AI 搜尋與安全
Perplexity Citation Audit Finds Three Affiliated Sites Generated More Than 210,000 Software Ranking Pages
Researchers collected 7,534 citations across 760 Sonar queries, finding that long-tail websites frequently appeared in the retrieval evidence used by recommendation systems. Three sites suspected of sharing the same operator accounted for just 2.4% of all citations, and the study did not establish that these pages actually changed the recommendations.
A [citation audit](https://trellner.com/reports/manufactured-sources-behind-ai-recommendations/) published on September 2 tested Perplexity's `sonar` and `sonar-pro`. The researchers predefined 380 purchase-intent software categories and queried each model once, asking it to return the top five products and their official URLs in JSON. All 760 calls produced parseable responses, yielding 7,534 citations across 2,055 domains.
When compared with Tranco traffic rankings, 59.8% of citations came from domains ranked outside the top 100,000, while 23.4% were not even among the top one million. This does not necessarily mean those sources were low quality, but it shows that the presence of citations does not mean the evidence has passed an authority filter. A more specific anomaly involved `wifitalents.com`, `worldmetrics.org`, and `gitnux.org`: together, the three sites listed 215,128 automatically generated `best <category> software` pages. Their templates, navigation, cross-references, and DNS infrastructure were highly similar, and two of the sites even described themselves on their homepages as a “Facts & Grounding Page.”
Together, the three sites supplied only 181 citations, or 2.4% of the total, so it would be inaccurate to say they controlled Perplexity's answers. The two models also did not constitute independent samples: in 289 of the 380 categories, they returned identical citation lists, while their URL sets had a Jaccard overlap of 0.898, suggesting that they may share a retrieval layer. The study did not perform counterfactual tests after removing the sources, and it used only a single prompt and a single-day snapshot.
The technical risk is that retrieval-augmented generation systems may mistake large volumes of structured pages designed for machine consumption for multiple independent pieces of evidence. Perplexity already provides [source labels](https://www.perplexity.ai/help-center/en/articles/20260806-understanding-source-labels) such as Government, Academic, and Trusted, but those labels do not replace item-by-item verification. Search and RAG teams should add cross-domain ownership clustering, near-duplicate template detection, source-diversity constraints, and verification that cited content genuinely supports the associated sentence. An earlier [cross-platform study](https://arxiv.org/abs/2605.23684) also found that all four generative search engines it examined cited suspected AI-generated content, indicating that the problem is not limited to a single product.