An analysis of 380 software categories revealed that AI models rely heavily on low-rank websites and vendor marketing blogs to generate product recommendations. The study, which examined 7,534 citations across 2,055 distinct domains, found that 59.8% of these citations point to domains ranked worse than #100,000 in the Tranco top-1M list, while 23.4% point to domains that are not in the top million at all.

The investigation, which used two Perplexity models, also identified a vendor's marketing blog, guideflow.com, as the third-largest source of citations, with 194 citations across 96 categories. Guideflow, which sells interactive product demos, is not a review site, directory, or publisher, and does not compete in any of the categories analyzed.

Furthermore, the study discovered three sites - wifitalents.com, worldmetrics.org, and gitnux.org - that appear to be under common control, with each site publishing thousands of machine-generated "best " pages. These sites, which were registered between December 2023 and May 2024, share the same DNS nameservers, page template, and navigation, and each has a blog with exactly six posts, all of which are about the other brands in the set.

The scale of these sites is notable, with each site listing over 100,000 URLs, mostly in the form of /best/-software/ pages. The self-description of these sites, which use titles like "Facts & Grounding Page" and meta descriptions that mention "machine-readable records," suggests that they are optimized for AI retrieval rather than human readers.

The study also found that the recommendations generated by the AI models often point to vendor homepages that are no longer reachable or have been redirected to different domains. In some cases, the models recommended products with domain names that resolve to unrelated websites, such as an Indonesian online-gambling portal.

Finally, the study highlights the importance of understanding the evidence base used by AI models to generate recommendations. While the study did not test whether removing these sources would produce different recommendations, it suggests that the use of low-rank websites and vendor marketing blogs may have significant implications for the quality and reliability of AI-driven product recommendations.

The full dataset and methodology used in the study are available online, along with a PDF version of the report. The study's findings have implications for the development of AI models and the optimization of websites for AI retrieval, highlighting the need for greater transparency and accountability in the use of AI-generated recommendations.

Este artículo fue escrito con la asistencia de IA.
News Factory APP - noticias agénticas para impulsar tu SEO y AEO.