what you are. State your category, product and audience in simple sentences on your own site, not only in slogans.
• Keep facts consistent. Names, prices, features and locations should match across your site, profiles, directories and press.
• Earn third-party coverage. Models learn from, and retrieve, what reviews, comparisons and articles say about you, not only what you say about yourself.
• Let AI crawlers in. A blocked crawler cannot collect or retrieve your pages. The free AI crawler checker shows which ones your site allows.
• Measure the answers. Ask the questions your customers ask, on the engines they use, and track the results over time.

Common mistakes

• Trusting one answer. A single chat session is one sample of a variable output. Repeat the question before drawing conclusions.
• Treating the LLM as a search index. The model does not store your pages and rank them. It writes an answer, sometimes from memory and sometimes from retrieved sources.
• Assuming every engine is the same. ChatGPT, Perplexity and Google use different models, retrieval systems and instructions, so your brand can do well on one and be absent on another.

How Mencoro tracks what LLMs say about your brand

Mencoro asks ChatGPT, Perplexity, Google AI Overview and Google AI Mode the prompts you choose, in the countries you choose, and records the text of each answer and the sources it returns. Because LLM answers vary, each check can run 1 to 3 passes of the same question to smooth out the noise. A language model then finds every brand named in the answer and classifies each mention by type and tone, and each mention is checked against the visible text. The result is your Coverage, Favorability and Share of voice per engine, explained step by step on how Mencoro works. See the AI brand monitoring feature, or run a quick sample with the free AI visibility checker.

Related glossary terms

• Retrieval-augmented generation (RAG): how AI engines add fresh sources to what the model already knows.
• Knowledge cutoff: the date after which a model’s training data stops.
• AI hallucination: when a model states something false as fact.
• AI visibility: how often and how well your brand appears in AI answers.