Being found is not the same as being recommended. ChatGPT can know you and still not recommend you. Ask for a recommendation and you get one name, sometimes three. Why those? How the search before that works is covered in “How Do I Become Visible in ChatGPT?” Five insights:
01. You don't see what ChatGPT really reads
The sources under an answer do not show everything ChatGPT took into account for that answer.
OpenAI says so itself: the citations in a ChatGPT answer are not all the sources ChatGPT took into account.
A recent analysis of real ChatGPT conversations shows it: only 13.3 percent of all returned search URLs end up cited.
What we do not know: how much an uncited source influences the answer. It may play a role. Nobody has measured it yet.
Being visible on the web and being visibly cited are two different things. The citations are only a slice. The rest is uncharted waters.
Sources and links
OpenAI: Web search, API documentation, Sources section
Amani et al.: Characterizing Web Search by Conversational LLM Agents, arXiv:2609.19244, preprint, September 2026, ChatGPT citation rate 13.3%
02. Different question. Different answer
Which source ChatGPT cites depends on the actual search query. And that can differ from the prompt you typed.
ChatGPT Search uses query rewriting: a user's question gets rewritten into one or more targeted searches. That produces different terms and a different focus than the original prompt. A page can be relevant for one search and drop out of the selection for another.
There is no fixed list of sources for a given search in ChatGPT.
The question is part of the selection. Change it, and you change the selection.
Sources and links
OpenAI: Searching the web with ChatGPT, query rewriting, Memory and location
Vishwakarma et al.: What Gets Cited, arXiv:2605.25517, 252,000 controlled pairwise trials across six models
Ahrefs: 15,000 long-tail queries, ChatGPT in-text citations 8.0% and references 6.1% in Google's top 10 for the original prompt
03. You describe yourself. So does the web
Your website is a first-hand description of your company. It is not the only one.
Trade media, news, directories, profiles, vertical sources, other websites: they all describe you.
And those are exactly the sources that show up in AI answers most often: official sources, news and vertical-specific websites.
Your website has no monopoly on you. ChatGPT finds what you say about yourself and what others say about you, and it leans more on well-known, authoritative domains than on the average page from the search.
For ChatGPT, the picture of an entity comes together from several descriptions. Yours is one of them.
Good Answer Engine Optimization and Generative Engine Optimization strengthen your own picture amid the many external signals.
Sources and links
Zhang, He, Yao: From Citation Selection to Citation Absorption, arXiv:2604.25707, 21,143 citations, 79.12 to 87.52% from official, news or vertical sources
Amani et al.: Characterizing Web Search by Conversational LLM Agents, more authoritative domains are cited more often than the average returned search result
Kumar, Palkhouski: AI Answer Engine Citation Behavior, GEO16 Framework, arXiv:2509.10762
04. The final answer is still a black box
OpenAI documents it: ChatGPT ranks search results using multiple factors. Placement is not guaranteed. It does not say which factors.
What we know: ChatGPT retrieves search results, opens pages and refines the search. What we do not know: how those results are weighted in the answer, and by what mechanism a source gets visibly cited or a company gets recommended.
In short: we know how ChatGPT searches. We do not know how the final selection is made from those results.
That is all.
Sources and links
OpenAI: Searching the web with ChatGPT, ranking factors and non-guaranteed placement
OpenAI: Web search, API documentation, search, open and find mechanics and search context
05. ChatGPT's favorite today. Not tomorrow
If ChatGPT recommends you the first time, that does not mean you get recommended the second time.
Conversation context within the chat, Memory, location and the specific search all change the starting point.
ChatGPT isn’t rolling the dice. But the conditions change from one request to the next.
There is no large repeat test for ChatGPT. What exists are individual studies of its search behavior. And they show: sometimes it searches, sometimes it does not.
A single AI answer is not a ranking. The more clearly your entity is structured, the clearer the signals ChatGPT gets about you. Good Answer Engine Optimization and Generative Engine Optimization strengthen exactly those signals.
And the stronger they are, the more often ChatGPT chooses you.
Sources and links
OpenAI: Searching the web with ChatGPT, conversation context, Memory, query rewriting and location
Amani et al.: Characterizing Web Search by Conversational LLM Agents, strongly varying search behavior across platforms and models
Strauss et al.: The Attribution Crisis in LLM Search Results, arXiv:2508.00838, search mode and answers without an explicit web fetch