ChatGPT does not choose citations through one public, fixed ranking formula. OpenAI confirms that ChatGPT Search can search the web, use third-party search providers and partner content, and attach inline citations to its answer. OpenAI has not published the complete ranking logic or the weight of individual factors. The most useful way to publish for this system is to treat citation as a five-stage path: search activation, query rewrite and fan-out, retrieval eligibility, consulted sources, and final inline citation selection.
That model helps you diagnose where a page drops out. It is an operating model for publishers, not OpenAI's disclosed algorithm.
What has OpenAI actually confirmed about ChatGPT Search?
OpenAI says ChatGPT can search the web automatically when a question might benefit from current information, and users can also choose Search directly. Search answers include inline citations, plus a Sources panel that lets a reader inspect the cited publishers. OpenAI also says its search product uses third-party search providers and content supplied by partners. Its publisher and developer FAQ names Bing and Shopify among those providers. That is not the same as saying every answer comes only from Bing.
OpenAI's crawler documentation separates three relevant user agents. OAI-SearchBot is used to surface sites in ChatGPT search features. GPTBot is for content that may be used to improve generative foundation models. ChatGPT-User can visit a page because a user requested it, but it is not the control for Search inclusion. A publisher can allow OAI-SearchBot and disallow GPTBot because those choices are independent.
The public Web Search API documentation gives a further useful distinction. A search-enabled response can expose all URLs consulted through its sources field, while inline citations show only the references selected as most relevant to the written answer. This is documentation for OpenAI's developer API, not a full disclosure of consumer ChatGPT ranking. It still confirms an important point: being consulted and being cited are different states.
What is the five-stage citation operating model?
The five stages below turn the available documentation into a practical publishing model. They describe observable gates. They do not claim to reproduce OpenAI's private system.
| Stage | What happens | What a publisher can inspect | Common failure |
|---|---|---|---|
| 1. Search activation | The product decides whether web search is useful, or the user explicitly invokes it | Whether the answer includes current web sources and citations | The question is answered from model knowledge without a live search |
| 2. Query rewrite and fan-out | The system may reformulate or expand the question into searches | Search-call data in supported API workflows, plus repeated prompt testing | The page covers only one exact keyword and misses adjacent questions |
| 3. Retrieval eligibility | Search systems identify pages that can be accessed and considered | robots rules, response codes, rendered HTML, indexing and crawl logs | The page is blocked, unavailable, stale or absent from the searchable pool |
| 4. Consulted sources | A larger set of pages can be read while the answer is formed | The API sources field in supported implementations, or crawler and visibility evidence | The page is discovered but offers weak or redundant evidence |
| 5. Inline citation selection | A smaller set of sources supports claims in the final response | Citation URLs and cited passages in repeated answers | Another page supports the claim more directly, clearly or credibly |
This separation matters because teams often measure only the last line. If your page receives OAI-SearchBot visits but never appears in a citation, the problem is not necessarily crawl access. If it is cited but receives no sessions, the next problem is click intent and attribution. Our guide to turning visibility data into citations explains that wider measurement loop.
When does ChatGPT activate web search?
OpenAI says ChatGPT may search automatically when a request would benefit from web information. A user can also choose Search manually. The exact classifier and its thresholds are not public, so there is no responsible list of words that guarantees activation.
Freshness, explicit sourcing and current facts make web access more useful. A question about today's price, a new regulation or current product availability clearly needs live evidence. A timeless request can be answered without it. That distinction changes the opportunity available to a publisher. No search means no new page can be retrieved and cited at that moment, even if the page is excellent.
Publishers should therefore map prompts by evidence need, not just by keyword volume. Current comparisons, validation questions, product availability, definitions that change over time and questions asking for sources are natural search-shaped classes. Test the actual questions buyers use across multiple runs instead of assuming that every prompt triggers the same pathway.
Does ChatGPT rewrite one prompt into several searches?
OpenAI's Web Search API documentation shows that a search call can expose a search action and, usually, the query that was searched. Reasoning models can search, open pages, inspect them and decide whether to continue. That supports the idea that a complex answer can involve more than one retrieval action. It does not publish a universal rewrite recipe for consumer ChatGPT.
AirOps has published an observational vendor study of retrieval and fan-out. It analysed 15,000 original prompts, 43,233 original plus follow-up queries and 548,534 retrieved pages. AirOps reported that 89.6 percent of prompts generated at least two fan-out queries, and that 32.9 percent of cited pages found in a top 20 result appeared only for a fan-out query rather than the starting prompt.
Those figures are useful directional evidence, not OpenAI internal data or causal proof. AirOps sells software for this problem, its observation set is not a transparent random sample of all ChatGPT use, and its results should not be converted into a fixed ranking rule. The sound publishing lesson is narrower: one buyer question can branch into definitions, constraints, alternatives, prices, evidence and use cases. A strong page answers the important adjacent questions without becoming an unfocused encyclopedia.
