What Is an LLM? Large Language Models, Explained for Brands
What is an LLM?
An LLM (large language model) is a neural network trained on enormous amounts of text to do one thing extremely well: predict the next word. Ask it a question and it does not look the answer up in a database. It generates the most plausible continuation, word by word, based on everything it absorbed during training. ChatGPT, Gemini, Claude and Grok are all products built on LLMs.
That single fact (prediction, not lookup) explains most of what marketers find confusing about AI answers.
How an LLM actually produces an answer
Training compresses patterns from billions of pages into model weights. When you ask "what's the best CRM for a small agency?", the model is not consulting a list. It is reconstructing what an answer to that question tends to look like, including which brand names tend to appear in it. Brands that were written about often, consistently and positively are simply more likely to be generated.
Modern assistants add a second step: retrieval. Before answering, they can run a live search and feed the results into the model: that is RAG and, in Google's case, grounding. This is where recent content and smaller brands get their chance.
LLM vs search engine
A search engine ranks documents and shows you links. An LLM writes the answer itself and mentions a handful of names inside it. The practical difference for a brand: on a results page you can be result eleven and still exist; in a generated answer you are either in the two or three names the model writes, or you are nowhere. That is why AI visibility is measured differently from rankings.
Where an LLM's knowledge of your brand comes from
Two places. Training data: the public web as it existed when the model was trained (old, frozen, and biased toward sources that wrote about you years ago). Retrieval: what the assistant's search step finds today. You cannot edit the first. You can absolutely influence the second, which is why most GEO work is really source work: the pages, reviews and reference sites an engine reads when your category comes up.
What LLMs get wrong
Because generation is probabilistic, an LLM will sometimes state things about your brand that are simply false: a wrong price, a feature you never had, a founder who does not exist. These are hallucinations, and they read exactly as confidently as the true statements around them. Brands that never check what models say about them tend to find out from a customer.
What a marketer should actually do about LLMs
Not prompt engineering. Three things: know which questions in your category trigger brand recommendations, measure whether you appear in the answers across engines, and work the sources that feed those answers. The measuring part is what prompt tracking does; the source part is mostly off-site GEO. Everything else about LLMs is interesting but not actionable.
The LLM landscape brands actually meet
The names buyers use are products; the models sit underneath. ChatGPT runs on OpenAI's GPT family. Gemini is Google's model wired into Search through grounding. Claude is Anthropic's, Grok is xAI's, and Perplexity is not an LLM itself: it is a search product that borrows several of these models under the hood. For visibility work the practical split is not by vendor but by behavior: engines that search before answering can discover you today, while pure-memory modes only know the web as it was at training time.
That split is why measuring one engine tells you little about another. Each product pairs a different model with a different search backend and different citation habits. The same question about your category returns differently sourced answers on each. Treating "AI" as one channel is the most common measurement mistake we see; measuring AI visibility exists because the engines genuinely disagree.
Common misconceptions about LLMs
"The model updates itself in real time." It does not. Weights are frozen between releases; freshness comes only from the retrieval step, which is why RAG matters more to your visibility than model release cycles do.
"It learned about us from our chats." Consumer conversations are not a reliable path into anyone's answers. What models repeat at scale is the public written record: pages, reviews, comparisons, reference sites.
"You can optimize the model." You optimize what the model reads. Every durable GEO tactic (off-site coverage, consistent facts, extractable pages) is source work wearing an AI name.
"LLMs killed search." The engines answering your buyers are search systems with a language model on top. The retrieval half still runs on crawling, indexing and ranking, which is why classic technical SEO quietly decides a surprising share of AI searchability.