Why do some ChatGPT answers about my brand have no sources?

ChatGPT works in two modes, and it picks one for every query. In the first mode it searches the web, reads pages and cites them. In the second it answers only from what the model absorbed during training. That second mode has nothing to link to, so the answer shows no sources.

The distinction matters because the two modes have different inputs and different fixes. We call them retrieval-on and retrieval-off answers.

• Search on: ChatGPT fetches pages at answer time. If your page is reachable and states the fact clearly, it can be part of the answer.
• Search off: ChatGPT recalls training data frozen at a knowledge cutoff. Anything you published or changed after that date is invisible to this mode.

So an unsourced answer with an old price or a retired product name is not reading your site at all. It is repeating what the model learned months ago. Our guide on why search-on and search-off answers differ walks through more examples.

As of October 2026, ChatGPT search is model-discretionary. The model decides whether a question needs fresh information. Questions about recent events, prices or local details are more likely to trigger a search. General or definitional questions are more likely to be answered from memory. OpenAI does not publish the rule.

When it does search, OpenAI's help article says ChatGPT "typically rewrites your query into one or more targeted queries" before sending them to search providers. You never see those rewritten queries.

The cutoff also varies by which model answers. OpenAI's model documentation, checked in August 2026, lists a February 16, 2026 knowledge cutoff for the GPT-5.6 models and August 31, 2025 for the Instant path (chat-latest). The same documentation warns that these snapshots are regularly updated. "What ChatGPT knows" is not one fixed thing, so the same brand question can draw on knowledge nearly six months apart depending on routing.

Nobody agrees on how often ChatGPT searches. Semrush's US clickstream data put it at 34.5% of queries in February 2026. Profound reported that about 18% of conversations triggered at least one web search. cloro reported 80.5% globally on a 100-prompt set it chose. These studies measure different things on different populations, and none answers the brand-level question you care about.

Citations also undercount what the model read. OpenAI's API documentation says the number of sources consulted is often greater than the number cited. Research on how language models cite (arXiv:2606.28358) found that citation output can be changed without changing the sources, so a citation is a noisy signal of influence.

What this means for you: any single number for "how ChatGPT sees my brand" averages two systems you can't separate after the fact. Treat search-on and search-off answers as separate results.

A missing citation usually means no search ran, but not always. If you didn't see citations, you don't know for certain what happened.

• Memory and custom instructions. OpenAI says ChatGPT may use saved memories when rewriting a search query. A logged-in test can return a different answer from a fresh session.
• Navigational links. OpenAI's crawler page says sites opted out of OAI-SearchBot "will not be shown in ChatGPT search answers, though can still appear as navigational links." A link is not the same as a cited source.
• User-triggered fetches. When a user pastes a URL, ChatGPT-User fetches it. OpenAI says robots.txt rules "may not apply" to that agent, and it is not used to decide what appears in search.
• Blocked crawlers. If your site blocks OAI-SearchBot, a search can run and still never use your pages. The answer then cites other sites, or none.

Sort every wrong answer by mode first. Then fix what you can reach.

For answers with sources

1. Check that robots.txt allows OAI-SearchBot. Blocking GPTBot only stops training use. OpenAI says each setting is independent.
2. Check that your CDN or firewall doesn't return a 403 to OpenAI's crawlers. The AI Crawler Checker requests a page as each AI crawler and shows which ones get turned away.
3. Put the specific fact (price, plan name, founding year) in the HTML before JavaScript runs. Vercel's crawler study found ChatGPT's crawlers fetch JavaScript files but don't execute them.
4. Look at which third-party pages get cited instead of yours, using this method for finding the pages AI engines cite.

For answers without sources

Nothing you publish this quarter reaches them. The only path in is a future training run, and nobody has measured whether allowing training crawlers changes what a model later recalls. Make sure your current facts are public and consistent, then retest when OpenAI ships a new model. Our page on fixing ChatGPT errors in trained knowledge vs search covers both paths.

Record search state in every test

Log the date, the model, whether you were logged in, and whether citations appeared. Through the API you can switch search on or off yourself and log retrieved vs cited sources. For a step-by-step setup, follow testing AI answers with search on and off.

Can I force ChatGPT to search before answering about my brand? In the ChatGPT app, a user can ask it to search, and questions about current prices or news tend to trigger search. You can't control what other people's sessions do. In OpenAI's API you can enable the web search tool directly, which makes search-on testing repeatable. That is why API tests with an explicit search setting are easier to compare over time than screenshots from the app.

Does blocking GPTBot remove my brand from ChatGPT answers? No. OpenAI documents GPTBot as the training crawler and OAI-SearchBot as the search crawler, and says each setting is independent. Blocking GPTBot signals that your content shouldn't be used for training. It doesn't remove you from ChatGPT search, and it doesn't erase what an existing model already learned. To appear in search answers, OAI-SearchBot has to be allowed and able to reach your pages.

How long until a training-data error about my brand goes away? Nobody can say. A training-data answer changes only when OpenAI ships a model with a newer cutoff, and OpenAI doesn't publish what each training run includes. Even then, nothing guarantees the new model recalls your corrected facts. The practical move is to keep accurate facts public and consistent, and retest after each model update with search off.

Are answers with sources more accurate than answers without them? Not automatically. A sourced answer can still cite a weak or outdated page, and citations don't always reflect what the model relied on. The advantage is that a sourced answer can be traced and fixed, because you can see which pages fed it. An unsourced answer has no page to correct. Judge accuracy per answer, and note the search state each time.