What is a large language model (LLM)?

The short answer

A large language model is a prediction machine for words. You train it on a giant pile of text, and it learns the patterns well enough to guess what comes next, one token at a time. String enough of those guesses together and you get something that answers questions, summarizes a report, drafts an email, or explains a tax rule. GPT, Gemini, Claude, and Llama are all LLMs. They are the engines under ChatGPT, Perplexity, Google AI Overviews, and most chatbots you have talked to this year.

Here is the part that matters for a brand. When one of these models answers a real question, it has to pick which sources to lean on. That pick is your whole game now. Rank high, rank low, doesn't matter if the model never quotes you.

How an LLM actually works (without the math)

You do not need the math. You need the intuition.

1. Training. The model reads an enormous amount of text and learns statistical patterns: which words tend to follow which, how ideas connect, what a good answer looks like. This is frozen knowledge, and it has a cutoff date.
2. Prediction. When you ask something, the model generates a response token by token, each one a best guess given everything so far. It is not looking anything up. It is reconstructing from patterns.
3. Grounding (the new bit). Modern AI search bolts a live retrieval step on top, so the model fetches fresh web pages and writes its answer from those. That mechanism is retrieval-augmented generation, and it is why AI engines can cite a page that did not exist when the model was trained.

So an LLM on its own is a closed book with a great memory and a bad sense of dates. An LLM plus retrieval is the thing that reads your page and decides whether to quote you.

Why this is different from a search engine

Old search ranked pages and handed you ten blue links. You clicked. An LLM does something else: it reads, synthesizes, and answers in its own words. There is no spot number one to win. Ask the same question twice and you can get two different answers, with two different sources cited.

That last row is the whole shift. The job moved from "be clickable" to "be quotable." For the wider playbook, our AI SEO guide and LLM SEO breakdown cover the tactics.

Why brands should care which LLM cites them

An LLM is the gatekeeper between your content and the person asking the question. If the model answers and never names you, you are invisible to that person, even if you would have ranked first in Google. And because the answer is generated fresh each time, you cannot just check once and relax.

What makes an LLM more likely to quote you? Mostly the boring, honest stuff:

• A direct answer up top. Models lift clean opening lines. Bury the answer under a wall of throat-clearing and they skip you.
• Real facts with named sources. Stat-dense, sourced pages get cited more. Vague filler gets ignored.
• Structure a machine can parse. Headings, short paragraphs, lists, and clear question-and-answer chunks.
• Being talked about elsewhere. Mentions on trusted sites and communities feed the patterns the model learned and the pages it retrieves.

If you want the mechanics behind the pick, how AI engines choose sources walks through it. Fair warning, and we will be honest here: nobody outside the labs fully knows the exact weighting. The signals above are what the public research and our own testing point to, not a leaked formula.

The catch: LLMs are confident and sometimes wrong

One honest limit. LLMs predict plausible text, and plausible is not the same as true. They will state a wrong fact with total confidence, invent a citation, or misremember a number. That is a hallucination, and it is baked into how prediction works, not a bug someone forgot to fix. Retrieval reduces it. It does not erase it. So if a brand's facts are messy or thin online, an LLM is perfectly capable of getting your details wrong in front of a customer.

Measuring whether LLMs cite you

Because LLM answers vary run to run, you cannot eyeball this once. You have to sample across many runs and look at the rate, not a single lucky hit.

That is what AI Citation Monitor does. It checks whether five engines (ChatGPT, Perplexity, Gemini, Google AI Overviews, and Microsoft Copilot) actually cite or recommend you, and it reports the result with confidence intervals so you are reading a real signal and not noise. It also shows competitor share of voice, which sources each engine pulls from, and prescriptive fixes for the pages that get read but never quoted. There is a free instant check, and plans run from Free at $0 up through Starter, Growth, and Agency.

LLMs are not a fad you can wait out. They are how a growing share of people ask questions now. So the move is plain: write the kind of clear, sourced, quotable page a model loves to lift, then measure honestly whether it is working.

FAQ