Where AI Answers Get Their Sources: 12,408 of Ours Beside Six Published Studies
Semrush, Profound, Ahrefs, Peec, Yext and Muck Rack each measured what AI engines cite. We set their Reddit figures beside ours and checked what each one counts.
"Reddit is the most-cited domain in AI search." Peec puts it in those words, summarizing 30 million cited sources. Profound, working from 680 million citations, gives brands this advice: "Active participation in Reddit communities is crucial for visibility." Semrush tracked more than 230,000 prompts and found ChatGPT citing Reddit in close to 60% of its responses.
We recorded 12,408 answers from ChatGPT and Perplexity in September and October 2026. ChatGPT cited Reddit in 3.9% of them.
That is not a rebuttal. Semrush's own sentence doesn't stop at 60%:
"ChatGPT cited Reddit in close to 60% of prompt responses in early August before collapsing to around 10% by mid-September."
Profound's report gives Reddit two shares of Perplexity's citations, from the same data: 6.6% and 46.7%. One study, one engine, two figures seven times apart. Both are correct.
What each figure counts
| Study | What it counts | |
|---|---|---|
| Semrush, Nov 2025: 230,000+ prompts, July to October 2025 | Share of responses citing the domain at least once | ChatGPT: close to 60% in early August 2025, around 10% by mid-September |
| Profound, Jun 2025 (updated Aug 2025): 680 million citations | Share of all citations | ChatGPT 1.8%, AI Overviews 2.2%, Perplexity 6.6% |
| Profound, the same data | Share of the citations going to each engine's 10 most-cited domains | ChatGPT 11.3%, AI Overviews 21.0%, Perplexity 46.7% |
| Ahrefs, Jun 2025: Brand Radar, June 2025 | Mention share, top 10 domains per engine | AI Overviews 7.4%; outside the top 10 on ChatGPT and Perplexity |
| Peec, Mar 2026: 30 million cited sources, US | Rank across five engines | First; no share in the summary we read |
| Yext, Oct 2025: 6.8 million citations, local searches | Share of citations | 2% ("general forums like Reddit") |
| Ours, Oct 2026: 12,408 answers | Share of answers citing Reddit at least once | ChatGPT 3.9%, Perplexity 2.0% |
| Ours, the same answers | Share of answer-site pairs (each site counted once per answer) | ChatGPT 1.0%, Perplexity 0.3% |
Share of citations counts every link. If an answer cites five sources and one of them is Reddit, Reddit gets one fifth of that answer. Profound's 6.6% and Yext's 2% are this unit. Ours is a near relative: we count each site once per answer, so an answer that links Reddit three times counts once.
Share of answers counts every answer once. One Reddit link earns Reddit the whole answer, however many other sources sit beside it. Semrush's 60% is this unit, and so is our 3.9%. Computed from the same ChatGPT answers, the two units put Reddit at 1.0% and at 3.9%.
Share of the top 10 counts only the links that went to an engine's ten most-cited domains. Profound says so plainly: "These percentages do not represent overall citation volume." Its own figures show how much that matters. If Reddit's 6.6% of everything Perplexity cited is 46.7% of the top 10, then the top 10 together hold about 14% of Perplexity's citations (our arithmetic: 6.6 divided by 0.467). Reddit leads a group that covers one citation in seven.
A rank says who came first. It doesn't say by how much.
The month moves these figures. Semrush recorded ChatGPT's Reddit figure falling from close to 60% to around 10% between early August and mid-September 2025. In June of that year, Ahrefs had Reddit outside ChatGPT's ten most-mentioned domains altogether.
So do the questions. Yext asked local questions, from "coffee shop hours" to "best coffee." Of our 331 prompts, 230 are buyer questions about which firm to hire, in seven industries. In our data the industry a question belongs to moves the source mix further than the engine does: two industries on the same engine would have to trade about a third of their citations to look alike (0.34 to 0.38, measured as total variation distance), while two engines answering for the same industry differ by 0.22. Peec makes the same point about its own data: "industries have their own dominant sources (in hotels, Booking.com and Airbnb often outrank Reddit)."
What our 12,408 answers cite
Here is what the answers did cite, by type of site.
| Type of site | ChatGPT | Perplexity |
|---|---|---|
| Company site | 49.5% | 49.7% |
| Data and research | 5.7% | 3.6% |
| Medical and academic | 4.1% | 2.4% |
| Directory | 4.1% | 5.3% |
| Reference (Wikipedia and similar) | 3.6% | 0.2% |
| Information portal | 3.4% | 9.8% |
| Review and affiliate | 1.7% | 3.9% |
| News and trade | 1.5% | 4.3% |
| Podcast | 1.3% | 1.5% |
| 1.0% | 0.3% | |
| 0.2% | 2.1% | |
| Unclassified | 13.1% | 12.0% |
Share of answer-site pairs: each site counted once per answer. Perplexity: the first 10 sources of each answer.
One row is missing: government. When we tested the gap between the two engines on government sites, it did not hold up, and the gap on Reddit failed the same test. Government is on the study page with the rest of the table. Reddit stays here because it is the subject, which is why this page gives each engine's Reddit figure and never sets the two against each other.
