Module 2 · How AI Systems Find Information Technical Foundations

Lesson 1 of 10

Large language models can produce remarkably fluent answers, but they do not work like a conventional database or search engine. Understanding the difference is essential before you can understand how AI systems discover, interpret and surface information.

At the centre of an LLM is prediction.

A large language model processes language as units called tokens. Given the tokens already present in its context, the model estimates which token is most likely to come next. It repeats that process again and again until it has generated an answer.

An LLM does not normally retrieve a complete pre-written answer from somewhere inside itself.

It constructs an output token by token using patterns represented within the model and any additional context supplied at the time of the request.

A simplified journey through the model

01User input

02Tokenisation

03Context processing

04Probability calculation

05Generated output

Tokens

Text is broken into smaller units the model can process mathematically. A token may represent a word, part of a word, punctuation or another text fragment.

Context

The model considers the information available within its current context: the user's question, instructions, previous conversation and potentially retrieved information.

Attention

Transformer models use attention mechanisms to identify relationships between different parts of the available context and determine which information is most relevant to processing the next output.

Probability

The model produces probabilities for possible next tokens. Generation continues by repeatedly selecting tokens according to the model's configuration and the surrounding context.

A model is not the same thing as a search index.

Model knowledge

Encoded through patterns learned during training
Does not behave like a catalogue of stored web pages
May be incomplete or outdated
Can generate plausible statements that are wrong

Retrieved information

Obtained from an external source at query time
Can provide fresher or more specific evidence
May include web pages, databases or documents
Can be supplied to the model as additional context

Why this matters for GEO and AEO

If an AI answer is being generated entirely from patterns already represented within a model, changing your website today does not instantly rewrite the model's underlying knowledge.

But many modern AI-search experiences can also retrieve external information. In those environments, the accessibility, clarity, authority and relevance of your current web content can become far more important.

That distinction explains why effective AI search optimisation cannot be reduced to one tactic. You need to understand both the information environment surrounding your brand and the retrieval systems that may bring current information into an AI answer.

Practical challenge

Take one question a customer might ask an AI assistant about your market. Write down which parts of an ideal answer could come from general background knowledge and which parts would require current, specific or business-level information.

Key takeaway

Large language models generate language probabilistically. Modern AI-search systems may combine that capability with external retrieval, which is why understanding the difference between model knowledge and retrieved information is fundamental to GEO and AEO.

Next: Training Data vs Live Retrieval →