When a buyer asks ChatGPT “which term plan should I buy?” or “which protein bar has the least sugar?”, the answer is usually a short list of brands. This guide explains what is known about how that list gets made, what isn’t known, and how you can see it for your own category.
Two ways ChatGPT can answer
ChatGPT can answer a question in two different ways, and the difference matters for brands.
From memory. Without searching, ChatGPT answers from what its model learned during training. That knowledge stops at a training cutoff, so a brand launched or renamed after that date may simply be unknown to it.
With web search. OpenAI says ChatGPT “will automatically search the web if your question might benefit from information on the web”, and that users can also switch search on themselves. Answers that use search carry inline citations and a list of sources. This is called web search, or grounding.
How often does it search? A study by Ahrefs of 66.6 million ChatGPT responses, collected from free, logged-out sessions in August and September 2026, found that about one in seven responses contained “fan-out” queries: the shorter searches ChatGPT writes for itself. (Ahrefs first described these as the responses that searched the web, then corrected that wording.) The figure covers all kinds of questions, not only buying questions, so we don’t know how often a buying question in your category triggers a search.
What happens when it searches
OpenAI hasn’t published a full description, but some parts are documented and others have been measured.
- It runs one or more searches. In the same Ahrefs study, responses that contained fan-out queries had about 1.6 of them on average. These are shorter searches written by ChatGPT, not the buyer’s exact words.
- It uses outside search providers. OpenAI’s help page says ChatGPT may share search queries with “third-party search providers such as Bing”.
- It can only show sites its crawler may read. OpenAI’s crawler documentation says sites that block its search crawler, OAI-SearchBot, “will not be shown in ChatGPT search answers”.
- It cites what it used. The answer links to the pages it drew on, and a Sources button lists them.
What OpenAI does not say is how it weighs one page against another, or how it decides which brands from those pages make the final list. Anyone who tells you the exact rules is guessing.
Why the same question gets different answers
Ask ChatGPT the same question twice and you will often get two different lists. The model writes each answer fresh, the searches it runs can differ, and the pages it finds can differ. A brand can be named in one answer and missing from the next, with nothing on its website having changed.
That is why one screenshot proves very little. To see where a brand really stands, you need to ask the same question several times and count. Citationly asks every question three times for exactly this reason.
Which websites the answers lean on
Published studies give part of the picture, and they don’t all point the same way.
- US data, all topics. Ahrefs tracks which domains ChatGPT cites across a broad set of US queries. In its September 2026 table, Reddit was the most-cited domain and Wikipedia second.
- A shift towards official sources. Ahrefs’ fan-out study also reports that, in its US citation data from March to September 2026, citations moved towards official sources such as regulators, research bodies and brands’ own pages, and that citations of Reddit and Wikipedia fell sharply, with YouTube’s falling too.
What our own runs found. Citationly runs for an insurer and two packaged-food brands in India recorded 222 answers in all. They came from OpenAI’s model through its API with web search on, between late September and early October 2026, in reply to buying questions. The most-cited websites were brands’ own sites, and for insurance, the regulator’s site. Reddit was cited in 2 of those 222 answers and YouTube in 1.
We don’t know which of three things explains the gap with the US studies: the API versus the ChatGPT app, the Indian market, or the style of question. So we don’t treat either result as a rule. The safe conclusion is narrower: the websites that matter differ by category and market, and you have to look at your own.
What this means for your brand
From the documented facts and the measurements above, four practical points follow. None of them promises a mention.
- Let OpenAI’s search crawler in. If your robots.txt blocks OAI-SearchBot, your pages can’t appear in ChatGPT’s search answers.
- Be findable through the search it uses. Make sure your site is indexed in Bing as well as Google. Our 15-minute checklist covers this.
- Answer the buyer’s question on your own pages, in the words buyers use, with the facts written as text a crawler can read.
- Be present on the other websites your category’s answers cite. If the answers that leave you out keep citing the same review site, comparison page or regulator’s list, that page is part of the decision.
There is also a slower path. Brands that are widely written about are more likely to be known to the model from training, but that only shows up when a newer model is trained, and nobody can schedule it.
How to check your own category
You can see this for yourself in half an hour:
- Write five questions a buyer in your category might ask, in their words.
- Ask each one three times, starting a new chat each time.
- Note every brand named and every website cited.
- Count how many of the answers named you, and look at which websites the answers used when they didn’t.
Our guide to running a basic AI visibility audit walks through it step by step.
How Citationly helps
Citationly does this at full scale: questions prepared for your category, each asked three times with web search on, every answer recorded word for word. The report shows which brands the answers name and which websites they cite when they leave you out. Book a demo to see one, or check whether ChatGPT can see your website for free.