An LLM does not use one universal information pipeline. What an answer draws on depends on the product, its tools, the available context, and the question.
Four common sources
- Existing model knowledge: learned patterns and information from training. Publishing a new website page does not immediately update those model parameters.
- Web search and page retrieval: a search-enabled assistant can find public sources and fetch relevant pages.
- Uploaded or connected documents: a product may use files, organization data, or other authorized sources.
- Conversation context: information supplied by the user or already included in the current interaction.
One possible retrieval design
A retrieval-augmented system can index documents, search for relevant passages, and supply selected material to the model. It may use keywords, embeddings, reranking, or several searches. Passage size and selection vary; some products also process entire documents or larger contexts. A fixed 100–300-word chunk is not a universal rule.
What website owners can improve
Publish clear, accurate HTML with descriptive titles and headings. Keep URLs stable and important pages accessible. Use reliable sources and update stale claims. A clean text version can help tools that use it, while the normal web page remains important.
Neither crawl access nor a text mirror guarantees that a particular answer uses your page. Measure retrieval requests and sampled citations separately with the visibility metrics guide.