Company sites hold about half on both engines, 49.5% and 49.7%. The industry moves that share by up to 15 points; the engine doesn't. Counted by answers instead, company sites appear in 50.4% of ChatGPT answers (95% interval 46.1 to 54.3) and 88.2% of Perplexity answers (85.1 to 90.8). ChatGPT's figure is held down by the 32.7% of its answers that cited no source at all.
Read "company site" carefully. We took the brands we track, and their partners, out of these tables. A company site here is some other company's own website. It does not mean your site gets cited half the time, and we don't publish the share that goes to the brands we track.
Where another study lands close to ours
Yext analyzed 6.8 million citations from more than 1.6 million answers on Gemini, OpenAI and Perplexity, and put 44% of them on websites. Its forum share, "general forums like Reddit," was 2%. Our company-site share is about half, and our Reddit share is 1.0% or below.
It is the nearest match on this page, and it proves less than it looks. Yext's questions were local, and the industries it reports are retail, finance, healthcare and food service. Its bucket is "Full Control (Websites)," sites a business runs itself, which is close to what we mean by a company site. We take out the brands we track; the article doesn't say whether Yext did anything similar. Two studies asking different questions and landing near the same place is mildly reassuring. That's all it is.
Where we disagree and can't say why
Muck Rack has published "What Is AI Reading?" three times: July 2025, December 2025 and May 2026. Across all three, its authors write, "journalism has accounted for 25% to 27% of links cited by ChatGPT, Claude and Gemini." Our nearest type, news and trade sites, is 1.5% on ChatGPT and 4.3% on Perplexity.
Put a different figure of ours beside theirs and it looks like agreement: news and trade sites appear in 24.0% of Perplexity answers. It isn't. Theirs is a share of links and that one of ours is a share of answers, the same trap as the Reddit numbers above.
We can list candidate reasons, and we can't choose between them. They ran Claude and Gemini; we didn't. Ours went through an API, not the apps people type into. We don't know their prompts. We haven't read how they define journalism, because the full report returned an error when we requested it, and a trade publication could fall on either side of their line.
Their authors raise an objection that applies to this page: "comparisons to findings from much smaller samples need to account for differences in methodology and scale." Their latest edition analyzed more than 25 million links. We have 122,018 citation rows, about a two-hundredth of that. The point is fair. A larger sample narrows the error around a figure. It doesn't change what the figure counts or which prompts produced it, and those are where we would look first. The same piece says, "We should keep questioning the numbers and comparing methodologies as more research comes out, including questioning ours." We are taking them at their word.
Who is selling what
Every study on this page comes from a company that sells something its finding points toward. Yext's article recommends its Brand Visibility Platform. Semrush and Ahrefs sell AI visibility tools, and Ahrefs' figures come from one of them, Brand Radar. Profound and Peec sell AI search tracking. Muck Rack sells software to PR teams, and its piece is written by its VP of Data & Intelligence and its VP of Communications. We sell AI visibility work, and this page is part of how we sell it.
None of that makes a figure wrong. It is a reason to read the unit before the headline.
Before you repeat the next citation figure
Ask four things of it.
- What is the unit? Share of all citations, share of the top 10, share of answers, or a rank.
- Whose questions? Local searches, buyer questions, a vendor's customer prompts. And are they published?
- Which surface? The app a person types into, or an API. Ours is the API.
- Which month? Semrush's ChatGPT figure went from close to 60% to around 10% in about six weeks.
If you publish citation data, send us the study and tell us its unit, and we will add it as a row in the first table on this page. If you think our prompts are the reason Reddit barely shows up for us, send the questions you would add before our next fill closes, around 15 November.
Method
Our figures. 12,408 answers, 6,204 from each engine. The 331 prompts were written for 12 sites in seven industries, and each was asked 12 times in every fill it ran in: 517 prompt-months per engine. ChatGPT (gpt-5-mini) and Perplexity (sonar), through an API, not the consumer apps. Fills ran 12 to 26 September and 1 to 5 October 2026: 122,018 citation rows across 6,158 hosts (a host is a site address, such as en.wikipedia.org). Source-type shares count each host once per answer, and only the first 10 sources of each Perplexity answer. The brands we track and their partners are left out of both tables. Host types: 66.2% typed by hand, 21.3% by rule, 12.5% untyped; a hand spot check of 343 rule-typed hosts puts expected mistyping at 2.7% of all pairs (range 1.2% to 6.4%). We don't name the sites or the industries. Full method, tables and intervals: What kinds of sources do ChatGPT and Perplexity cite? on bestseopodcast.com. Next refresh: after the November fill closes, about 15 November 2026.
Outside studies, each read at its source on 27 October 2026:
- Profound (Nick Lafferty), AI Platform Citation Patterns: How ChatGPT, Google AI Overviews, and Perplexity Source Information, June 2025, updated August 2025.
- Semrush, The Most-Cited Domains in AI: A 3-Month Study, 10 November 2025.
- Ahrefs (Patrick Stox), The 10 Most Mentioned Domains for ChatGPT, Perplexity, and AI Overviews Across 78.6M Searches, June 2025.
- Peec (Tomek Rudzki), March 2026, read through Peec's own documentation summary of the study.
- Yext, AI Doesn't Rank, It Cites. And 86% of Its Sources Are Brand-Managed, 9 October 2025.
- Muck Rack, as summarized by its authors Matt Dzugan and Linda Zebian in PR Daily, What does 'earned' mean in the age of AI?, 3 September 2026. The full report returned an error to our request, so we have not read its category definitions.