What is a large language model (LLM)?
A large language model (LLM) is an AI model trained on a very large amount of text to predict and write language, and it is the part of an AI assistant that writes the answer. ChatGPT, Claude, Gemini and the model behind Perplexity's answers are all built on large language models. The model writes the answer; search, if the engine runs one, only decides what the model gets to read first.
How does a large language model work?
A large language model works by predicting the next piece of text, one token at a time, from patterns it learned in training on a very large body of text. Today's models descend from the transformer architecture described in 2017 by Vaswani and colleagues in Attention Is All You Need.
What it learned is fixed at a knowledge cutoff: the model knows the web as it was when training data was collected, and nothing after. Anything newer reaches the answer only if the assistant searches the web and hands the results to the model while it writes.
What does an LLM know about your brand?
Whatever the training data said about you, blended and compressed: your category, your best-known claims, and how often other pages mentioned you. It does not know what changed since its cutoff, and it cannot tell you which page it learned a fact from.
That is why a brand can be described confidently and wrongly, and why the order of work matters: fix the pages engines can look up now, because what the model remembers changes only when a newer model ships (the post).
What does this look like in a real example?
In the public Tesla example (Oct 4, 2026), Claude answered "What is Tesla?" and "Is Tesla legit?" from what the model already knew, with no sources cited. A famous brand is well covered in training data, so a memory answer can be good.
On buyer questions the same engine searched: across 25 buyer questions Claude gave 351 citations, only one of them to tesla.com, and named Tesla in 11 answers. The model and the pages it read worked together, and the pages decided much of the shortlist.
Why does the same LLM give different answers?
Because generation involves sampling, and serving adds its own variation. OpenAI states that Chat Completions are non-deterministic by default, so outputs may differ from request to request (OpenAI API docs, read Oct 15, 2026).
For brands that means one answer is one draw. A fair reading of how a model treats you takes repeated questions and an interval on every figure; see AI answer sampling.
Which LLMs does Proofsource track?
Proofsource asks the consumer-facing engines buyers use, each built on its vendor's models, and records which model version answered where the engine reports it. See engines and models and how each engine chooses brands.
Common questions about Large language model (LLM)
Is ChatGPT an LLM?
ChatGPT is an assistant built on OpenAI's large language models, with tools such as web search around them. The LLM writes the text; the assistant decides whether to search first.
Can I update what an LLM knows about my company?
Not directly. What a model remembers changes when a newer model is trained. What you can change now is the pages engines look up when they search, which reach the answer much sooner.
What is a knowledge cutoff?
The date after which a model has no training data. Vendors publish it on their model pages, and it is often months before the model's release.
Do LLMs copy text from my website?
They write new text. When an engine searches, the answer can follow a retrieved page closely and cite it; when it answers from memory, it draws on patterns from training rather than a specific page.
Related terms
- Training data vs live search
Training data is what a model learned before its knowledge cutoff; live search is what it looks up on the web while it answers, and one AI answer can draw on either or both.
- Retrieval-augmented generation (RAG)
Retrieval-augmented generation (RAG) is a method where a language model first retrieves relevant documents and then writes its answer with them in view.
- AI hallucination
A hallucination is a confident statement in an AI answer that is false or unsupported, such as a wrong price, a feature that does not exist or a source that was never written.
- Answer engine
An answer engine is a search tool that replies to a question with one written answer, usually with sources, instead of a ranked list of links.
Worked example: Tesla, public AI answers, Oct 4, 2026: Proofsource's first scan of Tesla, 25 buyer questions and 5 brand questions, one answer per question on each of four engines. A small sample from one brand, shown as an example, not a benchmark.
See where AI leaves you out, and what to fix.
See which buyer questions name your brand, who gets named instead, and what to fix. 200 answers. 25 Prompts/Question. Top AI engines. 2 days. Free. No card.
Google AI OverviewsGoogle AI Mode
