The research we build on
Other people's studies of how AI assistants choose their sources. We lean on these where our own measurements stop, and every figure says whose it is. What we have measured ourselves is on what we know.
There is no single answer
One study looked at 11,647 domains that five AI assistants used as sources, 127,198 times in all. Only 2.7% of those domains were used by all five. Seven in ten, 69.6%, were used by exactly one. Between ChatGPT and Perplexity alone, a separate run of 100,000 prompts through both found 11.0% of domains shared, with 37.4% appearing only in ChatGPT and 51.6% only in Perplexity.
SurfacedBy, 127,198 citations, 11,647 domains, five AI assistants, 29 March to 27 June 2026. Profound, 100,000 distinct prompts run through both assistants.
So the question has five answers, and any article giving one is describing a single assistant without saying which.
What each assistant reads
Every figure below states what it is a share of. A share of an assistant’s ten most-used sources and a share of all the times it uses a source are different numbers, and they get quoted as if they were the same one.
- ChatGPT
Reddit and Wikipedia, and neither reliably.
Wikipedia accounts for 7.8% of all the times ChatGPT used a source and Reddit 1.8%, though Reddit is 11.3% of its ten most-used sources. That mix is unstable. Wikipedia appeared in roughly 55% of ChatGPT responses until mid-September 2025, then fell below 20%, and Reddit fell from about 60% to about 10% in the same fortnight.
- Perplexity
Reddit and YouTube.
Reddit accounts for 6.6% of all the times Perplexity used a source and 46.7% of its ten most-used sources, the heaviest concentration measured on any assistant. YouTube is 13.9% of that top ten, and Perplexity alone accounts for 38.7% of all the times YouTube was used as a source across the assistants studied.
- Claude
Documentation, vendor pages and prestige editorial. Almost no user-generated content.
Company and product domains account for 64.0% of all the times Claude used a source and news and media 14.9%, more than four times any other category. Of 379,321 times Claude used a source, spanning 16,406 domains, not one was Reddit. A separate sample across five assistants put its YouTube share at 0.0%.
- Google AI Overviews
Community platforms, weighted toward question sites.
Within its ten most-used sources: Reddit 21.0%, YouTube 18.8%, Quora 14.3%, LinkedIn 13.0%. Across all the times it used a source, Reddit is 2.2%, the same gap between top-ten share and total share that Perplexity shows.
3.7 sources per answer, the fewest of any assistant.
8.6 sources per answer.
6.8 sources per answer.
7.8 sources per answer.
Shares of all sources used and top-ten shares from Profound, 680 million citations across three AI assistants, August 2024 to June 2025. Sources per answer and the 0.0% YouTube figure from SurfacedBy as above. The Wikipedia and Reddit collapse from Semrush, 230,000+ prompts and over 100 million citations, 14 July to 12 October 2025. Claude’s category mix from Otterly, 379,321 citations across 16,406 domains, June 2026, on SaaS and technology queries. YouTube by platform from Otterly, over 100 million citations across six AI assistants in 30 days.
If your buyers use Claude, a Reddit and YouTube strategy is the wrong bet. Earned coverage in high-trust editorial and your own documentation is the right one. If they use Perplexity, the reverse.
The Reddit number, and its caveat
A study of 129,000 domains and 216,524 pages across 20 niches found that domains with heavy Reddit brand mention were used as a source by ChatGPT 7 times on average, against 1.8 for domains with minimal Reddit presence. A difference of 3.9 times.
SE Ranking, 129,000 domains, 216,524 pages, 20 niches, November 2025.
Now the part most articles leave out. An analysis of 200 million prompts over five months found that even the most-used domain on any platform rarely exceeds 5% of all the times that platform used a source, and that Wikipedia, Reddit, LinkedIn and YouTube combined rarely top 5%. The other 95% is spread across thousands of domains.
Evertune, 200 million prompts, five months.
Both are true, because they count different things. Reddit is 46.7% of Perplexity’s ten most-used sources and 6.6% of all the times Perplexity used a source. Same platform, same study, two numbers seven times apart. One says Reddit dominates the short list of usual suspects. The other says the usual suspects are a small slice of what gets quoted.
Almost nobody states which of the two they are quoting. When you read that a platform “drives 46% of citations”, find the denominator before you budget against it.
What this means for a local Australian business
Be careful with all of it. Only some of the studies above state what they asked about: Otterly’s Claude work says SaaS and software on its face, SE Ranking covers 20 niches, and Yext’s is local businesses. The rest report totals without a subject, and the domains that recur in them are Gartner, Forbes, TechRadar and Business Insider, which is not the neighbourhood your customers are asking about.
Ask an AI assistant for mobile bar hire in Melbourne and we expect it to reach for Maps data, local directories and review platforms more than Reddit threads or YouTube explainers. We have not yet measured a local business with web search switched on, so that is an expectation, not a finding.
The nearest published evidence is a study of local questions across more than 200,000 business locations. The businesses’ own websites were 44% of the sources used, listings such as Google Business Profile 42%, reviews and social 8%, and forums such as Reddit 2%.
Yext, 6.8 million citations from about 1.6 million questions per assistant, three AI assistants, 1 July to 31 August 2025, locations of Yext’s own clients and prospects, published 9 October 2025. Method.
Where a client’s questions include both local and other questions, our report shows the two rates side by side.
Which leaves one reliable method
If only 2.7% of the domains used as sources are shared across the assistants, no general answer can tell you what gets a specific business used as a source. The only thing that can is asking that business’s own buyer questions, on the assistants its buyers use, and reading which domains the answers used.
That is what our monthly AI checks read for a client: which pages the answers use, and whether the client’s facts are right on them.
Every study on this page was published by a company selling monitoring or SEO tooling. Our own work touches AI answers too, so read this page with the same caution. The method behind each figure is linked above so you can check it, and how pages become sources and why they stop covers the mechanism underneath. Our own standard of proof is a separate page, including the results that came back flat.
Corrections
The correction below said our work moved to bookings. It has moved again: we now set up and optimise affiliate programs and discount code visibility. This page stays as the research it always was.
The page said the source map is a section of our client report. We stopped selling that report on 15 September 2026, when our work moved to bookings. The same reading now happens in our monthly AI checks.
The page said nearly every study above ran on B2B and SaaS questions. Only some of them state a subject: Otterly’s Claude work says SaaS and software, SE Ranking covers 20 niches, and Yext’s is local businesses. The sentence now says that instead.
The page presented as observed that AI assistants answer local questions from Maps data, local directories and review platforms. We have not measured that, so it is now an expectation, beside a Yext study in which businesses’ own websites were the largest source category.
The page said every scoreboard reports the split between local and other questions. Our report shows it only where a client’s questions include both.
The page said the source map is in every report and is the section clients act on first. We hold no evidence for either, so it now says only that it is a section of our client report.
The page said Perplexity reads whatever was published this week and that freshness counts for more there than anywhere else.Ahrefs (16.975 million URLs, 28 July 2025) found ChatGPT favours fresh pages most, so both phrases are removed.