Large Language Model (LLM)
A large language model (LLM) is a neural network, typically a transformer, trained on enormous volumes of text to predict the next token in a sequence. That single skill, scaled up with billions of parameters and trained on trillions of tokens, is enough to produce fluent answers, summaries, and conversation. ChatGPT, Claude, Gemini, and similar assistants are products built around one or more LLMs; the model itself is the engine, not the app.
Why it matters for your visibility
People increasingly put their questions to an LLM-powered assistant instead of typing them into a search box. ChatGPT alone reported 900 million weekly active users in February 2026, Google's Gemini app passed 1 billion monthly active users in August 2026, and Google AI Overviews now reaches more than 2 billion monthly users across 200+ countries. Whichever businesses the underlying LLM names in its answer get seen, with no click required at all. Pew Research found that when an AI summary appears above Google's results, people click through to a normal search result in only 8% of visits, versus 15% when there's no summary, and Ahrefs measured a 58% average drop in click-through rate for the #1 organic result once an AI Overview shows up. Getting cited inside the model's answer is becoming the visibility event that matters, which is exactly what generative engine optimization is built to influence.
How an LLM works
An LLM splits text into tokens, small chunks that can be a word, part of a word, or punctuation. During training it processes huge quantities of tokens drawn from books, websites, code, and forums, adjusting billions of internal parameters so it gets better at predicting the next token given everything before it. The architecture behind almost every modern LLM is the transformer, introduced in 2017, which uses an attention mechanism to weigh how strongly each word in a passage relates to every other word. At inference time the model still isn't doing anything more exotic than repeatedly predicting the most probable next token; it has no built-in fact database and no understanding in the human sense. That's why an LLM's output can be extremely fluent while still being wrong.
Common mistakes and limits
An LLM can state something confidently and still be inventing it, a failure mode known as hallucination. Columbia Journalism Review's Tow Center tested eight AI search tools on 1,600 queries asking them to identify a news article's real source, headline, date, and URL, and got more than 60% incorrect responses across the board, with ChatGPT alone misidentifying 134 of 200 articles. Models also carry a training cutoff and don't know about anything after it unless they're paired with live web retrieval. The common mistake on the optimization side is assuming there's a technical trick to fix this: Google's own guidance states plainly that no special markup, schema, or AI-specific files are required for AI Overviews or AI Mode, and explicitly warns against writing separate content "for AI." The more reliable lever is making the facts about your business, correct pricing, correct hours, correct claims, so consistently available across the web that a model has little room to invent something else.
Relation to generative engine optimization
Generative engine optimization, first formalized in the 2024 paper "GEO: Generative Engine Optimization" by researchers from Princeton, IIT Delhi, and independent contributors, is the practice of shaping what an LLM has to work with so it cites you accurately. You can't retrain the model, but you can influence its inputs: your own site content, structured data, and, more powerfully, how often and how consistently other sites mention your brand. Ahrefs' analysis of roughly 75,000 brands found that mention frequency correlates with AI citation rate at about 0.664, roughly three times stronger than backlink count at 0.218. It's also worth knowing that an LLM's citation behavior doesn't mirror traditional search rankings: Ahrefs found only 6-8% overlap between URLs ChatGPT cites and Google's top 10 for the same query, and about 80% of ChatGPT's citations don't rank in Google's top 100 at all. Optimizing for one doesn't automatically optimize for the other.
Example
Imagine a small tax advisory firm in Leipzig looking for new clients. A prospect used to type a query into Google and click through the results. Now they ask an AI assistant: "Who can help me in Leipzig with a tax return for the self-employed?" The large language model behind that assistant phrases an answer and names specific firms. Whether this particular firm is one of them depends on how well the model's underlying sources describe its services: clear service pages, consistent address and contact details across the web, and articles that answer the exact questions a self-employed client would ask. Get that groundwork right and the firm gets recommended without ever running an ad.
Common questions
Is a large language model the same thing as ChatGPT?
No. ChatGPT is a product built around an LLM, complete with an interface, memory features, and safety layers. The large language model is the underlying engine that actually generates the text. The same model can power more than one product, and a single product can switch between several models.
Can I influence what an LLM says about my company?
Not by retraining the model directly. What you can influence is its raw material: clear, accurate, consistently repeated information about your business across your own site and third-party mentions. That's the same lever generative engine optimization works with, and it matters more than technical add-ons like schema markup or an llms.txt file, which Google has confirmed its systems don't read or act on.