What makes a page eligible for retrieval?
Eligibility begins with access. OpenAI says sites that opt out of OAI-SearchBot will not be shown in ChatGPT search answers, although they may still appear as navigational links. It recommends allowing the bot in robots.txt and allowing requests from its published IP ranges. OpenAI notes that robots changes can take about 24 hours to affect its systems.
Access is only the first gate. The page still needs a stable URL, a successful response, useful rendered text, a clear canonical and enough discoverability to enter a search provider's candidate set. Traditional search remains relevant because ChatGPT Search uses third-party search services as part of its wider supply chain. Our article on why websites fail to appear in ChatGPT covers those technical basics.
Do not confuse a crawler request with retrieval for an answer. A log line proves that a named user agent requested a URL. It does not prove that the page was selected for a particular prompt, consulted during answer generation, cited, clicked or converted. Those are separate measurements.
Why can a consulted page still be left uncited?
OpenAI's API makes the distinction explicit: the sources field can contain the complete list of consulted URLs, while inline citations represent the references selected as most relevant. A model can use several pages to understand a topic and cite only those that directly support claims in the visible response.
AirOps reported that only 15 percent of retrieved pages in its study received an inline citation. Again, that is vendor-run observational evidence, not a platform-wide citation rate. It does illustrate the size of the gap between discovery and final attribution.
A page is easier to select when a passage answers the precise question, states who is making the claim, gives current evidence, and can stand alone when quoted. Clear titles and question-led headings help a retrieval system locate the relevant passage. Tables help when a decision depends on comparable facts. First-party evidence helps when the claim is about your own product, while independent sources are stronger for validation and comparison.
This does not mean writing robotic fragments. It means giving each section a clear job. Start with the answer, show the evidence, state the limitation, then link to the primary source. Our practical guide to getting recommended by ChatGPT shows how that structure fits with entity clarity and independent proof.
How should you diagnose a citation problem?
Use the earliest failed stage instead of changing everything at once.
| Evidence you have | What it proves | Best next check |
|---|---|---|
| OAI-SearchBot can access the page | Automated search crawling is permitted | Confirm response, rendering, canonical and indexing |
| The URL appears in consulted sources | The page entered the research set for that run | Compare the passage against pages that received citations |
| The URL receives an inline citation | The answer used it as visible support | Check the claim, prompt class, engine and citation stability |
| An AI referral session lands on the site | A person clicked from an AI surface | Check landing-page intent and conversion events |
| A qualified event follows the session | The channel influenced a business outcome | Review attribution limits and repeat across a longer window |
Keep a fixed prompt set and record engine, location, date, answer, cited URL and the exact claim supported. Run it often enough to distinguish a repeatable pattern from one answer. Watch Bing Webmaster Tools too. Microsoft's AI Performance public preview reports cited pages and sampled grounding queries for supported AI experiences, while warning that citation counts do not indicate ranking, authority or placement inside an answer.
What should publishers change first?
Start with the stage you can prove is weak. Fix blocked or broken access before rewriting prose. If the page is retrievable but not cited, tighten the direct answer, make evidence current, separate claims into focused sections and strengthen primary sourcing. If citations occur but traffic does not, improve the reason to click and the landing path. If traffic arrives but does not convert, that is a proposition or page problem rather than a crawler problem.
Do not publish a hundred near-duplicate pages for imagined fan-out queries. One authoritative page can cover a coherent question cluster, supported by narrower pages where intent genuinely changes. Link them clearly so a search system and a person can move through the evidence.
Schmitdy is our own AI search service, so this is a disclosed maker's view. We use the five stages as a working diagnostic, then test the actual prompts and citations instead of claiming access to a secret formula. If you want to compare the measurement layer before hiring anyone, see our review of AI search visibility tools.
Frequently asked questions
Does ChatGPT use only Bing for search? No. OpenAI says ChatGPT Search uses third-party search providers and partner content, and its publisher FAQ names Bing and Shopify among providers. That does not justify saying that every search result or citation comes only from Bing.
Can I pay OpenAI to rank higher in ChatGPT Search? OpenAI's publisher FAQ says there is no way to guarantee top placement. Search visibility should be treated as an earned and measured outcome, not a purchasable organic rank.
Does allowing OAI-SearchBot guarantee a citation? No. It makes a site eligible to appear in ChatGPT search answers. A crawl request does not prove retrieval for a prompt, consultation, inline citation, referral traffic or conversion.
What is the difference between a consulted source and a citation? In OpenAI's Web Search API, consulted sources are the complete URLs returned in the sources field for a search action. Inline citations are the selected references attached to claims in the final response. The sources list can therefore be longer than the citation list.
Is the five-stage model OpenAI's official ranking algorithm? No. It is Schmitdy's operating model built from documented product behaviour and observable evidence. OpenAI has not published its complete ranking formula or factor weights.
