A large language model, or LLM, is an AI system trained on a vast amount of text so that it can predict which words should come next. That one skill, done at enormous scale, lets it write, summarise, translate, classify and answer questions in ordinary language. ChatGPT, Claude, Gemini and Microsoft Copilot are all built on LLMs.
How a large language model works
During training, text is split into small units called tokens, which are often parts of words. The model, a very large neural network, reads billions of examples and adjusts itself until it is good at predicting the next token in a sequence. Later stages of training teach it to follow instructions and to answer helpfully rather than simply continue text.
When you use one, your request (the prompt) and any documents you add are converted into tokens. The model then generates its reply one token at a time, each choice based on everything before it. How much it can consider at once is limited by its context window.
Three consequences follow from this design:
- It is not a database. It does not store facts in a lookup table, so it can produce fluent text that is wrong, which is known as hallucination.
- Its knowledge stops at a training cut-off date unless it is connected to search or to your own documents.
- Its output depends heavily on the instructions and examples you give it, which is why prompt engineering makes such a difference.
Why it matters
For a UK business, LLMs matter in two ways. The first is everyday work: drafting emails and product descriptions, producing ad copy variations to test, summarising call notes or reports, sorting incoming enquiries and turning rough notes into a first draft. Used well, they save time on tasks where a person still checks the result.
The second is search. ChatGPT search, Perplexity, Google’s AI Overviews and AI Mode all use LLMs to write answers, often citing a handful of web pages as sources. When a potential customer asks one of these tools for an accountant in Leeds or the best way to fix a damp wall, the model decides which businesses and pages to mention. Being a clear, trustworthy source for those answers is becoming part of SEO.
There is a UK-specific catch. Most models default to American spelling and context. Left unchecked, they use American spellings, quote prices in dollars, cite US law and use US date formats. Anything produced for a UK audience needs checking for British English, GBP and the right legal references, such as UK GDPR and PECR rather than US rules.
Common mistakes
- Treating the output as fact without checking names, numbers and sources.
- Pasting customer personal data into a consumer tool without checking how the provider stores and uses it.
- Publishing unedited AI text at volume. Google’s spam policies target content mass-produced mainly to manipulate rankings, however it was made.
- Expecting the model to know your prices, services and policies without telling it.
- Missing the American defaults in spelling, currency and law.
How to act on it
- Start with tasks where an error is cheap to spot: summaries, first drafts, headline options.
- Give the model context: who the audience is, that they are in the UK, what tone to use, and an example of good output.
- Supply the facts yourself, and check every claim before anything is published.
- Agree simple rules on what data may go into which tools.
- Make the facts about your business easy for AI search to find and quote accurately.
The guide on using AI for marketing without hurting SEO covers the content side. For visibility in AI answers, see my generative engine optimisation service.