Blog / AI & search /
ChatGPT search a month on: who is it citing?
It’s been a month, and people are already asking which sites ChatGPT Search cites, as if it were a mature search engine with a stable pattern. Fair question. But a month is not enough to turn a handful of answers into a ranking report. Results change with the wording, topic, location, time and product version.
OpenAI introduced ChatGPT Search on 31 October 2024. The launch announcement described answers with links to relevant web sources and said the search model used third-party search providers as well as information supplied directly by partners. Access began with paid plans and was rolling out, so not every reader necessarily saw the same product at the same time. So treat it as a new interface to watch, not a list of ranking factors.
What does “citing” mean here?
A citation is a link attached to a claim or answer. It lets the reader inspect the source, but a link doesn’t prove it was the only input, or that the answer works like a normal results page. Read the actual response and the linked page. Is the source relevant? Does it support the statement beside it? Is it the original source or a page repeating someone else’s work?
Also distinguish inline citations from additional links shown in a sources panel. A source can be offered as further reading without being the page that supports a particular sentence. Counting every visible link as an equal recommendation distorts what you think is happening.
Why one test cannot settle the question
Ask the same question two ways and you may get two answers. “Best running shoes” has many readings; a specific budget or use case narrows it. The tool may also decide it doesn’t need to search at all. One screenshot is a poor basis for claims about anyone’s visibility.
The interface and access were changing during the first month. OpenAI said the rollout was starting with certain paid plans, and the announcement described plans to expand access later. A person’s account, region, query and the day of the test can all matter. If you are recording results, note those conditions. Otherwise you may compare two different experiences and call the difference a ranking change.
Small samples bring a selection problem too. Run ten prompts about your own industry, remember the surprising ones, and you don’t have a neutral sample. Test only queries where you already rank in Google and you’ve framed the experiment around one channel. I’d want a clear question and a repeatable method first.
A modest way to inspect citations
Choose a set of questions real customers might ask, not only queries written to make your own pages appear. Keep the prompts and dates. Repeat them later, and save the response with its citations. Don’t publish a precise percentage from a small set as if it describes all of ChatGPT Search. Call it an exploratory check and say how you gathered it.
For each answer, inspect the links. Record whether the source is an original document, a publisher, a forum, a retailer or another kind of page. Check if the cited page actually supports the answer. If a page is absent, do not conclude it is excluded from the system. It may not have been retrieved for that phrasing or may simply have lost out among many relevant pages.
Compare what the cited pages offer. Are they clear, specific and easy to verify? Do they have first-hand detail? That isn’t a recipe for “optimising for ChatGPT”. It’s a way to see whether your own page serves the same need. Don’t copy competitors’ phrasing just because a cited page has it.
What site owners can do now
Keep the foundations sensible: publish accurate pages, make important information accessible to crawlers, use clear titles and headings, and link related pages so people can navigate. That work helps users and supports discoverability across systems, though it cannot guarantee a citation in any particular answer.
If the answer includes a citation to your page, check the page for accuracy and the linked context. If it does not, continue improving the page for the audience it serves. Search Console remains useful for Google Search performance; it does not report a ChatGPT citation count. Do not mix those measures or interpret a change in one as proof of a change in the other.
Early weeks invite bold claims, because the interface is visible and the mechanics aren’t. I’d resist turning observation into certainty. Keep records, test more than one prompt and keep your conclusions in proportion. A month raises questions worth chasing. It can’t tell us what the product will cite next year.
Related reading
- What do AI tools say about your site? A quick test
- AI agents for SEO: hype versus reality
- Using AI to write schema markup
If you test ChatGPT Search, what questions will you repeat so you can tell a real pattern from an interesting one-off?