In short. A large language model predicts token sequences, while retrieval can supply external passages for an answer.
Definition: A large language model (LLM) is a language prediction system trained on large text collections. It estimates probabilities for sequences of tokens, units of text such as words, word fragments or characters, and can generate text from those predictions.
Key facts
- Google's language model introduction defines language modelling through probabilities assigned to token sequences.
- Google's LLM lesson describes the Transformer as a widely used architecture for LLMs.
- The retrieval-augmented generation paper combines a language generator with an external document index.
- As of 17 October 2026, Google's AI features documentation says AI Overviews and AI Mode may issue multiple related searches to develop a response.
How does an LLM process text?
Tokenization divides text into the units a model processes. Google's tokenization example shows how the word unwatched can split into a prefix, root and suffix:
un (the prefix)
watch (the root)
ed (the suffix)
Context is the surrounding information used to interpret a token. In the same lesson, Google explains that context helps distinguish whether orange refers to a fruit or a colour.
An LLM uses a neural network, a model that learns patterns by adjusting numerical values called parameters. Google's Transformer lesson describes self-attention, a mechanism that weighs relationships between tokens in the input. It also describes decoder-only models, which generate new tokens from text already generated and can continue a supplied prompt.
Training adjusts parameters; generation uses the resulting model to produce text. Google describes instruction tuning as an additional training step that can improve instruction following.
How does retrieval relate to an LLM?
Retrieval selects information from an external collection. Retrieval-augmented generation (RAG) combines that information with a language model during answer generation.
The paper by Patrick Lewis and colleagues describes a language generator paired with an index of Wikipedia. Its generator uses retrieved passages when producing output. The paper also explains that pre-trained language models store factual knowledge in their parameters. These are separate sources of information for an answer: knowledge learned during training and passages supplied through retrieval.
For website analysis, this distinction separates a model's learned knowledge from the documents an answer system retrieves.
What does this mean for a website?
Google calls its related-search technique query fan-out: searches across subtopics and data sources help develop a response. Its website guidance specifies that a page must be indexed and eligible for a Google Search snippet to qualify as a supporting link in AI Overviews or AI Mode.
The same guidance recommends allowing crawling through robots.txt and hosting infrastructure, making pages reachable through internal links, and keeping important content available as text. These are concrete website checks relevant to Google's retrieval process.
Related terms
FAQ
Does generation always choose the most probable token?
Google's language model lesson describes both choosing the highest-probability prediction and randomly selecting among tokens whose probability exceeds a threshold.
Can an LLM produce incorrect facts?
Yes. Google's LLM lesson describes hallucination as predictions that contain mistakes.
Does a website need special LLM markup for Google AI features?
Google's AI features guidance says these features require no new AI text files or special schema.org structured data.
Sources
- Google for Developers, "Introduction to Large Language Models", 9 January 2026
- Google for Developers, "LLMs: What's a large language model?", 2 January 2026
- Patrick Lewis et al., arXiv, "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks", 22 May 2020
- Google Search Central, "AI features and your website", 10 December 2025
Check how a page reads to answer engines. Open the GEO score analyzer