Stage 1 — Retrieval: can the engine find you at all?

Before any intelligence is applied, an engine must retrieve candidate material. Retrieval happens two ways. Index-based retrieval queries an existing search index — Google AI Overviews and Gemini ground against Google's index; ChatGPT and Copilot lean on Bing's index supplemented by their own crawlers. Live retrieval fetches pages in real time — Perplexity is built this way, and ChatGPT's OAI-SearchBot fetches pages for search queries on demand.

The implications are blunt: if your robots.txt blocks GPTBot or PerplexityBot, if your content renders only via JavaScript, or if Bing has never indexed your English pages, you are absent from the candidate pool before quality is ever assessed. Retrieval is a binary gate — the most common reason brands are invisible to AI is not weak content but technical absence.

Stage 2 — Grounding: which passages does the model read?

Retrieved pages are not read whole. They are chunked into passages, embedded, and ranked by relevance to each sub-question. Google describes its approach as query fan-out: one user question becomes many hidden sub-queries, each pulling its own candidate passages. The passages that survive tend to share traits: they state a fact or definition directly, they sit under a clear heading, they carry numbers, dates and named entities, and they agree with what other credible sources say.

This is why answer-block writing — a 40–80 word direct answer immediately under a question-shaped heading — is the single highest-leverage GEO formatting technique: it produces exactly the passage shape grounding systems prefer.

Stage 3 — Synthesis: does the model trust you enough to name you?

During synthesis the model drafts the answer. Two forces decide whether your brand is named. Corroboration: claims repeated consistently across independent sources are stated confidently; claims found in only one place are hedged or dropped. Entity resolution: the model must be confident that "Haipengyun", "haiyingyun.net" and "Beijing Haipengyun Holding Group" are one thing — inconsistent naming across sources fragments your entity and dilutes every mention.

Stage 4 — Citation: which sources get shown?

Citation selection is partly a trust ranking and partly a product decision. Google shows link cards beside the AI Overview; Perplexity numbers its sources inline; ChatGPT shows a sources panel; Claude cites sparingly and prefers primary documents. Across engines, citation favors: the page the passage was grounded in, sources with recognizable authority in the topic, and pages that are safe to send a user to (fast, clean, not paywalled).

The GEO takeaway in one line: retrieval is won with technical hygiene, grounding with passage structure, synthesis with corroboration and entity consistency, citation with source authority. Four stages, four workstreams — that is the entire discipline.

Where engines differ

The pipeline is universal; the parameters are not. Google grounds in the world's largest index and its own Knowledge Graph. ChatGPT blends Bing's index, its own crawl and licensed corpora, with a memory of training data that can substitute for retrieval. Perplexity retrieves live on every query and exposes its source list. Grok adds the real-time X firehose. The Engines section documents each pipeline in detail, and How LLMs choose sources covers the shared citation logic in depth.

Frequently asked questions

What is query fan-out?
Query fan-out is Google's technique of expanding one user question into multiple hidden sub-queries, each retrieving its own candidate passages that are then synthesized into one AI Overview. It means a single buyer question like 'best CRM for a 20-person logistics company' fans out into sub-queries about features, pricing, company size fit and industry fit — your content needs to win passages across several of them, not just the head term.
Do AI engines read my whole page?
No. Pages are split into chunks or passages, each independently scored for relevance to sub-queries. A buried insight in paragraph 40 competes alone. Put the answer first, keep sections short, and make every H2 a question or claim a passage can stand behind.
If I block AI crawlers, am I safe from AI using my content?
Mostly, for training — but blocking crawlers also removes you from retrieval, which means competitors get cited for your category instead. For most brands the strategic question is not whether AI reads you, but whether AI reads you correctly and cites you when it answers.