Every grounded answer an assistant gives - the paragraph with citations that increasingly ends a buyer's research - is the output of a pipeline you can reason about even though its weights are private. The engines publish enough about their retrieval systems, and behave consistently enough under systematic measurement, that the pipeline's stages and preferences are observable. This article walks the four stages and what each one rewards - because every effective GEO tactic is an answer to one of them.
Opening the black box, honestly
A caveat worth stating plainly: nobody outside the engine companies knows the exact ranking functions, and anyone selling certainty is selling. What follows is built from three legitimate sources of knowledge: what the companies publish about their systems (retrieval-augmented generation is a documented architecture, not a rumor); what the academic literature measured (the GEO paper quantified which content properties move visibility); and what repeated measurement across models shows in practice. The pipeline below is the consensus shape.
Stage one: query formulation
The user's conversational prompt - often long, contextual, mid-thread - is not what gets searched. The model reformulates it into one or more retrieval queries: a question about "tools for tracking whether ChatGPT recommends us" might fan out into queries about brand monitoring, AI answer tracking and share-of-voice measurement. Two consequences follow. Coverage beats repetition: content that addresses a question's phrasings and neighbors intersects more query variants than content optimized for one exact string. And conversational context leaks in: the same question asked inside a startup-budget conversation retrieves differently than inside an enterprise one - which is why measurement uses fixed question sets rather than anecdotes.
Stage two: retrieval
Under every assistant sits a search layer - proprietary indexes, licensed indexes, or conventional engines - returning candidate documents for each query. The unglamorous implication is the most important one in this article: retrieval is where technical citability and classic search equity decide your fate. Pages that cannot be fetched, render empty to a plain HTTP client, or rank nowhere for any relevant query simply never enter the candidate pool. GEO does not bypass search; it inherits it - the argument the AEO vs SEO comparison makes at length.
Stage three: selection
With candidates in context, the model chooses which passages inform the answer - the stage where GEO's distinctive craft lives. Observed behavior across models is consistent about what wins:
| Property | Why it wins |
|---|---|
| Direct answers | A passage that answers in its first sentence beats one that builds to it |
| Self-containment | Chunks that survive extraction without losing meaning are usable; context-dependent prose is skipped |
| Specificity | Numbers, names and definitions give the model something to hold; adjectives give it nothing |
| Internal coherence | Passages that contradict themselves or their own page get discarded |
| Agreement | Claims corroborated by other candidates are safer to assert than outliers |
These properties are exactly the liftable-chunk formats - not because engines published a spec, but because selection is a model reading under time pressure, and models reading under pressure prefer what research on how users read online says any pressed reader prefers.
Stage four: synthesis and attribution
The answer gets written, and compression is severe: dozens of candidates become a few paragraphs with a handful of citations. Two attribution behaviors matter for strategy. Citation favors the source whose passage the answer actually leans on - being paraphrased without citation usually means your content informed the answer but another source stated it more quotably, which is a fixable content problem. Named mention ("according to X", or a brand recommended by name) favors entities the model holds confidently - built from entity consistency and repeated corroborated appearances, which is slower work than a content retrofit and worth starting earlier.
What happens when sources disagree
Disagreement handling also explains a failure mode teams misdiagnose: a brand whose own site says one thing while directories and old press say another is not "unlucky" in AI answers - it is presenting the engines with a contradiction, and contradictions lose to whichever competitor reads consistently.
Implications for getting cited
Mapping tactics to stages turns GEO from folklore into engineering. Absent from answers entirely? That is retrieval - run the AI search optimization checklist before touching content. Retrieved but never quoted (your topic ranks, your brand does not appear)? That is selection - retrofit the pages with direct answers and liftable chunks. Quoted but never named, or losing ties to a rival? That is synthesis - corroboration and entity work. The diagnostic sequence is exactly what the AI visibility audit playbook runs in an afternoon, and it is the difference between doing GEO and doing GEO-flavored activity.
Frequently asked questions
How do AI assistants pick which sources to cite?
Through a four-stage pipeline: the prompt becomes search queries, retrieval returns candidates, the model selects passages that answer directly and agree with other candidates, and synthesis cites the sources the answer actually leaned on.
Why does my content get paraphrased but not cited?
It usually informed the answer while another source stated the same point more quotably. The fix is content shape: direct first-sentence answers, self-contained passages, and specific claims worth attributing.
Do AI engines prefer fresh content?
On time-sensitive questions, observably yes - freshness is a visible tie-breaker between disagreeing sources. Honest published and modified dates, updated when content really changes, are the workable lever.
What makes an assistant name a brand directly?
Entity confidence: consistent descriptions of the company across its site, directories and third-party mentions, reinforced by repeated corroborated appearances. It builds slower than citation and is worth starting early.
Can you influence which sources are retrieved?
Only through the classic levers: technical accessibility and search equity for the queries your buyers’ prompts decompose into. Retrieval is conventional search under the hood, which is why GEO builds on SEO rather than replacing it.
Sources
- GEO: Generative Engine Optimization (Aggarwal et al., 2023) - measured visibility effects
- Lewis et al., 2020 - Retrieval-Augmented Generation, the architecture underneath grounded answers
Every playbook on this blog ships as a runnable mission.
Open a workspace and the playbook library is waiting - describe the outcome and the agents carry it end to end, on your plan's monthly credits.