AI Search 101
What is a large language model?
A large language model is a neural network trained on very large quantities of text to predict what comes next. From that single objective it acquires grammar, factual associations and reasoning patterns, which is why one model can translate, summarise, write code and answer questions without being built separately for each task.
The basics
| Abbreviation | LLM |
| Built from | A neural network using the transformer architecture |
| Trained by | Deep learning on large text corpora |
| Examples | GPT, Claude, Gemini, Llama, Mistral |
| Used in | ChatGPT, AI search, AI Overviews |
| Category | A form of generative AI |
How does a large language model work?
The training objective is deceptively simple. Given a stretch of text, predict the next fragment. Repeat across a corpus of hundreds of billions of words, adjusting billions of internal parameters each time the prediction is wrong.
Nothing in that objective mentions grammar, facts or reasoning. They emerge because predicting text well requires them. To finish the sentence "the capital of Finland is" the model needs an association it can only have acquired from the text it read. To continue a mathematical proof it needs something functioning like a procedure. The capabilities are a by-product of compression: the model cannot store the corpus, so it stores structure instead.
Text is handled as tokens, fragments roughly three quarters the length of an average English word. The model converts each into a vector, a list of numbers positioning it in a space where related meanings sit close together. Attention layers then let every token weigh every other token in the input, which is how the model resolves what a pronoun refers to or which sense of a word applies.
Why is the transformer architecture important?
Transformers, introduced in 2017, replaced architectures that read text strictly in sequence. The change that mattered was parallelism: a transformer processes an entire passage at once rather than word by word, which made training on internet-scale corpora practical for the first time.
Everything that followed depended on that. The scale-up from millions to hundreds of billions of parameters, and the capability jumps that came with it, were possible because the architecture could use large numbers of processors efficiently. Every widely used model today is a transformer variant.
What is a context window?
The context window is how much text the model can consider at once, measured in tokens. Everything within it is available; everything outside it does not exist as far as the model is concerned.
Early models held a few thousand tokens, a handful of pages. Current models hold hundreds of thousands, enough for a book or a substantial codebase. This is why a model can now be handed a long contract and asked about clause fourteen, and why an assistant loses the thread of a very long conversation: the earliest exchanges have fallen out of the window.
Why do large language models get facts wrong?
Because generating fluent text and stating true things are different operations, and the model is optimised for the first.
A model has no store of facts to check against. It has statistical associations, and when the associations are weak it produces text that fits the pattern of a correct answer without being one. This is usually called hallucination, and the term flatters the mechanism: the model is not perceiving something false, it is completing a pattern.
Confidence is no guide. The same fluent register carries a well-supported claim and an invented one. This is why retrieval matters so much in practice: giving the model source documents to work from converts a recall problem into a reading problem, which it handles far better.
What are the limits worth knowing?
A model's knowledge ends at its training cutoff, so anything later is unknown unless supplied at the time of the question. Training data carries the biases of the text it was drawn from. Inference is computationally expensive, which shapes what is economical to build. And the reasoning behind an answer is not inspectable: an explanation, when asked for, is generated text, not a trace of what happened inside.
How large language models are used in marketing
Production. Drafting, editing, adapting one piece into several formats, and translation. This is where adoption started and where it remains heaviest.
Analysis. Summarising research, clustering customer feedback, classifying inbound enquiries, and pulling structure out of unstructured text at a volume no team would attempt by hand.
Distribution. Increasingly the model is not a tool but the channel. Buyers ask an assistant about a category and read a composed answer, which makes how a company appears to a language model a distribution question rather than a production one.
How this impacts your business
A language model does not look a company up in a register. It composes a description from everything it read during training and whatever it retrieves at the moment of the question. Trade press, directories, forums, competitor comparison pages, and somewhere in that mix the company's own website too.
So the description a buyer sees is assembled, not published. No company controls it, but every company supplies a share of the material it is built from.
The traffic impact is still small, roughly one percent of sessions for most companies. The influence is not small, because the reader is often a buyer forming a first impression of what a company does and who it serves, before any page gets a click.
Checking takes ten seconds. Type the company name into ChatGPT and read the answer.
Frequently asked questions
What makes a language model large?
Parameter count and training data volume. Contemporary models have billions to hundreds of billions of parameters, trained on corpora measured in trillions of tokens. There is no formal threshold.
Is a large language model the same as artificial intelligence?
No. It is one approach within artificial intelligence, specialised in language. It is a narrow system in the technical sense, however broad its apparent competence.
Do these models understand what they are saying?
Contested, and it depends on the definition. They model relationships between symbols with great sophistication and have no perception or intention. Whether that constitutes understanding is a philosophical dispute rather than a technical one.
Can a company train its own?
Training from scratch costs millions and is rarely justified. The practical options are fine-tuning an existing model on domain data, or supplying context at query time through retrieval, which is cheaper and easier to keep current.
How do these models handle smaller languages such as Finnish?
Less well than English, because there is far less training text. Quality has improved substantially but the gap persists, particularly on idiom and specialist vocabulary. See does ChatGPT work in Finnish.
Related definitions
- What is generative AI?
- What is a neural network?
- What is deep learning?
- What is ChatGPT?
- AI Search 101
Written by Lari Numminen, Generate More. Also available in Finnish. Updated 31 August 2026.