What is a Large Language Model (LLM)?

At a glance

• An LLM predicts text. Every answer, summary and recommendation is that one operation repeated.
• GPT, Claude, Gemini and Llama are all large language models, with different training data and different habits.
• A model alone has no access to the live web. Search is a separate step bolted on around it.
• The same prompt gives different wording twice, because the prediction is sampled rather than looked up.

Prediction, one fragment at a time

The model reads everything so far, the question plus whatever it has already written, and produces a distribution over what could come next. One fragment is picked. The whole thing runs again with that fragment added.

There is no plan and no draft. A model does not settle on a conclusion and write towards it. Coherence across a long answer comes out of each step being conditioned on every step before it.

The capability came from scale and from an architecture that weighs every part of the input against every other part, set out in Attention Is All You Need. Text is split into tokens and mapped into numbers, the representation behind an embedding. Nothing here is a lookup, and no stored copy of your site is being consulted.

Fluent by default, accurate only sometimes

Fluency and accuracy come from different places, and only one of them is guaranteed.

A model produces a clean, confident paragraph about your pricing whether or not it has seen your pricing page. Grammar is learned reliably. Facts about a mid-sized company are learned patchily, from whatever the training data happened to hold, and the model cannot tell the two apart.

This is why retrieval matters so much for brands. When an answer engine fetches your live page and puts it in front of the model, the model is summarizing a real document. With nothing fetched, it reconstructs you from a blurry memory, and the output looks identical either way. Your content is not competing to be memorised. It is competing to be the document handed over at the moment somebody asks.

Give a model less room to improvise

You cannot change the model. You can change how much guessing it has to do.

1. Put the fact in one plain sentence. “Plans start at 49 dollars a month.” A model summarizing that has nothing to invent. A pricing table rendered as an image gives it everything.
2. Keep claims next to their qualifiers. If a price depends on annual billing, say so in the same sentence. Split across two sections, the qualifier gets dropped in the summary.
3. Cut the marketing phrasing. Superlatives and vague benefit copy compress badly. A model paraphrasing “the fastest platform on the market” returns something you would not have signed off.
4. Repeat the core facts on every page that gets fetched. Product, pricing, about. Each stands alone, because each gets retrieved alone.
5. Read your page as a stranger would. If a person skimming for thirty seconds could get your category wrong, so can a model.