If you are building tools that rely on AI search APIs for data retrieval, you need to know that your source of truth is currently a house of cards. A new preprint submitted to arXiv by Qi Liu and colleagues, titled 'From Public Posts to AI-Search Citations: Measuring the Fragility of AI Search,' demonstrates that the citation layer in AI search is surprisingly brittle. Unlike traditional search engines that rank pages based on established authority metrics, AI search systems filter, select, and generate answers in a way that can be manipulated by ordinary users with minimal effort.

The Citation Bottleneck

The core issue identified in the study is the concentration of citations. The researchers analyzed 17,211 citation instances across 6,356 unique source domains on 10 different AI-search platforms. They found that the top 20 domains accounted for between 20.5% and 70.8% of all citations on a given platform. This extreme concentration means that if you can get your content onto one of these high-preference domains, you have a significant chance of influencing the AI's output. Worse, the study found that 15 out of 22 tested publication platforms tied to these cited domains had low or medium barriers for both account setup and posting.

Experimenting with Influence

To test this theory, the team conducted marker-controlled publication experiments. They created fabricated concepts and published them on these low-barrier platforms. The results were alarming: within seven days, 8 out of 10 platforms cited the fabricated concepts. In one specific case, a single article on a high-preference platform generated more citation impact than over 20 matched posts on low-preference platforms. This suggests that the platform choice matters more than the volume of content. For developers and product managers, this means that 'search engine optimization' is evolving into 'AI citation optimization,' where the platform you choose to host your docs or blog determines your visibility in AI answers.

The Commercialization of Manipulation

Perhaps the most practical finding for builders is that this manipulation is already a service. The study highlights a $14 Generative Engine Optimization (GEO) purchase that resulted in 13 public posts. One AI-search platform cited this GEO-posted content, complete with designed markers, within just one hour of publication. This speed and low cost indicate that bad actors or aggressive marketers can currently outpace legitimate content creation in influencing AI search results. If your product relies on AI search to direct traffic or validate information, you are competing with a market where a few dollars and a low-barrier account can skew the results.

Key Takeaways

  • AI search citations are heavily concentrated, with top 20 domains capturing up to 70.8% of citations.
  • Low-barrier publishing platforms are the primary entry point for manipulation, with 15 of 22 tested platforms offering easy access.
  • Fabricated content can appear in AI search results within seven days, and sometimes as quickly as one hour.
  • Commercial GEO services are already exploiting this fragility for as little as $14.

The Bottom Line

We are currently treating AI search as a reliable oracle, but this study proves it is easily gamed by anyone with a free account on a preferred domain. Until these platforms tighten their citation selection logic, assume that any AI-generated answer is potentially influenced by cheap, low-effort content.