# How AI search works

A practical mental model for retrieval, synthesis, citations, and why answers vary.

You do not need to reverse-engineer every model. You do need a working model of the path between a buyer's question and the answer they see.

## The simplified answer path

1. The system interprets the question and its likely intent.
2. It may rewrite the question into several searches or sub-questions.
3. It retrieves pages or passages from an index, the live web, or connected sources.
4. It assesses relevance and usefulness among the retrieved material.
5. A model synthesizes an answer from its prior knowledge and retrieved context.
6. The interface may attach citations, products, maps, or follow-up actions.

## Why this matters for content

A useful source must clear several gates. It needs to be accessible, understandable without excessive context, relevant to the sub-question, supported by evidence, and consistent with what credible sources say elsewhere.

## Why results vary

Answers can change with prompt wording, conversation history, geography, time, account context, retrieval availability, and the model version. Measure with a stable prompt set and repeated observations. A single screenshot is an anecdote, not a benchmark.

  
### Do schema and llms.txt guarantee inclusion?

No. Structured data can reduce ambiguity and explicit machine-readable files can help certain crawlers, but neither substitutes for accessible, useful, corroborated information.