When an AI assistant cites or recommends a brand, the visible answer is the end of a longer process.
The tempting story is that there is a secret list of "LLM ranking factors." The more useful model is that a source must pass through several stages: it needs to be available, relevant information needs to be retrieved, the information needs to be useful to the answer, and the generated response must decide how to attribute or present it.
That distinction explains why GEO experiments often produce inconsistent results. Improving a paragraph can help after retrieval while doing nothing to make the page discoverable in the first place.
A four-stage model of an AI citation
Use this model when diagnosing why a brand is missing.
Stage 1: eligibility and discovery
Can the system or its search layer access the page at all?
OpenAI tells publishers who want to appear in ChatGPT search to allow OAI-SearchBot. Perplexity recommends allowing PerplexityBot. Google says supporting links in AI Overviews and AI Mode need normal Search indexing and snippet eligibility. Bing says search and Copilot grounding share core crawl and index foundations.
If the source is inaccessible, wrongly canonicalized, noindexed, buried without internal links or otherwise absent from the relevant index, content tweaks further down the pipeline cannot compensate.
Stage 2: retrieval
When a user asks a question, the system needs candidate information relevant to the intent.
Retrieval is not the same as "ranking in the final answer." It is the process of finding material that can help answer the question.
A page can improve its chances of being useful at this stage by being specific about the topic, entities, facts, comparisons and use cases it covers. Clear language helps because the retrieval system has less ambiguity to resolve.
Stage 3: evidence selection and attribution
A retrieved source may be used, ignored, summarized or cited.
This is the stage most early GEO experiments examine. The foundational GEO paper tested content changes inside controlled generative-engine settings and reported visibility improvements from some interventions.
But the scope matters. A 2026 critical survey warns that results observed after content is already supplied to a system should not be treated as proof of durable organic discoverability across platforms.
In practical terms: making a source more quotable can help only if the source gets into the room.