An AI answer is produced in four stages: the query is understood (sometimes rewritten or fanned out into related sub-queries), candidate information is retrieved — from the model’s training knowledge, from live web search, or both — a subset of those candidates is selected to ground the answer, and the answer is generated as prose with citations attached to the selected sources. Which stages run, and which sources get selected, differs by surface and by how the question is phrased; the selection stage is where your citation is won or lost, and it is the stage least documented publicly. Everything below is what the surfaces themselves say plus what published studies have observed — labeled accordingly.

Stage 1: query understanding

The question you type is rarely processed as-is. Two documented behaviors:

  • Deciding whether to search at all. ChatGPT “will choose to search the web based on what you ask, or you can manually choose to search” (OpenAI’s 2024 launch post for ChatGPT search, https://openai.com/index/introducing-chatgpt-search/). This is not trivia: in Profound’s study of roughly 700,000 ChatGPT conversations, only about 18% triggered at least one web search (Profound, published February 2026, summarized in what ChatGPT cites). The other ~82% were answered from the model’s internal knowledge — where nothing you publish in 2026 can appear.
  • Fan-out into sub-queries. Google documents that its AI features “may use a ‘query fan-out’ technique — issuing multiple related searches across subtopics and data sources — to develop a response” (Google Search Central, “AI features and your website,” https://developers.google.com/search/docs/appearance/ai-features). One user question becomes several retrievals, which is why a page can be irrelevant to your exact query and still be pulled in — or pulled in and still not cited.

Stage 2: retrieval — model knowledge vs live web

The engine’s candidate pool comes from one or both of:

  • The model’s training knowledge. A snapshot in time, no citations attached, no way to know what’s current. For evergreen category facts this pool dominates; for anything current it is unreliable and engines generally know it.
  • Live web search. The retrieval step where citable pages enter the pool. The surfaces describe it slightly differently:
    • ChatGPT: “fast, timely answers with links to relevant web sources,” with a Sources sidebar under the response (OpenAI, 2024 launch post; the ChatGPT search FAQ describes the same for current users). OpenAI’s developer documentation for its web search tool states plainly: “Web search allows models to access up-to-date information from the internet and provide answers with sourced citations” (https://developers.openai.com/api/docs/guides/tools-web-search).
    • Perplexity: real-time by default — it “uses advanced AI to search the internet in real-time, gathering insights from top-tier sources,” and “each answer includes numbered citations linking to the original sources” (Perplexity Help Center, “How does Perplexity work?”, https://www.perplexity.ai/help-center/en/articles/10352895-how-does-perplexity-work). Its Pro/Research modes run multiple deeper retrieval passes over many sources.
    • Google: AI Overviews are “help[ing] people get to the gist of a complicated topic or question more quickly, and provid[ing] a jumping off point to explore links to learn more,” and “are only shown when our systems determine that it is additive to classic Search, and as such, often don’t trigger” (Google Search Central, https://developers.google.com/search/docs/appearance/ai-features). To be a supporting link, a page “must be indexed and eligible to be shown in Google Search with a snippet” — no special markup, no AI-specific file, by Google’s own statement.

Stage 3: selection — where citations are decided

Retrieved candidates are not all cited. This is the least transparent stage, and the published evidence on it is observational:

  • Only a subset of retrieved candidates ends up cited, and retrieval frequency predicts citation. The closest published evidence is an observational 2026 study of production ChatGPT web search across four model configurations (GPT-5.2 through GPT-5.5) and both Instant and Thinking modes: “candidates retrieved more often are also more likely to be cited” (“Understanding ChatGPT Citations,” arXiv:2610.15382, https://arxiv.org/abs/2610.15382). Observational, not causal — but it means the selection stage is reachable, in part, through the retrieval stage.
  • External proxies match citations better at hostname and entity level than exact URL. The same study found that Google Search results plus OpenAI Responses API candidates, taken together, “match ChatGPT Web better at the hostname and entity levels than at the exact-URL level” — and a supervised model on those proxies reached AUC above 0.74 for predicting citation at all three levels. The practical consequence: being the entity a claim binds to matters as much as owning one perfect URL, and presence across your domain (not one page of it) is the more robust target.
  • The proxies recover only about a quarter of cited URLs. Google Search plus the Responses API together “recover only about one quarter of cited URLs” (arXiv:2610.15382). The engine’s live retrieval pool is largely private. This is the single most important number for anyone who has tried to game the system by watching what one proxy returns: you are seeing roughly 25% of the candidate pool through the wrong lens.

Stage 4: generation with citations

The answer is prose generated from the selected grounding, with citations attached as it is written. The documented behavior that matters:

  • Citations are attached to claims, and claims bind to entities. A named product or company in the answer is the model’s resolved entity, not your URL. If your entity is ambiguous — inconsistent naming, no structured data, facts that conflict across your own pages — the claim attaches to the nearest resolvable entity or is dropped. (Full treatment in the content & corroboration layer of the series.)
  • Multi-round retrieval changes the selection set. In the arXiv:2610.15382 study, Instant settings “remain mostly single-round, while GPT-5.4 Thinking frequently performs multiple retrieval rounds” — the deeper mode retrieves more and re-weights, so the cited set differs by mode for the same question. One probe of one mode is one data point, not the system.
  • Answers triangulate across sources. Cited conversations average ~6 unique citations (Profound, February 2026), and models treat vendor content as one source type among several — review, community, comparison, primary. What third-party presence actually does for your share of that set is the subject of why AI answers triangulate.

What this means for AEO/GEO work

Three implications, in order of how often they get violated:

  1. Do not reverse-engineer one engine. The retrieval pool is mostly private (the ~25% proxy result), the selection logic is unpublished, and it changes with every model generation. Work that optimizes to a snapshot of one engine’s behavior — one proxy, one mode, one month — is work that decays. The durable targets are the inputs the studies say move selection: retrieval frequency (reachability, entity presence, corroboration), extractability of the passage, and hostname-level consistency.
  2. Measure across surfaces and modes. ChatGPT, Perplexity, Gemini and Claude retrieve and cite differently; even within ChatGPT, Instant and Thinking modes differ in how many retrieval rounds they run, so the set one mode cites is not guaranteed to be the set the other will. The measurement setup — GA4 referrals, GSC’s generative-AI reports, Bing’s AI Performance, and the cross-surface question probes — is in how to measure AI search traffic.
  3. Distinguish retrieval visibility from citation visibility. Being fetched is not being cited; being cited is not being mentioned favorably; being mentioned is not being clicked. The research term for the first two is the subject of what is GEO, and the distinction is where most “we’re invisible to AI” diagnoses go wrong — the failure is usually in one specific stage, and the fix differs by stage.

The honest limits of this picture

What we do not know, and what no vendor can honestly show you: the exact ranking function inside the selection stage, the size of the live retrieval pool, or how training-time knowledge and retrieved content are weighted in the final prose. Published evidence is observational and changes with every model generation. The defensible position is the one this post takes: the four-stage pipeline is well supported, the stage-level evidence is real but single-source or observational, and the work that survives engine changes is the work on your side of the retrieval boundary — content, entities, corroboration, reachability — measured against a standing diff instead of a snapshot.

The full series — definitions, infrastructure, content and corroboration — is in AEO and GEO in one page.