Large Language Models: Basics for Marketers

Large language model in one sentence

A large language model (LLM) is an AI system that has learned from huge amounts of text to continue language statistically: it predicts the most likely next word each time and thereby generates coherent answers. Exactly such models sit behind ChatGPT, Google Gemini, Perplexity and the AI Overviews – and thus behind the AI search for which you optimize your content.

For GEO you do not have to build an LLM. But you should understand how it "thinks", because from that follows directly why some content is cited and other content is not – and why an AI sometimes claims something false about your brand. This article explains the basics as far as they count for marketing.

How an LLM processes language

An LLM does not understand text like a human. It calculates with probabilities. Put simply: the model has learned from billions of text examples which words typically follow one another. When you ask it something, it assembles word by word the statistically most plausible continuation.

This explains two things at once. First, why LLMs sound astonishingly fluent and convincing – they are trained to phrase things naturally. Second, why they can be wrong without noticing it: a plausible-sounding sentence is not automatically a true sentence. The model optimizes for language flow, not for truth.

Tokens and context windows

Two terms come up again and again, and both are quickly explained.

Tokens are the small units into which an LLM breaks text – often whole words, sometimes word parts or characters. The model works not with letters or sentences, but with these token building blocks. As a rough rule of thumb, a token in German corresponds to roughly a syllable up to a short word.

The context window is the amount of tokens a model can keep "in view" at the same time – that is, the conversation history plus your current question plus the retrieved sources. Once the window is full, older material drops out of view. For GEO this means: your most important statement should stand compact and self-contained, so that it still carries even when only an excerpt of your page makes it into the context window.

Knowledge cutoff: the model does not know the present

An LLM learns its knowledge during training – and this training ends at a fixed point, the knowledge cutoff (the "knowledge cutoff"). Everything that happens after that, the pure model does not know. If you ask an LLM without web search about an event from last week, it simply cannot know anything about it.

That is why modern AI search systems combine the model with a live retrieval from the web. They pull in current sources and let the LLM answer from them. This procedure is called Retrieval Augmented Generation and is the basis for your fresh content being able to land in an answer at all – covered in detail in the article RAG: Retrieval Augmented Generation explained.

Why LLMs reproduce facts incorrectly

The most important point for brands: LLMs hallucinate. That is what it is called when a model invents something false and sounds completely convinced while doing so. This is not a bug in the classic sense, but a consequence of how it works – the model strings plausible words together, even when it lacks the robust information.

Typical triggers are:

• Knowledge gaps: For a niche question or a smaller brand, the training lacks robust data. The model fills the gap with something plausible.
• Outdated knowledge: Without live retrieval it answers at the state of its knowledge cutoff – for example with an old price or a long-changed product name.
• Contradictory sources: If different information about your brand circulates on the web, the model may mix it into a false statement.

What this means for your brand

For your marketing this has an uncomfortable consequence: an AI can claim something false about you – an outdated location, a discontinued product, an invented statistic. And because AI answers sound convincing, users often believe them without checking.

You can influence this above all via the factual basis the AI draws from:

1. Create consistency: Keep central facts – name, services, figures – uniform across your website, directories and profiles. Contradictions on the web raise the hallucination risk.
2. Maintain currency: Actively correct outdated information so the live retrieval finds fresh, correct values.
3. Phrase clearly and self-containedly: Sentences that convey a statement completely and without context leave less room for interpretation – and thus less room for error.
4. Be recognizable as an entity: The more clearly your brand is established on the web as an independent, consistent source, the more reliably the AI assigns information to you.

How generative systems then concretely select and cite from the available material is deepened in the article How LLMs select and cite content. The larger marketing context is placed by Generative AI in marketing.

Conclusion

A large language model generates language by predicting, word by word, the most likely continuation – fluent, but not inherently true. It works in tokens, has a limited context window and a fixed knowledge cutoff, which is why without live retrieval it is neither current nor reliably faithful to facts. Hallucinations are the flip side of how it works. For brands this means: you cannot change the model, but you can change the factual basis it draws from. Consistent, current and clearly phrased content is the best protection against the AI telling something false about you – and the precondition for being cited correctly.