What is a large language model (LLM)?

A large language model is a neural network trained on a very large body of text. Its core job is narrow: given some text, predict the most likely next token. A token is a word or a fragment of one. Repeat that prediction thousands of times and you get a paragraph, an answer or a product comparison.

"Large" refers to the size of the training data and the number of internal parameters the model adjusts during training. "LLM" and "large language model" mean the same thing. This entry is a short definition. For how LLMs find and use web content, read the guide to LLM SEO.

An LLM can produce an answer in two ways.

From memory

During training, the model absorbs patterns from text. What it retains is called parametric knowledge. It stops at a knowledge cutoff. OpenAI's model documentation lists a February 16, 2026 cutoff for the GPT-5.6 family and August 31, 2025 for the Instant path behind chat-latest. So "what ChatGPT knows" is not one fixed thing.

From live search

When retrieval is on, the app runs searches, fetches pages and places their text in the model's context window. The model then writes an answer using that text and may cite it. OpenAI's help center says ChatGPT search typically rewrites your question into one or more targeted queries before sending them to search providers.

Whether search fires is decided per query. As of October 2026, ChatGPT searches at the model's discretion, and Claude.ai search is a user toggle. On many APIs, web search is something a developer has to switch on, as covered in do AI APIs search the web by default.

The two modes give you two different brand stories.

• Memory answers reflect old training data. If you changed your pricing or launched a product after the cutoff, a memory answer cannot know. Nothing you publish this quarter reaches it.
• Search answers can reflect your current site, but only if the model's crawlers can reach and read it.

The second mode is where you have levers. OpenAI's crawler documentation says OAI-SearchBot controls whether a site appears in ChatGPT search, while GPTBot governs training only, and each setting is independent. Vercel found that ChatGPT and Claude crawlers fetch JavaScript files but don't execute them, so content that only appears after JavaScript runs may never reach the model.

The next move: check that retrieval crawlers are allowed in robots.txt, that your firewall doesn't block them, and that key facts like prices sit in the raw HTML. Then compare answers per platform, as described in LLM visibility tracking.

• "ChatGPT is one model." Apps route between models with different cutoffs and different search behavior.
• "The LLM reads my website every time." Only when retrieval fires. Otherwise it answers from memory.
• "Allowing GPTBot gets me cited." GPTBot is a training crawler. ChatGPT search uses OAI-SearchBot.
• "I can fine-tune ChatGPT to know my brand." Fine-tuning changes a copy of a model that a developer controls. It does not change what the public app says.
• "Special phrasing makes LLMs cite me." A 2026 critical survey of GEO studies found no technique with a stable, cross-platform causal effect on discoverability.

• Parametric knowledge: what a model knows with no web access.
• Context window: the text a model can read while answering.
• Fine-tuning: extra training on a narrower dataset.
• How a brand ends up in training data: the slow route into memory answers.

What does LLM stand for? LLM stands for large language model. It describes an AI system trained on large amounts of text to predict the next token in a sequence. Chat assistants such as ChatGPT, Claude and Gemini are products built on top of one or more LLMs, with extra layers for search, safety and conversation handling.

Is an LLM the same as an AI search engine? No. The LLM is the text-generating model. An AI search engine combines an LLM with a retrieval system that runs queries, fetches pages and passes them into the model's context. Without that retrieval layer, the model can only answer from its training data, which stops at a knowledge cutoff.

How can I tell whether an answer came from memory or search? Look for citations. If the answer links to sources, retrieval almost certainly ran. If there are no citations, you usually cannot know for sure, and the answer may come from training data with an older cutoff. Testing the same question several times shows how often each mode appears.

Can I change what an LLM remembers about my brand? Not quickly. Memory changes only when a vendor trains a new model, and no vendor documents how a given page affects that. What you can influence sooner is the search mode: make sure retrieval crawlers can reach your pages and that accurate facts are in the HTML they read.