A new working paper from researchers at Fudan University puts a number on a fear many marketers already had: AI search tools can be steered by content planted on the sites they already trust. Across 10 tools, eight cited a made-up concept within seven days of it being posted, including ChatGPT, Perplexity, Grok and Google AI.
The paper, posted on arXiv on Oct. 8, describes tests run in March and April. The team first mapped which domains each AI search tool cited across 120 questions, then posted the invented Cactus Gravity Protocol to platforms where new accounts could publish. Before posting, the researchers checked that none of the tools returned informative answers on the topic. Within a week, that changed.
What the tools picked up
All four English-web tools cited the fictional concept. Grok linked to a planted post in 30 of 32 answers, Google AI in 15 of 33, Perplexity in 13 of 35 and ChatGPT in 7 of 20. When the researchers looked for planted details, the numbers fell: 53% of Grok’s responses repeated fabricated specifics, 10% for ChatGPT, 6% for Google AI and none for Perplexity. Across all 10 tools and three topics, 176 of 1,004 answers linked to the planted posts.
The research does not prove that a fake brand or product would get the same treatment. The made-up concept succeeded, while an opinion question and a fictional keyboard mostly did not. The authors also do not isolate whether the topic, competition or question type drove the difference.
Placement beat volume
In paired tests for five lesser-known vacation destinations, a post on a site the tools cited often produced the first AI citation every time. The best result on a rarely cited site required 20 postings, and a self-hosted WordPress site received zero references after 40 postings in seven days. A separate volume experiment saw total citations rise from 24 to 128, but ChatGPT never cited those posts and Google AI did only for the lowest-volume topics.
The paper even documents a $14 Chinese GEO package that delivered 13 posts in about 28 minutes; one Doubao reference appeared within an hour. That is a single confirmed instance, not a repeatable win.
What this means for your AI visibility work
This is less a hack and more a warning. The quick win is clear: the domains tools already cite carry outsized weight. The risk is equally clear: Google’s May 15 spam policy update explicitly covers attempts to manipulate generative AI responses in Search. Its AI guide says chasing inauthentic mentions ‘isn’t as helpful as it might seem.’
A practical audit looks like this:
- Run your category’s key questions through the AI tools you actually care about.
- List the domains that keep appearing in citations and answers.
- Build genuine expertise and content on those surfaces, instead of mass-posting thin material.
- Log AI answers before and after, not just links and rankings.
The study’s exact rates will drift as the tools and policies change. Treat it as evidence that AI search citations reward pre-existing source authority, not volume alone.
Source: Search Engine Journal
