What a large language model means
A large language model (LLM) is a neural network trained on enormous quantities of text — web pages, books, reference works, forums, code — to do one deceptively simple thing: predict what text comes next. At sufficient scale, that objective produces a system that can answer questions, summarize, reason through problems, and hold conversations in fluent natural language. LLMs are the technology inside ChatGPT (OpenAI's GPT models), Google's Gemini, Anthropic's Claude, and every other engine now answering your customers' questions.
"Large" refers to both the model — billions to trillions of learned parameters — and the training data it absorbed. The key mental model: an LLM doesn't look facts up in a database. It generates answers from statistical patterns compressed into its parameters during training, which explains both its fluency and its characteristic failure mode.
How LLMs work, briefly
An LLM's life has two phases:
- Training — the model reads its corpus and adjusts parameters to predict text well, absorbing language, facts, and associations up to a knowledge cutoff date. Later stages (fine-tuning and alignment) shape it into a helpful assistant.
- Inference — the trained model generates answers to user prompts, working within a context window that bounds how much text it can consider at once. Modern engines often augment this step with live web retrieval so answers aren't limited to training-era knowledge.
Generation is probabilistic: the same question can produce differently worded answers — and different brand lists — from run to run.
Why LLMs matter for marketing
Everything a marketer cares about in AI search reduces to LLM behavior. When an engine recommends products in your category, an LLM is drawing on two sources: brand associations memorized from training data ("this name keeps appearing next to 'best CRM'"), and whatever live sources were retrieved at answer time. That has three practical consequences:
- Your brand's web footprint is now model input. Years of coverage, reviews, and mentions determine what models "know by heart" — the mechanics are covered in how AI models know about your brand.
- LLMs can be confidently wrong. Pattern-based generation produces hallucinations: fluent, plausible, false claims about prices, products, and companies.
- Visibility is measurable but statistical. Because outputs vary, understanding how LLMs present a brand requires repeated sampling across engines — the premise behind prompt tracking.
LLMs and adjacent terms
An LLM is the model; an answer engine like ChatGPT or Perplexity is the product wrapping a model with retrieval, tools, and an interface. Generative engine optimization is the practice of influencing what those products say about a brand — which ultimately means influencing what LLMs read, both during training and at answer time.