Skip to the lesson
0%

How Language Models Find and Use Sources

Beginner 12 min read Lesson 2 of 12

An AI answer comes from two places: what the model learned during training, and what it retrieves from the web at the moment you ask. GEO has a different lever for each.

Training knowledge

A model is trained on a large snapshot of text up to a cut-off date. What it "knows" about your brand from training depends on how often and how consistently you were described in that text. This changes slowly: only when a new model version is trained.

Retrieval at answer time

Most assistants now search the web when a question needs current or specific information. The system turns the question into one or more searches, fetches the top results, reads the relevant passages and writes an answer from them, often with citations. This is called retrieval-augmented generation (RAG) or grounding.

Retrieval depends on ordinary search: if your page is not indexed and ranking in the search system the assistant uses, it cannot be retrieved.

Query fan-out

For a broad question an assistant may run several narrower searches behind the scenes ("best ERP for manufacturers", "ERP pricing India", "ERP implementation time"). Pages that clearly answer those sub-questions get pulled in. Google describes this behaviour for AI Mode as query fan-out.

What the model does with a source

It extracts passages that answer the sub-questions, compares them, and writes a synthesis. Clear, specific, well-sourced passages are easier to use and to cite. Vague marketing copy contributes little.

How a grounded answer is built
1QuestionUser asks theassistant2Fan-outSeveral searchesare run3RetrieveTop pages arefetched4ExtractRelevantpassages pulled5Write andciteAnswer withsource linksRetrieval-augmented generation in five steps

Pro tips

  • For current facts (prices, offers, staff), focus on retrieval: make sure the right page ranks.
  • For brand reputation, focus on consistency across the web, which feeds future training.
  • Write pages that answer the likely sub-questions of a broad query.

Common mistakes

  • Assuming the model reads your site directly whenever someone asks about you.
  • Updating a fact on one page and leaving the old fact on five others.
  • Writing only broad pages that answer no specific sub-question.
Summary

Answers combine training knowledge and live retrieval. Retrieval runs on search, so ranking and clear passages matter; training reflects how consistently you are described.

Quick quiz

1. What is RAG?
2. What is query fan-out?
3. If your page is not indexed, can retrieval use it?

Stuck on something in this lesson? Ask our team on WhatsApp. We answer course questions for free.

Ask on WhatsApp