What is an LLM?
According to Google, a language model estimates how likely a token or sequence of tokens is to occur within a longer text. A token is the smallest unit the model works with: a whole word, part of a word or a single character.
An LLM uses this ability to continue text. It predicts the next token, appends it and then predicts the one after that. Many such steps produce a sentence, a paragraph or an entire text. A simple example: “The capital of France is” is followed with very high probability by “Paris”.
“Large” refers to size: LLMs are trained on very large amounts of text. From this they learn language, relationships and a great deal of factual knowledge, but not verified knowledge in the sense of a database.
How an LLM works
Many LLMs are built on the Transformer, an architecture introduced in 2017 in the paper “Attention Is All You Need”. According to Google, the Transformer is now the leading architecture for many language applications, such as translation.
Its core is called self-attention. For each token, the model works out how strongly the other tokens influence its meaning. This is how it recognises that “bank” means something different in “we had a picnic on the river bank” than in “I opened an account at the bank”.
In practice, the context window matters most. It covers all the text the model can take into account when answering: your question, the conversation so far, any documents you have added and the answer itself. Anthropic describes it as the model’s working memory, separate from its training data. More context is not automatically better: the more text there is in the window, the more likely the model is to overlook details.
Strengths and limits
LLMs summarise texts, translate, write drafts, answer questions and bring unstructured information into a fixed format. Even so, you should know their limits:
- Hallucinations: an LLM generates probable text, not necessarily true text. It can invent facts, figures or sources that sound convincing. See AI hallucinations for more.
- Knowledge cut-off: what the model learned in training ends with the state of its training data. It only knows newer information if it is given to it in the context window via a search or a document.
- Bias: like all machine learning models, LLMs can adopt biases from their training data, such as one-sided portrayals.
- Cost: according to Google, training and running large models requires a lot of computing power and electricity.
LLMs and search: why this affects your website
More and more people ask their questions directly to AI assistants or read the AI answers above the search results, such as Google’s AI Overviews. To give current information, the service retrieves suitable pages and passes them to the model together with the question. The model then formulates an answer and, ideally, names its sources.
For your website, this means it must be accessible to these services’ search crawlers, and its content must be written clearly enough for a model to reproduce it correctly. OpenAI, for example, states that websites that block its search crawler OAI-SearchBot will not be shown in ChatGPT’s search answers. You control which crawlers you allow in your robots.txt. How to prepare content for AI answers is explained in the article on generative engine optimisation.
Using LLMs sensibly in marketing
- For drafts, not final versions: outlines, variants of ad copy or summaries of long documents are good uses. The expert review remains your job.
- Provide sources: paste the documents the answer should be based on directly into the request and ask for verbatim quotes from them. Anthropic explicitly recommends this to reduce hallucinations.
- Allow uncertainty: tell the model it may answer “I don’t know”. Otherwise it tends to fill gaps with guesses.
- Ask precisely: the more clearly you describe the task, target group and desired format, the more useful the result. See prompt engineering for more.
- Protect confidential information: agree in advance which customer data and internal documents may be entered into which service.
Would you like to know whether AI assistants know your company and describe it correctly? Then get in touch with us or read more about our GEO services!
