Honest measurement · October 26, 2026 · 13 min read
ChatGPT Citations: Which Sources It Used in 1,200 Sampled Answers
What ChatGPT cited in 1,200 sampled answers to our own buyer questions: how many links, which kinds of sites, how little it overlaps with Perplexity, and the limits of one small sample.
ChatGPT attaches source links to an answer when it searches the web. In our sample of 1,200 ChatGPT answers, those links concentrated on a small set of domains, leaned toward vendor sites, and overlapped little with what Perplexity cited for the same questions. The sample is small and it is ours. Read the numbers as one category’s picture, not as a rule about ChatGPT.
The sample, up front. 6,069 AI answers from seven engines, 1,200 of them from ChatGPT, to 48 prompts, from 23 August to 1 October 2026. Every answer is about brands we track in our own workspace. About 92% of the runs are in one developer software niche, and 78% are for one brand. Each prompt repeats daily, so the effective sample is the 48 prompts, not the thousands of answers. Results in your category will differ.
How ChatGPT citations work
ChatGPT answers a question in one of two ways. It answers from memory, from what the model learned in training, and shows no links. Or it searches the web, reads pages, and writes an answer with links to the pages it used. Those links are AI citations: the sources an assistant shows for an answer.
Some ChatGPT answers come with a web search and some do not, and a user can also ask it to search. Buying questions such as “best X for Y” often come back with links, but not every time. We cover both modes in ChatGPT SEO.
A citation is not a mention. ChatGPT can name your brand with no link to your site, or link your site and not name you. We count the two separately.
The sample
| Answers, all engines | 6,069 |
| ChatGPT answers | 1,200 (321 through the API, 879 from the app) |
| ChatGPT citations | 8,141 |
| Prompts | 48 (ChatGPT: 46 through the API, 41 in the app) |
| Dates | 23 August to 1 October 2026 |
| Source | Brands we track in our own workspace |
| Niche | About 92% of runs in one developer software niche |
| Largest brand | 78% of runs |
We read ChatGPT in two ways. The API: two OpenAI models (gpt-5-mini and gpt-5.6-luna) with web search turned on, 23 August to 4 September. The product: the answer the ChatGPT app itself shows for our prompt, 1 September to 1 October. The tables label this route “The product”, as they do for Copilot and the Google surfaces. The two ran in different weeks, on different models and a different brand mix (the largest brand is 37.7% of API runs and 84.4% of app runs). So a gap between the API row and the app row is not a clean API-versus-app effect.
What ChatGPT cited
Nearly every answer had links, about six each
| ChatGPT | Answers | Answers with citations | Mean citations | Median citations | Median distinct domains |
|---|---|---|---|---|---|
| The API | 321 | 94.7% | 6.08 | 6 | 4 |
| The product | 879 | 98.9% | 7.04 | 6 | 4 |
The API rate reflects our setup, because we turned web search on. For the app, we see only the answer it shows. In this sample, 98.9% of its answers to our commercial software questions carried links. ChatGPT also repeats domains: the app averaged 7.04 links but only 3.99 distinct domains per answer. One app answer carried 108 links, so the median of 6 is the safer figure.
Check in your category: ask your buyer questions in fresh chats, several times each, and count how many answers carry links. If most do, the cited pages are where your work goes. If few do, ChatGPT answers from memory, and a consistent description of you across the web matters more.
A small set of domains takes most of the links
| ChatGPT | Citations | Distinct domains | Top domain | Top 10 domains | Top 50 domains |
|---|---|---|---|---|---|
| The API | 1,951 | 311 | 16.0% | 46.3% | 77.9% |
| The product | 6,190 | 315 | 13.7% | 64.3% | 90.5% |
In the app, ten domains took 64.3% of all links, and fifty took 90.5%. Its top-10 share was the highest of all the routes we read. A small prompt set asked daily returns the same sources again and again, which inflates every concentration figure. Still, in this sample a short list of domains took most of ChatGPT’s links.
Do this: list the domains ChatGPT cites for your questions and count them. If ten domains carry most of the links, those ten are your priority list.
Vendor sites took a large share
We sort every cited domain into three groups: the tracked brand’s own site, a competitor site we configured, and every other site (third parties).
| ChatGPT | Own site | Competitor sites | Third parties |
|---|---|---|---|
| The API | 6.1% | 17.2% | 76.7% |
| The product | 14.0% | 28.0% | 58.0% |
In the app, 42% of links went to the tracked brand’s own site or a configured competitor’s site. Only three of the seven tracked brands have a competitor list, so the competitor share is understated. On those three brands alone, the app split was 14.7% own, 31.6% competitor and 53.6% third party. The third-party group also holds rival vendors nobody configured.
One bias pushes the other way. We wrote the prompts about the tracked brands’ niche, so their own sites appear more often than they would for neutral questions.
For contrast, Perplexity’s API sent 91.0% of its links to third parties, and Microsoft Copilot 92.1%.
Do this: if ChatGPT leans on vendor sites in your category too, your own pages are a real lever. Publish a page that answers each buyer question directly: pricing, comparisons, “best X for Y”.
Reddit, YouTube, Wikipedia and review sites barely appeared
Share of ChatGPT answers that cite the platform at least once:
| Platform | The API | The product |
|---|---|---|
| reddit.com | 4.0% | 0.2% |
| youtube.com | 0.0% | 0.0% |
| wikipedia.org | 1.2% | 0.3% |
| linkedin.com | 0.3% | 0.1% |
| g2.com | 0.9% | 1.4% |
| capterra.com | 0.0% | 0.0% |
| github.com | 44.2% | 26.7% |
In this sample, ChatGPT almost never cited Reddit or YouTube. Google AI Overviews cited reddit.com in 36.8% of answers and youtube.com in 32.3%, and Perplexity’s API cited reddit.com in 67.9%. G2 and Capterra were under 1% of citations on every route. Wikipedia appeared 14 times in 39,442 citations across all engines, which says more about our commercial software prompts than about Wikipedia.
GitHub needs a warning. Across the sample, 67.3% of github.com links have a tracked brand’s or competitor’s name in the URL. Most link to the tracked brands’ or their competitors’ own GitHub pages. That is an artifact of the sample’s niche, not a sign that ChatGPT favours GitHub.
Check in your category: do not assume that the platforms that matter on other engines matter in ChatGPT for you. Count them. A review-site push makes sense only if that site shows up in your citations.
Linked is not the same as named (read with care)
| ChatGPT | Named in the text | Own site cited | Cited, not named | Named, not cited |
|---|---|---|---|---|
| The API | 28.7% | 19.0% | 1.6% | 11.2% |
| The product | 21.2% | 37.9% | 17.3% | 0.6% |
In the app, 17.3% of answers linked the brand’s own site but did not name the brand in the text. Only 54.4% of app answers that cite the brand’s site also name the brand. On most other routes that figure is 86% or more.
On 6 September 2026 we changed how we detect a name, so that link text no longer counts, and we re-checked every earlier answer under the new rule. The API and app rows still ran in different weeks, on different models and brand mixes, so compare them with care.
Do this: track named and cited separately. If you are cited but not named, ChatGPT trusts the page but the answer does not say who you are. Put your brand name and a one-line description in the passage that gets quoted.
ChatGPT and Perplexity cite different sites
For the same 43 prompts in the same weeks, through both APIs, we compared the domains each engine cited. The measure is Jaccard overlap: the number of domains both engines cited, divided by the number of domains either engine cited.
| Comparison | Base | Mean overlap | Median overlap | Detail |
|---|---|---|---|---|
| API answers per prompt, pooled | 43 prompts | 0.114 | 0.100 | 2.7 shared domains, of 14.5 (ChatGPT) and 17.9 (Perplexity) |
| One answer each, same prompt, same day | 575 pairs | 0.137 | 0.091 | 26.4% of pairs shared no domain |
That is about 1 domain in 9 in common. ChatGPT’s API leaned toward vendor sites and large platform owners. Perplexity’s API leaned toward reddit.com (67.9% of answers, against 4.0% for ChatGPT), linkedin.com (13.4% against 0.3%), list-style blog posts and software directories.
Do this: pull a separate citation list for each engine your buyers use.
All engines, side by side
| Engine | Read through | Answers | With citations | Median citations | Top 10 share | Third-party share | Answers citing reddit.com |
|---|---|---|---|---|---|---|---|
| ChatGPT | The API | 321 | 94.7% | 6 | 46.3% | 76.7% | 4.0% |
| ChatGPT | The product | 879 | 98.9% | 6 | 64.3% | 58.0% | 0.2% |
| Claude | The API | 908 | 94.5% | 5 | 34.0% | 89.4% | 0.0% |
| Gemini | The API | 320 | 100.0% | 5 | 46.4% | 80.6% | 0.0% |
| Perplexity | The API | 321 | 100.0% | 10 | 38.7% | 91.0% | 67.9% |
| Microsoft Copilot | The product | 866 | 97.0% | 4 | 41.8% | 92.1% | 0.0% |
| Google AI Overviews | The product | 838 | 98.1% | 7 | 47.1% | 79.5% | 36.8% |
| Google AI Mode | The product | 828 | 98.8% | 6 | 53.6% | 85.0% | 20.8% |
The API rows ran 23 August to 4 September (Claude to 27 September), the product rows 1 September to 1 October. The rows also differ in prompts and brand mix: the largest brand is about 38% of the ChatGPT, Gemini and Perplexity API runs, 78% of the Claude API runs and about 84% of the product runs. So a gap between two rows is not a clean engine or route effect. Gemini’s API never returned more than five links, which looks like an API limit. Google AI Mode’s top domain is google.com itself (24.0% of its links). Perplexity read through its app is left out: that data is too thin to quote. Gemini read through its app is left out too: it covers one brand and 13 prompts.
What it means for your work
The cited sources are your to-do list. Each domain is a place ChatGPT read before it answered your buyer. Work down the list from the most cited domain: your own answer-first pages, better comparison pages than the rival pages that get cited, and the third-party articles, directories and review sites that appear in your list.
Do not copy our list. Reddit, review sites and Wikipedia may be large in your category even though they were small in our developer software niche. Measure your own category before you plan.
How to run the same analysis
By hand, it takes an afternoon:
- Write 15 to 25 buyer questions. Leave your brand name out of most of them.
- Ask each one in a fresh ChatGPT chat, with memory off, several times on different days.
- Copy every source link into a spreadsheet: question, date, URL, domain.
- Label each domain: your site, a competitor, a community, a review site, a reference site, or other.
- Count answers with links, links per answer, the top 10 domains’ share, and your own share.
- Repeat for Perplexity and compare.
AI citation tracking in CitePrism does this on a schedule. It stores every raw answer and groups every cited link by domain and category: own, competitor, review, social, institutional. It reports citation share per engine and over time, counts named and cited separately, and greys out metrics below 10 answers. ChatGPT brand monitoring reads ChatGPT through the API and as the app shows the answer.
What this sample can and cannot tell you
The sample: 6,069 answers from seven engines, 1,200 from ChatGPT, to 48 prompts, from 23 August to 1 October 2026. All answers are about brands we track in our own workspace. About 92% of runs in one developer software niche, 78% for one brand. There is no customer data in it.
It can tell you what ChatGPT cited for our prompts in this period, the large contrasts inside this sample (ChatGPT against Perplexity, vendor sites against third parties), and how to run the method in your own category.
It cannot tell you:
- What ChatGPT cites in general, or in your category. Results in another category will differ.
- Anything about one industry against another. Three of the four industries in the sample have about 5% of runs or less each.
- Statistical significance. The answers repeat the same prompts on consecutive days, so they are not independent. The effective sample is 48 prompts, not 6,069 answers.
- A pure API-versus-app effect, because the two ran in different weeks on different models and brand mixes.
- Trends over time.
- What real users see in their own chats. No tool can read real conversations; we sample, as every AI visibility tool does.
FAQ
What sources does ChatGPT use?
ChatGPT uses what its model learned in training and, when it searches the web, the pages it retrieves for that question. Only search answers show links. In our sample of 1,200 answers to commercial software questions, a short list of domains took most of the links, and 42% of the app’s links went to the tracked brands’ or configured competitors’ own sites. It rarely linked to Reddit, YouTube, Wikipedia or review sites. Sample your own questions before you act.
How does ChatGPT choose which sources to cite?
OpenAI does not publish a ranking formula for citations. When ChatGPT searches, it reads the results and links the pages it used. From outside, you see only the result: which pages it cited, how often, and for which questions. Sampling many answers turns that into a usable list.
Who does ChatGPT cite most often?
In our sample, the top 10 domains took 64.3% of the links in the ChatGPT app and 46.3% through the API. Vendor sites (the tracked brands and configured competitors) took 42% of app links. These figures come from one developer software niche and our own prompts, so they describe that niche, not ChatGPT in general.
Do ChatGPT and Perplexity cite the same sources?
Not much, in our sample. Through their APIs, for the same 43 prompts in the same weeks, they shared about 1 cited domain in 9. Perplexity cited reddit.com in 67.9% of answers and ChatGPT in 4.0%. Check each engine your buyers use, because work aimed at one engine’s sources may not move the other.