Eight AI search tools cited a made-up research topic within a week

Any brand that wants to appear in an AI answer today is trying to get itself written about on sites those tools already read. A new Fudan University paper tested how little work that actually takes, by publishing about a fabricated research topic on nine of those sites and measuring who cited it. Eight of ten AI search tools cited the invention within seven days, Search Engine Journal reports.

The paper, "From Public Posts to AI-Search Citations: Measuring the Fragility of AI Search", is on arXiv and credited to Qi Liu, Geng Hong, Xinyang Zhang, Pei Chen, Yutong Li and Min Yang. It maps 17,211 citation instances across ten tools, finding that the top twenty domains for each tool capture 20.5 to 70.8 percent of all its citations. The controlled publication experiment then tests whether a new post on one of those preferred domains gets picked up.

How the fabricated topic test ran

The authors ran the experiment on three fabricated topics across nine publication platforms including Medium, GitHub, LinkedIn, Facebook, YouTube, Reddit and Sohu, picking for each AI tool the most-cited site that accepted posts from new accounts. They queried each tool daily for seven days after posting and recorded citations.

Of ten tools tested, eight cited the fabricated content within a week. Once a tool cited the content for the first time, it kept doing so on later queries. Reported from Search Engine Journal's write-up of the paper's table of citation rates after the first citation: Grok cited the invention in 30 of 32 answers, Google AI Mode in 15 of 33 answers, Perplexity in 13 of 35 answers, and ChatGPT in 7 of 20 answers. DeepSeek did not cite the content at all.

What it costs to try

A second experiment tested paid distribution. The authors spent $14 on a content-posting service, which produced posts on five platforms. Doubao cited the first one within an hour. The paper also tested a deletion step on day 11 of a 15-day run and found that most tools kept citing the content for days after the source URL stopped resolving.

The authors frame the finding as a security problem for the AI search layer, since the same path works for anyone with a cheap posting budget and a target domain already in the tool's preferred set. A team working on getting its brand into AI answers reads the same paper as a method sheet: pick the domain the tool already trusts, publish on it, wait a week.

Source

Search Engine Journal, Matt G. Southern. Primary source: Liu et al., arXiv:2610.11932.