How ChatGPT chooses its sources, from 45,000 citations
A data-led look at how ChatGPT chooses its sources, which pages and domains it cites, and what to check when your own site rarely appears.
By CiteGraph · 12 Sept 2026 · updated 2 Oct 2026 · every number below has a source
The short answer to how ChatGPT chooses its sources: in our sample, it cited pages that matched buyer questions, especially listicles, roundups and competitor product pages. Those three types accounted for 91.9% of citations.
The rest of this post shows which page types recur, which domains keep appearing, and how to check the gap for your own site. The figures come from CiteGraph's research page, based on a measured sample rather than every ChatGPT answer. We cannot see what happens outside that sample or inside the model's selection process.
The citation pattern is heavily concentrated
We recorded 45,309 citations across 7,260 AI answers. Those answers came from 60 scans of 41 sites in 44 categories, covering 442 distinct buyer questions. The engines cited 7,624 domains and 16,506 distinct pages.
Listicles and roundups accounted for 67.5% of all recorded citations, or 30,585 citations. Competitor product pages supplied another 24.4%, or 11,052 citations.
Together, those two page types made up 91.9% of citations. Every other category shared the rest. The page-type data in full is in The pages AI cites.
Two page types dominate the recorded citations: listicles and roundups, followed by competitor product pages.
That points to a practical priority: if your buyers ask for options, your product needs to appear in the comparison pages already answering those questions. Your own site can supply product facts, but it is competing with pages that have already organised alternatives for the reader.
The rest of the sample was scattered. Community threads, including Reddit and forums, made up 3.4% of citations. Documentation accounted for 1.1%, social posts for 0.8%, review platforms for 0.8%, YouTube for 0.4%, news for 0.1%, and Wikipedia for 21 citations, recorded as 0% of the total share.
ChatGPT sources recur by domain and format
Page type explains the broad shape. Domain recurrence shows where usable material keeps appearing across different questions.
Reddit was the most cited domain, with 1,482 citations across 44 categories. GitHub followed with 481 citations across 26 categories. LinkedIn had 338 citations across 32 categories, and Apple had 323 across 21 categories.
Several product and publishing domains also appeared repeatedly. Virlo.ai was cited 304 times across 3 categories. Otterly.ai was cited 285 times across 1 category. We do not know why these domains appear so narrowly, but their category counts suggest they should be read as specific signals rather than broad evidence of market demand.
Medium.com appeared 267 times across 31 categories, while TikTok appeared 230 times across 5 categories. G2 had 203 citations across 30 categories, and YouTube had 201 across 31 categories.
The most cited domains recur across different categories, led by Reddit and GitHub.
The ten most cited domains accounted for 9.1% of all citations. Across the full scan, 885 domains had at least ten citations and 84 had at least fifty.
That is a long tail, with a small group of familiar domains sitting above it. For your category, start with the domain type that matches the question: comparison pages for buying decisions, community threads for lived experience, or documentation for implementation detail.
The domain list also makes one point clear: ChatGPT sources are not limited to company websites. Public communities, platforms, publishers and product databases all contribute material, while a brand's own domain accounted for 1.5% of citations in this sample.
The methodology page explains how CiteGraph measures the scans and classifies the results.
Your own site is a small part of the answer
The scanned product's own site accounted for 663 citations. That is small beside the 30,585 citations from listicles and roundups, and the 11,052 from competitor product pages.
This does not mean your website has no role. It means the site was rarely the main source of citations in the sample.
A product page can still be useful when the question asks for a specific feature, price, integration or limitation. Documentation can answer implementation questions. Your site may also be named without being cited, or cited in a different category from the one you care about.
Across the sampled answers, the ChatGPT family cited an average of 3.4 sources per answer. The scanned product was named in 4.4% of ChatGPT-family answers. A name appearing in an answer and a page receiving a citation are related outcomes, but they are not the same measurement.
One objection: the page-type split may reflect the questions we asked, rather than how ChatGPT picks sources. That is fair, because buyer questions naturally produce comparisons, roundups and product pages. It still gives you a useful starting point: for the questions your buyers ask, inspect the pages already appearing in the answers.
Check the gap before you publish more
You can turn this data into a quick review without starting a large content project. The aim is to find whether your site is missing because of page type, coverage or simple absence from the cited set.
Collect the buyer questions your sales team hears most often, then group them into comparisons, alternatives, use cases and product facts.
Search the answers you care about and record the cited domains, page types and whether your product is named or cited.
Compare your pages with the recurring source types, then make a page easier to quote rather than publishing a broad batch.
Recheck the same questions after the page has had time to be found, and keep the question wording unchanged.
This fits into under an hour if you start with a small set of high-value questions. The free AI visibility checker can help you inspect where your brand appears, while the AI readiness check gives you a separate view of basic availability.
Look for a specific mismatch. You may have product pages but no clear comparison page. You may have documentation but no plain answer to the buyer's question. Or you may have the right page, but it is not being cited by the sources that dominate your category.
Do not treat the domain list as a list of sites to chase blindly. Study the reason each source fits the question. A forum thread may contain the exact complaint buyers use. A roundup may provide the comparison structure. A product page may state a fact with less ambiguity than a long article.
Questions founders ask about ChatGPT sources
How does ChatGPT choose its sources?
In our sample, citations were concentrated in listicles and roundups, competitor product pages and a smaller set of recurring domains. The pattern suggests that fit with the buyer's question matters more than ownership of the brand.
How do we know if the source is reliable?
Check whether the page supports the claim, fits the buyer's question and agrees with other credible sources. Citation frequency shows recurrence, not quality, so a frequently cited domain still needs review.
Why doesn't ChatGPT cite my website?
Your site may be competing with comparison pages and competitor product pages, which together accounted for 91.9% of citations in our sample. The scanned product's own site accounted for 1.5%, so check whether your pages cover the questions and formats that already attract citations.
Rewrite a comparison or product page so it answers a high-value buyer question plainly.
Set a date to recheck those same questions, using the same wording.
Questions people ask
What types of pages does ChatGPT cite most often?+
Listicles and roundups accounted for 67.5% of citations in the sample, while competitor product pages supplied 24.4%. Together, these page types made up 91.9% of all recorded citations.
How can I check whether a ChatGPT citation is reliable?+
Check whether the page supports the claim, fits the buyer's question and agrees with other credible sources. Citation frequency shows recurrence, not quality, so frequently cited domains still need review.
Why is ChatGPT citing competitor pages instead of my website?+
Listicles, roundups and competitor product pages accounted for 91.9% of citations in the sample, while the scanned product's own site accounted for 1.5%. Check whether your pages cover the buyer questions and formats that already attract citations.