Which sources ChatGPT cites for software recommendations: a study of 100 answers

When ChatGPT recommends software with web search on, it links to websites. Which ones? We looked at every link in the 100 answers behind our AI rankings: 5 buyer questions in each of 20 SaaS categories, asked to a ChatGPT model (gpt-4.1-mini with web search) on 26 October 2026.

This is one run of one model, so treat it as a sample, not a law. But it's a real sample, and it doesn't look like the "just get on G2" advice you usually hear.

The short answer

• 83 of the 100 answers linked at least one website. The other 17 named products without linking anything.
• Those 83 answers linked 424 different websites, 650 links in total (counting each site once per answer). A typical answer with links linked 6 to 10 sites; the median across all 100 answers was 8.
• Only 204 of the 424 sites were cited inline to back a claim about a product. The rest showed up only as image or link cards, which are often unrelated to the claim next to them.
• One site stood out: techradar.com. It was linked in 32 of the 100 answers, across 13 of the 20 categories, and cited for a claim in 31 of them. toolradar.com came second, in 14 answers across 9 categories.
• After those two, it's a long tail. 303 of the 424 sites appeared in exactly one answer, and 193 of the 204 claim-cited sites appeared in only one category.

Method

For each category we asked the same five kinds of question: "What are the best [tool] for [audience]?", "I'm choosing [tool] for [audience]. Which products should I shortlist and why?", "What [tool] do most small teams use right now?", "Recommend an affordable [tool] with good support." and "Which [tool] is easiest to set up for a small team?".

For each answer we recorded the products named, every linked website (by domain), and which of those links sat next to a specific claim about a product. A link that appears only as an image or link card counts as linked, but not as cited for a claim. We also recorded when a product's only cited source was its own website. The per-category numbers are on each category page in the AI rankings.

The sites cited most often

Sites cited inline for a claim, by number of answers (out of 100) and categories (out of 20):

• techradar.com: 31 answers, 13 categories
• toolradar.com: 14 answers, 9 categories
• tomsguide.com: 5 answers, 3 categories
• schedulingkit.com: 4 answers, 1 category (scheduling software)
• toolvs.co: 3 answers, 3 categories
• softwarecompare.net: 3 answers, 3 categories
• techrepublic.com: 3 answers, 1 category (AI note takers)
• simplifylearn.com: 3 answers, 1 category (team password managers)
• dupple.com: 3 answers, 2 categories
• geekchamp.com: 3 answers, 2 categories

Only 11 sites were cited for a claim in more than one category. Most categories have their own set of sources: in scheduling software it was schedulingkit.com, in AI note takers techrepublic.com, in project management aversusb.net and spotsaas.com. A general "top review sites" list would miss most of them.

What wasn't there

Some sites you might expect were rare or absent in this run:

• G2 was linked in 3 answers (CRM, e-signature and form builders) and cited for a claim in 1.
• Capterra, Reddit, YouTube, Wikipedia, Forbes, PCMag and ZDNet weren't linked in any of the 100 answers.

That doesn't mean those sites don't matter. ChatGPT may still have learned from them, and other engines, other models and other wordings cite differently. It does mean that "which sites does ChatGPT cite for my category?" is a question to check, not assume.

Linked isn't the same as cited

About half the links (295 of 650) were cited for a claim. The rest were image or link cards, and some of the most-linked sites never backed a claim at all:

• freepng.com was linked in 8 answers across 4 categories, always as an image card and never for a claim.
• prnewswire.com (5 answers) and businesswire.com (4 answers) are press release wires. They were never cited for a claim.
• zapier.com and appsumo.com were each linked in 4 answers and never cited for a claim.

If you're building an outreach list from AI answers, count claim citations, not every link. A card for a stock image site won't help you get recommended.

Your own website can be the only source

In 29 of the 100 answers, at least one product's only cited source was its own website.

The question wording mattered a lot. For "Recommend an affordable [tool] with good support", 64 of the 99 products named were backed only by their own site. For the other four questions it was 29 of 396. Narrow, specific questions are where smaller products got in, and their own pages (pricing, support, feature pages) did the work. That's a reason to state your pricing and support plainly as text on your site.

Some answers cite nothing

17 answers named products without linking any website. They were spread across 12 of the 20 categories, with no more than 2 in any one. Those answers come from what the model already knows, so they can be out of date: one email marketing answer with no links still recommended "Sendinblue" and "ConvertKit", the old names of Brevo and Kit.

What to do with this

1. Find your category's sources. Look up your category in our AI rankings, or ask ChatGPT your buyers' questions with search on and write down every domain cited for a claim.
2. Prioritize repeat appearances. A site cited in several answers for your category matters more than one cited once.
3. Ignore card-only links. Focus on sites that back claims.
4. Make your own pages quotable. Plain-text pricing, who it's for and what support you offer, since your own site may be the only source cited for you.
5. Re-check over time. This is one run. Sources change as lists get updated and models change. Our rankings re-run on a weekly schedule.

For the practical steps behind each of these, see How to get your SaaS recommended by ChatGPT. To see which sites ChatGPT cites when buyers ask about *your* category, run a free check.