Where ChatGPT Citations Come From

OpenAI does not run a web search for every prompt. As of October 2026, ChatGPT search is model-discretionary: the model decides, question by question, whether to search or to answer from what it already knows.

The two paths draw on different material:

• Retrieval on. The answer is built from pages fetched for this question, and those pages can be cited.
• Retrieval off. The answer comes from training data frozen at a knowledge cutoff. OpenAI's model docs list Feb 16, 2026 for gpt-5.6-luna and Aug 31, 2025 for the Instant path (chat-latest). No citations appear, and nothing you publish today reaches this answer.

OpenAI does not publish how often search fires. Third-party estimates disagree: Profound reports about 18% of conversations trigger a web search, Semrush's US clickstream put it at 34.5% in February 2026, and cloro measured 80.5% globally on a 100-prompt set. They measure different things on different populations, so none is a reliable rate.

What it means for you: if an answer you're checking shows no citations, read it as a parametric answer. It says nothing about whether your pages can be retrieved.

OpenAI's help center says ChatGPT search "typically rewrites your query into one or more targeted queries" and sends them to search providers. The same page says ChatGPT search "sometimes partners with other search providers". The provider named in OpenAI's documentation is Microsoft, but that list traces to 2024-era docs, and OpenAI does not publish the current mix. Our breakdown of OpenAI's help article on searching the web walks through the wording.

Query rewriting has three practical consequences:

• The queries that actually retrieve your page are model-generated and invisible to you.
• Memory can change them. OpenAI says that "if memory is enabled, ChatGPT may use relevant saved memories when rewriting a search query." Logged-in results are not reproducible.
• Any tool that tracks a fixed prompt list is measuring a proxy. That is also why AI search rarely searches the exact question a user types.

Being in a search index is necessary, not sufficient. In OtterlyAI's 2026 experiment, two AI-generated sites stayed in Bing and kept ChatGPT's crawlers visiting, yet ChatGPT citations were effectively zero. It was one pair of low-stakes sites, so treat it as a warning against "ChatGPT uses Bing" as a strategy, not as a measured rule.

ChatGPT reads more pages than it cites. OpenAI documents this in its API: the Responses API web_search tool returns a sources field with "the complete list of URLs the model consulted", separate from the inline citations. OpenAI adds that "the number of sources is often greater than the number of citations."

So every grounded answer has two sets:

1. Retrieved. Every page pulled into the model's context for that answer.
2. Cited. The subset the model chose to link.

The consumer app shows only the second set. The API is the only documented way to see the first, and the API is not the consumer app: retrieval behavior and model routing differ.

Citation is also an imperfect signal of influence. Mechanistic research on retrieval-augmented generation (arXiv 2606.28358) repaired over 90% of missed citations by amplifying specific attention components without changing the sources. A page can shape an answer and still go uncredited. Our guide to cited versus retrieved sources covers what that means for measurement, and the playbook on logging retrieved and cited sources via the API shows how to capture both sets.

OpenAI's complete public statement on ranking is this: "ChatGPT ranks search results using multiple factors intended to help users find relevant, reliable information. Placement is not guaranteed." Nothing is enumerated. As of October 2026, there is no documented ChatGPT ranking factor.

That makes most ChatGPT-specific advice unverifiable. Schema markup, llms.txt files, "AI-friendly" writing and chunking have no OpenAI documentation behind them. The July 2026 critical survey of 45 GEO studies (arXiv 2607.14035) found no technique with a stable, cross-platform causal effect on discoverability.

Two things have some support:

• Topical relevance. The same survey and a 252,000-trial factorial (arXiv 2605.25517) both found topical relevance and list position drive citation among pages already retrieved. You influence relevance, not position.
• Specific, quotable facts. A price, a number or a one-line definition stated on a fetchable page gives the model a sentence it can use.

For how these stages fit together across engines, see how AI search picks sources, and compare with how Claude picks sources, which cites text spans rather than whole pages.

You can't steer ChatGPT's ranking. You can make sure nothing blocks it from reaching you. OpenAI runs three crawlers, and "each setting is independent of the others."

A site can allow search and opt out of training at the same time:

User-agent: OAI-SearchBot
Allow: /

User-agent: GPTBot
Disallow: /

robots.txt is only half the gate. OpenAI's help center says eligibility requires allowing OAI-SearchBot and confirming your host or CDN allows traffic from OpenAI's published searchbot IP addresses. A firewall returning 403 removes you silently. OpenAI also warns against relying on IP allowlisting alone, so verify by user agent plus your CDN's verified-bot settings.

Rendering is the third gate. Vercel found ChatGPT's crawlers fetch JavaScript files but don't execute them, so content that appears only after JavaScript runs may be invisible. The AI Crawler Checker requests a page as each AI crawler and shows which ones your firewall turns away.

Work from the gates you control toward the ones you don't:

1. Read your robots.txt per bot. Confirm OAI-SearchBot is allowed. A blanket AI block often removes ChatGPT search eligibility by accident.
2. Test reachability. Request a key page with the OAI-SearchBot user agent and check for 200, not 403, 429 or a challenge page.
3. Diff raw HTML against the rendered page. Prices, product names and key facts should exist before JavaScript runs.
4. State facts plainly. Put the specific claims you want quoted on pages that can be fetched.
5. Check citations with retrieval visible. Only count answers that show citations, and remember the user's prompt is not the query ChatGPT ran.

For a quick snapshot, the free checker on this site shows whether AI answers name you for up to three buyer questions. What models say after that is not in your control. Access is.

Does blocking GPTBot remove my site from ChatGPT search? No. GPTBot governs training use only. ChatGPT search inclusion is controlled by OAI-SearchBot, and OpenAI states that each crawler setting is independent. You can disallow GPTBot to opt out of training while allowing OAI-SearchBot so your pages stay eligible for ChatGPT search answers. Allowing GPTBot, likewise, does not help you get cited.

Can I see which sources ChatGPT read but didn't cite? Not in the consumer app, which shows only citations. OpenAI's Responses API exposes a separate sources field listing every URL the model consulted, and OpenAI says that list is often longer than the citations. The API is a different surface from the app, so treat what it returns as an approximation of consumer behavior, not a copy of it.

Why does ChatGPT cite different sources for the same question? Several documented behaviors stack up. The model decides per question whether to search at all. When it searches, it rewrites the prompt into its own queries, and saved memories can shape that rewrite. OpenAI also says placement is not guaranteed. Two runs of the same prompt can retrieve different pages, or none, so single checks are snapshots.

If my site ranks well in Bing, will ChatGPT cite it? Not reliably. OpenAI says ChatGPT search sometimes partners with other search providers but does not publish the current mix or how results are chosen. In OtterlyAI's 2026 experiment, sites that stayed in Bing got effectively zero ChatGPT citations. Being indexed and reachable makes you eligible. It does not decide whether you get cited.