What is a large language model, and how does it answer?

A large language model is software trained on a vast amount of text until it can write new text that answers a question. LLM is its short form, and engines such as ChatGPT, Gemini and Claude are built on one. The short definition is in the glossary.

OpenAI explains how it works in plain terms: during training the model learns how words typically appear together in context, then uses that understanding when it writes a response to predict the next most likely word, one word at a time. It adds that the model does not store copies of the text it was trained on; instead its internal values are adjusted as it learns.

So it does not open a memory holding your page and copy it out. It writes from patterns it has learnt, which is why it can produce a sentence that is fluent and wrong, what is called an AI hallucination.

What it knows from training is fixed up to a certain date; what it reads at the moment of the question is as current as the pages it finds. The engines your customer asks combine the two: they write from training when that is enough, and search the web when they need current information or a source.

OpenAI says ChatGPT may search the web automatically when a question would benefit from current information, and that responses that use web search may include citations. Google says grounding in its search lets Gemini cite verifiable sources beyond its knowledge cutoff.

That principle, retrieving pages and then writing from them, is called retrieval-augmented generation, and in AI engines it takes the form of grounding. Either way, your site can only be cited when the engine reads your page at the moment of the question.

A knowledge cutoff is the last date a model's training data reaches; whatever happened after it, the model does not know on its own. Anthropic publishes two dates for each Claude model, a training data cutoff and a reliable knowledge cutoff, and defines the second as the date through which the model's knowledge is most extensive and reliable.

If you open a branch, change your prices or launch a service after that date, the engine learns of it only if it searches and finds a page that says so. Otherwise the model may write, from memory, an old price or an address that has since changed, which is why everything that changes in your business needs a current, dated page.

Because the road to a citation runs through search, not through training. Four things follow:

• The link comes from search: a model writing from its training alone does not read your page; the link to it comes when the engine reads it at the moment of the question.
• The page must be readable: what the crawler cannot reach cannot be cited, as in can AI engines read your site, and how do you check?
• A clear sentence travels more easily: what we see in our measurements is that the engine usually quotes the shortest passage that answers, so write an answer paragraph under every question, as in how do you write a paragraph AI engines will quote?
• A date beats memory: a current, dated page gives the engine something newer than its training, and nobody knows the age of a page with no date.

An illustrative example: a restaurant chain opens a branch in Dubai Marina after the model's knowledge cutoff, and a customer asks where the restaurant's nearest branch in Dubai Marina is. If the model answers from its training alone, it may not know the branch exists; if the engine searches the web, it can find the branch page, provided it exists, is indexed and states the address and hours as text. One page can be the difference between a wrong answer and one that cites you.

In our measurement of Arabic UAE buying questions in October 2026, the answers came in Arabic, with business names often left in Latin letters; the likeliest explanation is that the engine writes a name the way it found it in its sources. So write your business name in Arabic and in Latin letters on your page. What makes an engine understand that the two names are one business is in what is an entity in search?

The model is the writer; the engine is the product your customer reaches, which combines the model with search, sources and the way they are shown. ChatGPT, Gemini and Perplexity are engines, each with one or more models behind it; how Gemini builds its answer on Google Search is in how Gemini chooses its sources in Arabic. The definition of an engine is in what is an AI engine?, and the eight we measure are in the AI engines we measure in Arabic.

That is why you measure the engine, not the model: your customer does not ask a bare model, but opens ChatGPT or Gemini and reads what it wrote after searching or not. The work of getting an engine to cite your page is called optimising for AI answers (GEO).