The global lists are real. They just mix every topic together
If you search for where to get mentioned so ChatGPT recommends you, the answer is a ranking of the most cited domains. Ahrefs' September 2026 list for ChatGPT puts Reddit first at 16.8% and Wikipedia second at 7.0%, across broad US queries on all topics. Similarweb's count for January and February 2026 has Wikipedia at 13.15% and Reddit at 11.97% of ChatGPT's citation events.
Two things get lost when those numbers are passed around. Ahrefs' 16.8% is a share of the citations going to its top 50 sources, not of everything ChatGPT cited. And both lists pool every kind of question: dictionary lookups, symptoms, car reviews, recipes. A buyer asking for the best financial close software is asking something else. So we looked at what gets cited for that kind of question, one category at a time.
We counted 12,889 citations across 15 categories
Between 4 and 6 October 2026 our sampler put unbranded buyer questions to four engines: our ChatGPT engine on OpenAI's API, Google AI Overviews, Perplexity and Claude. There were 15 brand setups, one per category, running from steel bars and plywood to cardiac surgery, hostels and finance software. We kept the first answer per question per engine. That left 1,227 answers carrying 12,889 citations to 4,190 distinct domains. Every figure here is ours and belongs to that sample.
Seven domains were cited in at least 12 of the 15 categories: LinkedIn, Facebook, YouTube, Wikipedia, Reddit, Instagram and Quora. That's the universal list, and it overlaps heavily with the global rankings.
- Together the seven took 679 of the 12,889 citations, 5.3%. 95% interval 4.9% to 5.7%.
- 888 of the 1,227 answers, 72.4%, cited none of them.
- The most cited single domain was LinkedIn, at 1.1%. Reddit was 0.6%.
94% of the cited domains showed up in one category only
The rest of the pool was local. 3,940 of the 4,190 domains were cited in exactly one category, and those domains took 9,726 citations, 75.5% of the total. 1,212 of the 1,227 answers cited at least one of them.
Inside each category the seven universal sites took between 2.6% and 8.4% of citations. In 8 of the 15 categories not one of them made that category's top ten.
In finance software the top ten were vendors, G2 and two roundup sites
Here's what a local list looks like. For the financial close software setup, 100 answers carried 1,081 citations. The ten most cited domains took 24.5% of them:
- Six were software companies' own websites.
- One was G2, and the cited pages were its category rankings, head to head comparisons and alternatives pages.
- One was a large vendor's product documentation.
- Two were sites that publish "best X software" roundups, one page per sub-category.
- None of the seven universal sites.
Google AI Overviews is where the universal list does matter
One engine behaved differently. The seven universal sites took 19.5% of Google AI Overviews' citations (interval 17.9% to 21.2%), mostly YouTube, Reddit and LinkedIn. For Claude it was 3.2%, Perplexity 1.9% and ChatGPT 0.7%. ChatGPT cited Reddit 8 times in 1,449 citations.
So the advice to work on YouTube or community threads isn't wrong. It's engine specific. If your buyers search Google, those sites are part of your list. If they ask ChatGPT for a shortlist, our data says they mostly aren't.
What we'd actually do
Build your category's list before you pitch anyone.
- Write 20 or so unbranded questions a buyer would ask, such as "best financial close software for a mid-size company". Leave your name out.
- Run them on the engines your buyers use and collect every cited URL. Rank the domains.
- Open the cited pages, not just the domains. A G2 alternatives page and a G2 review page need different work.
- Sort by what you can influence. Rivals' websites are out. Roundup publishers, review sites, trade directories and trade press are in.
- Start with domains cited by more than one engine and for more than one question.
Where this could be wrong
Three days is short, and each category had one setup with 15 to 29 questions. A different set of questions in the same category could surface a different list.
Several of our setups are Indian businesses, which is why IndiaMART, TradeIndia and Justdial rank high. A US sample would have a different set of local sites, though we'd expect the same shape.
Our ChatGPT and Claude answers come through the providers' APIs. Similarweb measured ChatGPT's web browsing mode, and if the consumer app leans on Reddit and Wikipedia more than the API does, that would explain part of the gap. We haven't tested it.
The intervals treat each citation as independent. Citations cluster inside answers, so the real uncertainty is wider than shown.
We counted domains after stripping www, so subdomains of one company count separately.
