Perplexity’s web-grounded answers for software buyers may be leaning on a source base that is easy to manufacture at scale.
A new analysis of 760 Perplexity API calls—covering 380 software categories and two Sonar model tiers—found that 59.8% of the 7,534 returned citations pointed to domains outside the top 100,000 sites in the Tranco ranking. Nearly a quarter pointed to domains outside Tranco’s top million altogether.
The study does not establish that the resulting product recommendations are wrong. But it does show that the retrieval layer behind an AI answer can elevate vast amounts of SEO-style content alongside established review platforms, vendors and publishers.
A small cluster, a very large content footprint
The most striking finding involved three sites: WifiTalents, WorldMetrics and Gitnux. Their sitemaps collectively listed 215,128 generated pages formatted as “best `<category>` software” guides.
The domains appear to share infrastructure and an identical site template, according to the report, though common ownership was not confirmed. All were created after December 2023. WorldMetrics and Gitnux used the HTML title “Facts & Grounding Page,” language that appears tailored to AI retrieval systems rather than conventional visitors.
Together, the three sites accounted for 181 citations across 41 of the 380 tested categories. Their pages also produced inconsistent rankings for the same category. In one example, “project estimation software,” the three sites named different winners and substantially different top-five lists despite using nearly identical layouts and editorial-process claims.
That inconsistency is not proof of manipulation. It is, however, a practical warning: content volume and machine-readable formatting can create visibility in AI answer pipelines without providing a clear signal of editorial independence or research quality.
The retrieval mix matters more than the model label
Perplexity’s two tested tiers, Sonar and Sonar Pro, were highly similar in their sourcing. They returned byte-identical citation lists for 289 of 380 categories, while their overall citation URL sets had a 0.898 Jaccard overlap. They selected the same top-ranked product in 290 categories.
For operators, that means choosing a more expensive model tier may not materially diversify the research substrate behind a software recommendation. The tiers should be treated as variants of a shared retrieval stack, not as independent checks.
The third-most-cited domain in the test was Guideflow, a vendor of interactive product demos. Its blog was cited 194 times across 96 categories, ahead of Gartner. Guideflow is not a review publisher and does not operate in most of the categories it was cited for; its large content-marketing archive appears to have become an important source of evidence nonetheless.
What buyers and software companies should do
AI-generated shortlists should be treated as a discovery input, not a procurement decision. Teams using them should verify product claims against official documentation, current pricing, security materials, reference customers and hands-on evaluations. This is especially important in specialized categories where a plausible-looking listicle can be difficult to audit quickly.
Software vendors should take a related lesson. Being well represented in AI answers is increasingly a retrieval and information-architecture problem: accurate product pages, structured documentation, current comparison material and reputable third-party coverage all matter. But publishing generic pages at extreme scale may also become a competitive pressure—and a reason for AI platforms to improve provenance and source-quality controls.
What to watch next
The report measured Perplexity only, on one day and with one prompt formulation. It did not test ChatGPT, Gemini, Copilot or Google, nor did it test whether removing the cited sites changes recommendations.
Still, its core finding is consequential: AI answer quality depends not only on model reasoning, but on which documents are retrieved as evidence. As more buyers begin their software search in conversational interfaces, transparency about that evidence base will become a product requirement, not an implementation detail.
