A large language model (LLM) is an AI model trained on very large amounts of text that generates language by repeatedly predicting the most likely next piece of text. The model families behind ChatGPT, Gemini and Claude are well-known examples. In legal education, LLM (or LL.M.) also stands for Master of Laws; this entry covers the AI meaning.
At a glance
- Type
- AI model for language (language model with a very high number of parameters)
- Architecture
- usually Transformer (Vaswani et al. 2017)
- Used in
- ChatGPT, Gemini, Claude and AI search
Last checked:
How does an LLM work?
Today’s LLMs are built on the transformer architecture, introduced by Google researchers in the paper “Attention Is All You Need,” first posted on June 12, 2017. A transformer uses self-attention to weigh, for every part of the input, which other parts of the context matter for it.
Text is split into tokens before processing. A token can be a whole word, part of a word or a single character. In pretraining, the model learns to predict tokens on large text collections. It is then usually fine-tuned with examples and human feedback so it follows instructions and gives helpful answers.
An LLM does not store a copy of the web pages it was trained on. It produces answers from statistical patterns learned in training. That explains hallucinations, statements that sound plausible but are wrong. Its knowledge also stops at the date of its training data, unless the application around it fetches current information.
Why do LLMs matter for SEO?
Website content can reach an AI answer in two ways:
| Path | What happens | Example at OpenAI |
|---|---|---|
| Training | Crawlers collect content that may be used for future models | GPTBot |
| Retrieval at answer time | The system searches the web for a question and uses the pages it finds as sources | OAI-SearchBot (ChatGPT search), ChatGPT-User (user-triggered actions) |
OpenAI documents that each of these bots can be controlled separately in robots.txt. A site can block GPTBot and still appear in ChatGPT search through OAI-SearchBot. For ChatGPT-User, OpenAI notes that robots.txt rules may not apply because a user triggers the request. Anthropic uses a similar split with ClaudeBot (training), Claude-User and Claude-SearchBot.
Retrieval at answer time works much like a search engine visit. The page has to be reachable, readable without heavy client-side hurdles, and not blocked for the bot in question.
- OpenAI releases ChatGPT as a free research preview. The model is based on the GPT-3.5 series.Source: openai.com
- Google announces its conversational AI service Bard and opens it to trusted testers first.Source: blog.google
- Microsoft launches the new Bing with AI chat and complete answers inside search.Source: blogs.microsoft.com
- Google introduces the Search Generative Experience (SGE) as an experiment in Search Labs. At launch it is limited to the US and English.Source: blog.google
- Microsoft renames Bing Chat to Copilot. General availability follows from December 1, 2023.Source: blogs.microsoft.com
- At Google I/O, Google announces AI Overviews for everyone in the US, with more countries to follow.Source: blog.google
- Bing previews its generative search experience. At first it appears for a small share of queries only.Source: blogs.bing.com
- OpenAI tests SearchGPT as a temporary prototype with a small group of users and publishers.Source: openai.com
- AI Overviews roll out in more than 100 countries and territories. Google expects more than 1 billion monthly users.Source: blog.google
- OpenAI launches ChatGPT search, first for paying users and the SearchGPT waitlist. Answers link to their sources.Source: openai.com
- Google launches Deep Research in Gemini Advanced. It browses the web on its own and compiles a report.Source: blog.google
- ChatGPT search becomes available to all logged-in users in regions where ChatGPT is offered.Source: openai.com
- OpenAI introduces deep research in ChatGPT, starting with Pro users.Source: openai.com
- ChatGPT search becomes available without signing up in all regions where ChatGPT is offered.Source: openai.com
- Google introduces AI Mode as an experiment in Search Labs. Google One AI Premium subscribers get access first.Source: blog.google
- Anthropic adds web search with citations to Claude. It launches as a preview for paid users in the US.Source: anthropic.com
- AI Overviews launch in Germany in German and English, together with other European countries. They are shown to signed-in users aged 18 and over.Source: blog.google
- At Google I/O, Google rolls out AI Mode to everyone in the US, with no Labs sign-up needed.Source: blog.google
- AI Overviews are available in more than 200 countries and territories and in more than 40 languages.Source: blog.google
- Web search in Claude becomes available globally on all plans.Source: anthropic.com
- Data from AI Mode counts towards the totals in the Search Console Performance report.Source: developers.google.com
- AI Mode launches in Germany, Austria and Switzerland as part of an expansion to more than 40 new countries and territories.Source: blog.google
- Search Console adds dedicated performance reports for generative AI features such as AI Overviews and AI Mode, first for a subset of sites.Source: developers.google.com
- The Search Console reports on generative AI features become available to all websites worldwide.Source: developers.google.com
Dates follow the publisher's announcement; rollouts often took days or weeks. Google's English post on AI Overviews in Europe carries the US date of March 25, 2025, the German post March 26, 2025. Last checked: .
Example
A user asks an AI assistant about the return policy of an online shop. Without web access, the model guesses from older training data. With search enabled, the assistant fetches the shop’s current returns page and cites it. If that page blocks the provider’s search bot, the answer has to rely on third-party sites instead.
FAQ
Is ChatGPT an LLM? ChatGPT is an application by OpenAI that runs on LLMs from the GPT series. The model generates the text, while the application adds web search, file uploads and conversation history.
What is the difference between LLM and GPT? LLM is the general category. GPT (generative pre-trained transformer) is the name of OpenAI’s model series within that category.
What does LLM stand for? Large language model. “Large” refers to the amount of training data and the number of model parameters.
Sources
- Vaswani et al.: Attention Is All You Need, arXiv:1706.03762 (submitted 2017-06-12)arxiv.org
- Machine Learning Glossary: large language model – Google for Developersdevelopers.google.com
- Ouyang et al.: Training language models to follow instructions with human feedback, arXiv:2203.02155 (2022)arxiv.org
- Overview of OpenAI Crawlers – OpenAIdevelopers.openai.com
- Does Anthropic crawl data from the web? – Claude Help Centersupport.claude.com
- Large language model – Wikipedia (secondary, for training stages and terminology)en.wikipedia.org
Last updated: