ChatGPT Sources by Domain: Five Citation Studies, One Merged Table

Reddit and Wikipedia are the two most cited domains in every ChatGPT sources study we could find, and that is about all the studies agree on. Reddit's measured share runs from 1.8% of all citations in one dataset to 16.8% in another. We put five published domain-share tables side by side, with each study's sample and metric next to its numbers, so you can see which differences are real and which are arithmetic.

ChatGPT sources: the short answer across five studies

If you need one sentence for a slide, use this one. ChatGPT cites Reddit and Wikipedia more than any other domains, followed by a long tail of publishers, retailers, reference sites and the vendors themselves.

Direct answer

Which sites does ChatGPT cite most? Across five studies published between June 2025 and September 2026, Reddit and Wikipedia rank first and second every time. Behind them sit Forbes, YouTube, LinkedIn, Reuters and large retailers. The exact percentages differ by up to 9 times between studies, because each one divides by a different total.

That last point is the reason this post exists. Most coverage of ChatGPT's most cited websites quotes one study's number as if it were the share of the web. It is not. Each figure is a share of something specific, and the something changes from study to study.

Our view is blunt. A single ChatGPT citation percentage, quoted without its denominator, is close to useless for planning. The ranking of domains is reliable. The size of each domain's slice is not, and anyone selling you a strategy built on the size of the slice should be asked which study, which month, and divided by what.

How the five ChatGPT citation studies were built

We used every study we could find that published a domain-level table for ChatGPT and stated at least its date window. Four are run by companies that sell AI visibility tools. One is run by an agency that sells generative engine optimisation services. None is peer reviewed. That is the state of the evidence, not a choice we made.

Compiled from each study's own methodology statement, retrieved 28 October 2026. Similarweb's total is our calculation: 118,285 Wikipedia events at 13.15% implies about 899,500 events. Its intro gives a different figure of nearly 600,000 across ChatGPT and Google AI Mode combined.

Profound's study is the largest by volume. It covers 680 million citations from August 2024 to June 2025 across ChatGPT, Google AI Overviews and Perplexity. It publishes two tables for ChatGPT, and confusing them is the most common error in this whole topic.

Semrush tracked more than 230,000 prompts in weekly snapshots across 13 weeks. Its value is the time series. It caught a sharp change in September 2025 that no single snapshot could show.

Similarweb and Ahrefs both publish from commercial monitoring tools. Similarweb reports raw citation counts, which lets you rebuild its denominator. Ahrefs reports share within its own top 50 list and publishes no sample size, which makes it the hardest of the five to check.

The Ahrefs page is also regenerated monthly by an AI agent, according to the page itself. The figures here are the September 2026 snapshot we retrieved. They will have moved by the time you read this.

DerivateX is tiny by comparison: 40 prompts, one per software category, run in ChatGPT with web search enabled. It is the only study of the five that isolates B2B buying queries. Its small sample is the price of that focus, and we treat its percentages as directional for that reason.

The merged ChatGPT domain-share table

This is the table the topic has been missing. Every cell carries the study's own number in the study's own unit. Read across a row to see how consistent a domain's position is. Do not add or average across a row, because the units differ.

Sources: Profound overall-citation table (2025), Semrush text and chart commentary (November 2025), Similarweb top 20 table (April 2026), Ahrefs top 50 table (September 2026 snapshot), DerivateX source list (June 2026). Semrush gives exact values only in chart images, so its cells quote the text. Retrieved 28 October 2026.

We summed each study's published rows to see what share of all citations its table actually covers. Profound's top 10 domains add up to about 16.3% of all ChatGPT citations. Similarweb's top 10 add up to 47.9%, and its top 20 to 63.6%. Ahrefs covers 100% by definition, because its denominator is the list itself.

That gap matters more than any single domain's rank. In Profound's data, 5 of every 6 ChatGPT citations go to sites outside the top 10. In Similarweb's monitored prompts, the top 10 take nearly half. Those are two different pictures of how concentrated ChatGPT's sourcing is, and they cannot both describe the same population of queries.

Where the ChatGPT citation studies agree

Strip out the percentages and keep only the rank order, and the five studies converge. Three findings hold in every study that measured them.

Reddit and Wikipedia are first and second everywhere

Profound, Semrush, Similarweb and Ahrefs all place Reddit and Wikipedia in the top two for ChatGPT. Semrush's text says that before September 2025, ChatGPT cited them about five times as much as the next most cited domains. Even after the September drop, both stayed at the top of its ChatGPT list.

The order between the two is what moves. Wikipedia leads in Profound and Similarweb. Reddit leads in Ahrefs, by a wide margin. We return to why below.

Forbes is the most consistent publisher

Forbes is the only commercial publisher named in all five studies. It holds 1.1% of all citations in Profound, 1.38% of events in Similarweb, and 3.1% of the Ahrefs top 50. Semrush lists it among ChatGPT's biggest gainers after September 2025, and DerivateX found it cited 4 times on software buying prompts.

That consistency is worth more than any single high number in the table. A domain that appears across different windows, prompt sets and metrics is telling you something about how ChatGPT ranks sources, not about one study's sampling.

The tail is long, and it is where most citations go

Every study that published enough data to check shows the same shape. A few domains sit well clear at the top, and the bulk of citations spread across thousands of smaller sites. Profound's figure is the starkest, with 83.7% of citations landing outside its top 10.

For a B2B marketer, that is the encouraging finding in this data. You are not competing with Wikipedia for a fixed pie. You are competing for a place in the tail on the queries your buyers actually type, which is a much smaller fight. How cited pages get picked from that tail, including the finding that most rank poorly on Google, is the subject of our analysis of what ChatGPT citations follow instead of Google rank.

Why ChatGPT's most cited websites disagree across studies

The disagreements have three causes, and only one of them is a real change in ChatGPT's behaviour.

Cause 1: the denominator

Profound's Wikipedia figure is the cleanest example. The same study reports Wikipedia as 47.9% of ChatGPT's top 10 citations and 7.8% of all citations. Both are correct. The 47.9% is the number that travels, usually without the words "of the top 10".

Semrush adds a third unit. Its percentages count the share of responses that cite a domain at all. A domain can appear in 55% of responses and still hold a modest share of citations, because one response carries many links.

Cause 2: the query set

The ratio in the chart above removes the denominator problem, because it compares two domains inside the same study. Reddit gets 0.23 citations per Wikipedia citation in Profound, 0.91 in Similarweb, and 2.4 in Ahrefs. That tenfold spread is still there.

Part of it is what each panel asks. Ahrefs describes a broad, search-backed US query set, and its top 50 is full of shopping, health and home sites: Consumer Reports, Walmart, Home Depot, Healthline. Product questions pull in Reddit threads. Definitional questions pull in Wikipedia. A panel heavy on one kind of question will tilt the ratio, and none of the five studies publishes its prompt list.

Cause 3: ChatGPT itself changed

The one disagreement that is clearly real is time. Semrush watched Reddit fall from close to 60% of ChatGPT responses to about 10% between early August and mid-September 2025. Wikipedia fell from roughly 55% to under 20% in the same weeks.

Semrush's own analyst says the data does not explain why. One proposed cause, Google's removal of the parameter that returned 100 results per page, was judged not to be the root cause, or at least not the only one. What it does show is that a citation share measured before September 2025 describes a different ChatGPT from the one your buyers use now.

So the sequence in the chart, Reddit rising from Profound to Ahrefs, is consistent with a real trend. It is equally consistent with different query sets. With this evidence, you cannot tell those two explanations apart, and we would not plan a budget on either one.

Which B2B sites appear in ChatGPT sources at all

The general studies are dominated by consumer queries. B2B sites appear in them, but mostly at the edges. Here is what the five tables show for domains a software buyer would recognise.

The general studies: LinkedIn, GitHub, tech press and review sites

LinkedIn is the strongest B2B domain in the general data. It holds 2.42% of Similarweb's ChatGPT events and 1.1% of the Ahrefs top 50, and Semrush puts it in the top five across all three platforms it tracked. GitHub appears once, at 1.62% in Similarweb. Tech publishers TechRadar and Tom's Guide sit in the Ahrefs top 50 at 2.1% each, though those are largely consumer gadget queries.

Similarweb also found openai.com in third place at 6.21% of ChatGPT's citation events. ChatGPT cites its own maker's site more than it cites YouTube or LinkedIn.

G2 is the interesting case. It held 1.1% of all ChatGPT citations in Profound's 2024 to 2025 data, level with Forbes. It does not appear in Similarweb's 2026 top 20 or the Ahrefs September 2026 top 50. Trustpilot does, at 1.5% in Ahrefs. A fuller treatment of review sites across six studies sits in our breakdown of G2's citation share in AI answers, so we will not repeat it here.

The B2B study: vendors cite vendors

DerivateX is the only one of the five built around software buying prompts. Of the 188 cited sources it classified by hand, 96, or 51.1%, were vendor-owned content. Independent and niche sites supplied 22.9%, recognised media 14.9%, community sources such as Reddit and LinkedIn 9.6%, and review aggregators 1.1%.

Read the definition carefully. "Vendor-owned" means any software company's site, which includes one vendor's blog listing ten competitors. The study's companion report found that ChatGPT cites a recommended tool's own website only 11.6% of the time. So vendor content is the dominant source type, but it is mostly other vendors' comparison pages, not your own product page.

What this means for a B2B content plan

Our position is that B2B teams have been reading the wrong table. The general studies tell you Reddit and Wikipedia matter, which is true for consumer questions and much less true for "best expense management software". The one B2B-specific study says comparison content on software sites is the main source, and review profiles barely appear.

If we ran a B2B marketing team, we would fund honest comparison pages on our own domain before buying another review-site package. The format that seems to win, with a dated title, a comparison table and plain verdicts, is laid out in our template for B2B comparison pages. Community presence comes second, and our Reddit citation playbook covers where legitimate participation ends.

Brand mentions across the web still count, even if they are not cited links. The correlation between mentions and AI visibility is covered in our post on brand mentions versus backlinks.

Where this comparison is weakest

This post asks every study to state its weak points, so here are ours.

First, every source here is commercial. Four studies come from companies selling AI visibility software, and one from an agency selling optimisation services. Each has a reason to find that citations are measurable, changeable and worth paying to track. We found no peer-reviewed domain-share study for ChatGPT to set against them.

Second, the samples cannot be audited. No study publishes its prompt list. Ahrefs publishes no sample size. Similarweb's stated total of nearly 600,000 events does not match the roughly 900,000 its own ChatGPT table implies. Semrush gives exact values only in chart images. DerivateX ran one prompt per category, which makes its B2B shares sensitive to wording.

Third, the windows do not overlap. The studies span August 2024 to September 2026, and Semrush shows ChatGPT's sourcing can shift sharply within weeks. Some of what looks like disagreement between studies may simply be ChatGPT changing between them.

Finally, the strongest counter-case deserves a hearing. If citation shares are this unstable, perhaps no domain-share table is worth acting on, and you should measure only your own prompts. We think that view is mostly right. The rank order is the one finding stable enough to plan around, and our guide to building your own AI visibility dashboard covers the measurement you would need instead.

Frequently asked questions

Where to start

Pull the cited domains for your 10 highest-intent buying questions this week, and sort them into the six DerivateX buckets. If vendor comparison pages dominate your list, as they did in that study, your next content budget should go to a comparison page on your own domain. Your review-site package can wait one quarter while you see whether that page gets cited.

When someone quotes you a ChatGPT citation percentage, ask what it was divided by. If they cannot answer, use the rank order and ignore the number.

Related reading