How an AI Picks an Answer

Reading intent: identifying the scenario, constraints and real decision goal. This depends on how the user asks — not controllable.

Decomposition: Google's guidance names query fan-out alongside RAG as how AI Overviews and AI Mode work — one question becomes several sub-queries retrieved separately. A page need not answer the whole question; being the better answer to one sub-query is enough.

Retrieval: candidates come back by semantic distance, not keyword presence. A page focused on one topic has a clear vector position and is retrieved accurately.

Reranking: beyond relevance, authority, freshness and cross-source agreement act here. This is the gap between being retrieved and being used.

Generation: the best-supported content gets written in. The wording is the model's — not controllable.

For retrieval: make every paragraph readable on its own and keep each page to one subject. Passage retrieval takes paragraphs, and whoever reads one has no surrounding context.

For reranking: supply evidence. The Princeton paper measured a 30–40% relative lift from adding statistics, cited sources and quotations precisely because those act on whether the content is worth believing.

Citations are long-tail: even the most-cited domains rarely exceed 5% share on a single platform. The goal is reliable presence in the source list, not owning it.

The same question returns different answers across time, region and device, so a single observation proves nothing. Sample repeatedly and read the distribution.