ChatGPT search works by classifying whether your prompt needs live information, expanding it into multiple sub-queries, retrieving candidate pages from Bing’s index, then selecting a small fraction of those pages to cite in the final answer. It is a retrieve-then-cite pipeline, and each stage filters hard: studies show ChatGPT cites only about 15 percent of the pages it retrieves. Understanding where pages get dropped is the difference between guessing at AI visibility and engineering it. This guide walks the pipeline stage by stage, separates search answers from training-data answers, and shows what each stage means for your content.
Does ChatGPT use Bing or Google for search?
ChatGPT uses Bing, not Google, for live web retrieval, and this single fact reorders most brands’ AI visibility priorities. When a prompt triggers search, ChatGPT queries Bing’s index for candidate pages; there is no Google integration anywhere in the pipeline, per the architecture breakdown at AI+Automation. Microsoft’s stake in OpenAI makes the dependency durable, and it extends to Copilot, which shares the same index.
The evidence is measurable: Seer Interactive’s analysis of thousands of SearchGPT citations found 87 percent matched Bing’s top results, per Conbersa’s citation research. The practical consequence: a page that ranks well on Google but is missing or weak in Bing’s index is nearly invisible to ChatGPT search. Sites routinely discover they have Bing crawl problems they never noticed because Google traffic looked fine, the failure mode we covered in why your website is not showing in AI search.
What happens when you type a prompt into ChatGPT search?
The pipeline runs four stages: trigger classification, query fan-out, retrieval, and citation selection. First, the model decides whether your prompt needs live data at all; timeless questions get answered from training data with no search and no citations. Second, prompts that trigger search get expanded: ChatGPT rewrites your question into multiple sub-queries covering different facets, and 89.6 percent of prompts trigger two or more additional searches before an answer returns, per erlin.ai’s 2026 optimization study.
Third, each sub-query retrieves candidate pages from Bing in parallel. Fourth, and most brutally, the model selects which retrieved pages actually inform and get cited in the answer. An AirOps study analyzing 548,534 pages across 15,000 prompts found ChatGPT cites only 15 percent of what it retrieves. Retrieval gets you considered; structure, authority, and freshness get you cited. The fan-out stage matters most for strategy: your page can win a citation by answering a sub-query the user never typed, the same mechanic behind Google AI Mode’s query fan-out.
Why does ChatGPT cite only 15 percent of what it retrieves?
ChatGPT cites so little of what it retrieves because the citation stage optimizes for answer construction, not ranking, and most retrieved pages do not hand the model a usable answer block. Research from SE Ranking, AirOps, Authoritas, and Fortis Media across 2025 and early 2026 converges on three citation drivers: content structure, domain authority, and freshness, per erlin.ai.
Structure means the answer is extractable: a direct answer high on the page, headed sections that map to sub-questions, tables and lists the model can lift without reconstruction. The position of the answer inside the page matters too: citation studies consistently find the first third of a page supplies the large majority of cited passages, so an answer buried after 800 words of preamble effectively does not exist for this pipeline. Authority means the model has reason to trust the domain when sources conflict; third-party corroboration and consistent entity signals do that work. Freshness means the page carries a recent date and current data; stale statistics lose to pages updated within months. A page can rank third on Bing for the sub-query and still lose the citation to the seventh result that states the answer in its first paragraph. The full playbook for winning that selection stage is in how to get cited by ChatGPT.
When does ChatGPT answer from training data instead of search?
ChatGPT answers from training data when the prompt looks timeless, and those answers carry no citations and cannot be won by fresh content. “What is answer engine optimization” may never trigger search; the model already holds an answer. “Best AEO agencies 2026” triggers search because the model detects recency demand. This split creates two distinct visibility layers: the retrieval layer, where the pipeline above applies, and the training layer, where your brand either exists in the model’s weights or does not.
The training layer moves slowly and rewards durable third-party coverage: publications, directories, and reference content that survive into future training corpora. The retrieval layer moves fast and rewards indexable, structured, current pages. A brand can dominate one and be absent from the other; most measurement confusion in AI visibility comes from conflating them. Watch which of your target queries return citations (retrieval) versus unlinked mentions (training), and split your strategy accordingly: pages and Bing indexing for retrieval, earned media and entity building for training, the two-track approach behind how to get your brand mentioned by ChatGPT.
Which crawlers does OpenAI run and what does each do?
OpenAI runs three crawlers with separate jobs, and your robots.txt treats each independently. OAI-SearchBot builds the search index that powers ChatGPT search results; blocking it removes you from retrieval. ChatGPT-User fetches pages in real time when a user’s conversation calls for a specific page visit. GPTBot collects training data for future models; blocking it affects the training layer, not search visibility.
The distinction matters because blanket AI-bot blocks are common and often accidental. Many sites blocked “AI crawlers” as a category in 2024 and 2025 without separating training collection from search indexing, and CDN-level bot protection quietly blocks these agents on roughly 27 percent of B2B sites. If OAI-SearchBot cannot reach your pages, nothing else in this pipeline matters: you are filtered at stage zero. Verify access before optimizing anything downstream; the crawler-by-crawler decision framework is covered in our companion piece on AI crawler policy, and the indexing mechanics are in how to rank in Microsoft Copilot, which shares the Bing dependency.
How should you optimize for each stage of the pipeline?
Optimize stage by stage: Bing indexing for retrieval, sub-query coverage for fan-out, and answer-first structure for citation selection. Concretely, that means verifying your site in Bing Webmaster Tools and submitting URLs via IndexNow so new content enters the index in hours rather than weeks. Then map each target query’s likely fan-out: a “how much does X cost” prompt spawns sub-queries about ranges, factors, and comparisons, and each deserves a section or page.
For the citation stage, write the direct answer in the first 40 words under every heading, use question-form H2s that match sub-queries, include current-year data with named sources, and keep pages dated and refreshed. Tables outperform prose for comparison content by wide margins in citation studies. One more stage-specific note: because fan-out runs sub-queries in parallel, a single deep page that covers a topic’s main question plus its natural follow-ups can earn multiple retrieval slots for one prompt, which compounds its odds of surviving the 15 percent citation cut. None of this is exotic; it is the discipline of answering rather than teasing, applied consistently. The 30-step version is in the 2026 GEO checklist, and the numbers behind each tactic are in AI search statistics 2026.
Frequently asked questions
Does ChatGPT search personalize results?
Results vary by conversation context, memory, and phrasing, so two users can get different sources for the same question. This is why single-prompt spot checks mislead; track your citation share across repeated queries over time instead.
Is ChatGPT search the same as SearchGPT?
SearchGPT was the prototype; its capabilities merged into ChatGPT as the integrated search experience. The retrieval architecture, Bing dependency included, carried over.
Does ranking number one on Google help with ChatGPT?
Only indirectly. ChatGPT retrieves from Bing, so Google positions do not enter the pipeline. Research has found only 6.82 percent of ChatGPT-cited pages appear in Google’s top 10 for the equivalent query. Bing indexing and answer structure are the direct levers.
How fast can a new page get cited by ChatGPT?
Pages entering Bing’s index via IndexNow can surface in ChatGPT search within days. Citation consistency builds over weeks as the page accumulates authority signals, on the timeline covered in how long GEO takes to work.
How do I know if ChatGPT is citing my site?
Track referral traffic from chatgpt.com in GA4, monitor your citation share with AI visibility tools, and check Bing Webmaster Tools’ AI Performance report, which shows Copilot citations from the same index.
If you want a query-by-query map of where your site enters and exits this pipeline, request a free analysis and we will run it engine by engine.
Tagged