What is a large language model?

Google's machine learning course starts with the smaller idea: "A language model estimates the probability of a token or sequence of tokens occurring within a longer sequence of tokens." A token "could be a word, a subword (a subset of a word), or even a single character" (Google ML crash course: introduction to LLMs, last updated 2026-01-09, read 1 November 2026).

A large language model does the same thing at a much bigger scale. Google's short version: "LLMs are essentially autocomplete mechanisms that can automatically predict (complete) thousands of tokens." Most of them use the transformer design, which Google calls "the state-of-the-art architecture for a wide variety of language model applications" (Google ML crash course: LLMs and transformers, last updated 2026-01-02).

How does an LLM work?

  1. Training: the model reads a very large amount of text and learns which tokens tend to follow which.
  2. Prompt: you send it text, such as a question.
  3. Prediction: it predicts the next token, adds it, and repeats until the answer is complete.
  4. Tools (optional): in many assistants the model can search the web first and answer from what it finds. This works like retrieval-augmented generation.

What can't an LLM do on its own?

Training stops at a cutoff date. Without search, the model does not know what happened after it. Google's Gemini docs say grounding with Google Search lets the model "cite verifiable sources beyond its knowledge cutoff" (Gemini API docs, last updated 2026-10-28).

It can also be wrong with confidence. Google's course says it plainly: "LLMs hallucinate, meaning their predictions often contain mistakes." See hallucination.

Why do LLMs matter for B2B SaaS?

Buyers now ask LLM-based assistants for software advice. G2's 2026 survey of 1,076 B2B decision-makers found that 51% of B2B software buyers now start their research in an AI chatbot more often than in Google.

What an assistant says about your product comes from two places: what the model learned in training, and what it finds when it searches. You cannot edit the first. You can work on the second: let the search crawlers in, publish clear answers, and be named on the pages assistants cite. When Prefer checked 188 answers from ChatGPT, Perplexity, Gemini and Claude in July 2026, 40% of the sources cited for buyer questions were independent lists.

That work is LLM optimization. The LLM SEO guide covers it in full.

What does this look like? (hypothetical example)

A buyer asks an assistant with web search off: "What does Acme Payroll cost?" Acme Payroll is a made-up product. The model answers from training data, so it may give a price from before its cutoff. With web search on, the same model finds the current pricing page and a recent comparison article and cites both. Same model, two different answers.

  • An AI assistant, such as ChatGPT or Claude, is a product. The LLM is the model inside it, plus tools like web search.
  • Generative AI is the wider term. It also covers models that make images, audio or video.
  • Retrieval-augmented generation adds a search step before the LLM writes, so the answer can use current sources.
  • A hallucination is a confident answer that is wrong. All LLMs can produce them.