What is an LLM (large language model)?

LLM stands for large language model, and here it has nothing to do with the law degree written LL.M. It is the technology underneath ChatGPT, Gemini or Claude: the part that writes. A store cares for one concrete reason: every assistant that talks about your products is an LLM writing, sometimes from what it reads on your product page and sometimes from what it thinks it remembers about it. What the model is doing in each case explains almost every mistake you will see in its answers.

How it works

A language model estimates the probability of a token, or a sequence of tokens, appearing within a longer sequence. A token can be a word, a piece of a word or even a single character.⁠^{[1]Google for Developers, Machine Learning. Introduction to Large Language Models} An LLM does the same thing at scale: it predicts tokens one after another, sometimes whole paragraphs, like a very sophisticated autocomplete.⁠^{[2]Google for Developers, Machine Learning Crash Course. LLMs: What's a large language model?} What makes it "large" is that it has far more parameters than earlier models and makes use of far more context, and the architecture that dominates today is the Transformer, with a self-attention mechanism that weighs how much each word matters relative to the others.⁠^{[2]Google for Developers, Machine Learning Crash Course. LLMs: What's a large language model?}

Context is what lets it disambiguate. Google gives an example: the surrounding words help the model decide whether orange is a fruit or a color.⁠^{[1]Google for Developers, Machine Learning. Introduction to Large Language Models} Training one takes enormous amounts of text, months of work and a great deal of computing power.⁠^{[2]Google for Developers, Machine Learning Crash Course. LLMs: What's a large language model?}

Why it makes things up

OpenAI explains where hallucinations come from, the plausible but false sentences a model generates. In pretraining, the model learns to predict the next word across enormous amounts of text, and that text carries no true or false labels: it only sees examples of fluent language.⁠^{[3]OpenAI (September 5, 2025). Why language models hallucinate} Regular patterns, like spelling or closing a parenthesis, are learned well as the model grows; arbitrary, infrequent facts, like a pet's birthday, cannot be predicted from patterns, and that is where hallucinations come from.⁠^{[3]OpenAI (September 5, 2025). Why language models hallucinate} On top of that, the usual evaluations reward getting it right more than admitting you don't know, so the model learns to guess.⁠^{[3]OpenAI (September 5, 2025). Why language models hallucinate}

For a store, the consequence is direct: the exact weight of a backpack, the size of a glove or the wattage of a light are arbitrary, infrequent facts. They are the kind of fact a model cannot deduce and, if it does not read it, may fill in with something that sounds right. Google says it in its machine learning course: LLMs hallucinate, and their predictions often contain errors.⁠^{[2]Google for Developers, Machine Learning Crash Course. LLMs: What's a large language model?}

What it reads versus what it remembers

An LLM has two sources of information. The first is what it learned in training, frozen in time and without sources. The second is what it is given to read at the moment of answering. AI search uses the second: Google describes retrieval-augmented generation (RAG) as a technique to improve the quality, accuracy and freshness of answers by retrieving pages from its index.⁠^{[4]Google Search Central. Optimizing your website for generative AI features on Google Search} When the model reads your page, what matters is that the fact is there and can be read: Microsoft warns that AI systems can skip content hidden in tabs or dropdown menus, and that key data should be in HTML and not only inside an image.⁠^{[5]Microsoft Advertising (Krishna Madhavan, October 8, 2025). Optimizing Your Content for Inclusion in AI Search Answers}

Even when it reads, the model rewrites. OpenAI explains that ChatGPT can generate simplified titles and product descriptions from what it receives, and that the tags it adds to products, like "Budget-friendly", are generated by the model and are not verified claims.⁠^{[6]OpenAI Help Center. Shopping with ChatGPT Search}

In a store

A fictional example: a camera backpack whose page says "20-liter capacity" and "water-resistant fabric", and nothing more about weight or rain. This is what can happen to each fact when an assistant answers about it:

The defense is not writing for the AI, but putting the specific fact in text, with its unit, on the page. Clione works with the model the store configures (OpenAI, Anthropic, Gemini or Ollama, with its own key) and, in each enrichment, passes it the merchant's answers about the product: what it is for, what it is made of, what warranty it has.

How to check it

1. Ask about a specific fact of your product to an assistant without search and then with search. Compare the two answers with your page and note the differences, with the date.
2. Go through the page looking for facts that live only in images, PDFs or closed tabs. Move them into the page text, with their unit.⁠^{[5]Microsoft Advertising (Krishna Madhavan, October 8, 2025). Optimizing Your Content for Inclusion in AI Search Answers}
3. Review ambiguous attributes: "water-resistant" versus a waterproofing standard, "lightweight" versus a weight. The model cannot guess the figure you did not write.⁠^{[3]OpenAI (September 5, 2025). Why language models hallucinate}
4. Look at how ChatGPT summarizes your products when they show up in its shopping answers: the title, tags and review summary are written by the model.⁠^{[6]OpenAI Help Center. Shopping with ChatGPT Search}
5. If an assistant shows a false fact about your product, report it. In ChatGPT, products with incorrect information can be reported from the product's own menu.⁠^{[6]OpenAI Help Center. Shopping with ChatGPT Search}

Related concept: RAG

An LLM is the model; RAG is a way of using it. The LLM only has what it learned in training. With RAG, the system first retrieves documents and then the LLM writes with them in front of it, which improves the accuracy and freshness of the answer and makes it possible to cite sources.⁠^{[4]Google Search Central. Optimizing your website for generative AI features on Google Search} ChatGPT with search, AI Overviews and AI Mode are systems of this kind: an LLM plus a retriever. When someone says "the AI read my site", they are almost always talking about the second step.

To keep going, the ChatGPT Search entry explains how ChatGPT searches, and the AI search reading path puts the rest of the terms in order.