In September we wanted to know what kind of page ChatGPT-style answers lean on when someone asks a buying question. So we put 24 questions through OpenAI's API with web search on and through Perplexity's API. Half were about software, things like help desks and CRMs. The other half were services, accountants and business insurance. Each ran once or twice, and then we opened every cited page, 308 of them, and wrote down what kind of page it was.
Software answers were mostly built from lists. 118 of the 167 software citations were roundups or alternatives pages, about 71%. And 85 of those 118 lists were published by a company that sells in the same category, so help desk vendors writing the best help desk lists that then get quoted back.
Services went the other way. Only 21 of 141 citations were roundups. Explainers got 53 and the firms' own service pages got 45. For CPA for real estate investors in Denver, both engines went to individual firms' real estate pages instead of directories.
The engines didn't behave the same either. Perplexity gave us 230 of the 308 citations and about half of what it cited was a roundup. OpenAI was at 28% roundups and cited 12 comparison pages, one of them Intercom's own Intercom vs Help Scout page.
If you want to try this for your own category, it's not complicated, just slow. Pick 10 or so questions a buyer would really type, run them on two engines a couple of times, open every source and tag it. Pay attention to prices too. One answer in our run took Help Scout's $30 from its Shopify App Store listing and $25 from its pricing page and just gave a range.
Doing it every week by hand gets old, which is why RankControl's AI visibility tracking runs a fixed prompt set across engines and logs who gets named and where the answer came from.
Full write-up on Medium · The data as CSV on GitHub · RankControl