A concept that never existed became citable material for Grok, ChatGPT, Perplexity, and Google AI after Fudan University researchers wrote about it on sites those tools already trust. All four referenced the planted posts within seven days. The study, covered by Search Engine Journal on Oct. 19, shows how readily AI answers can inherit claims from the open web.

The team first mapped where the tools look. In March, it put 120 questions to 10 AI search tools and logged every site they cited. Reddit topped the list for Google AI and Grok, Wikipedia for ChatGPT, and YouTube for Perplexity. For each tool, the researchers then chose the most-cited site where a brand-new account could publish, which yielded nine sites that included Medium, YouTube, and Sohu.

Three topics were tested: a fictional idea called the Cactus Gravity Protocol, an opinion question about whether AI would replace junior programmers within five years, and an invented keyboard. ChatGPT produced first drafts, which the authors revised by hand and seeded with fabricated names and numbers so any repetition could be traced. Before posting, they confirmed that no tool answered informatively about the fictional topics or echoed the planted details. After posting, they asked the same five questions per topic, every day, for a week.

Eight tools out of 10 referenced the invented protocol, all four English-web tools among them. Counted from each tool’s first citation through day seven, the rates of linking to a planted post were:

Repeating the invented specifics was rarer. Grok reproduced a planted detail in 53% of its 32 answers, ChatGPT in 10% of 20, Google AI in 6% of 33, and Perplexity in none of 35. Across every tool and topic, 176 of 1,004 answers linked to the researchers’ posts.

The opinion and product topics fared poorly. Among the English-web tools, only Perplexity cited the opinion post, in 11.4% of its 35 responses, and none cited the keyboard posts. The authors cannot say whether topic, existing competition, or question type explains the gap, because all three varied at once.

Placement mattered in tests built around five obscure vacation destinations. Each got one post on a frequently cited source and one on a site rarely or never cited. The well-cited location always produced the first citation or planted detail. On the weaker side, the strongest result took 20 postings, while 40 on a WordPress site the authors hosted themselves earned zero references in that window. The authors present these as findings tied to particular topics, not general rules.

Volume was murkier. Across six fictional subjects, Sohu post counts ranged from 1 to 20, and total citations across all tools rose from 24 to 128. ChatGPT never referenced those posts. Because each volume used a different subject, the authors call the rise an association, not an effect.

The team also bought a Chinese web posting package from a GEO (generative engine optimization) vendor for $14. All 13 posts went live in about 28 minutes, and Doubao cited one containing a planted detail within an hour. The authors count that as one confirmed instance, not a success rate.

The limits are substantial. The paper labels one tool “Google AI” and points to a generic Google AI page without naming the product tested. Testing took place in March and April, and arXiv received the paper on Oct. 8. None of the topics were real brands or products, so the findings say nothing about whether brand searches behave the same way, including thinly covered ones. The authors add that the tools will change, so precise rates may shift.

Google’s own guidance on AI features says chasing inauthentic mentions “isn’t as helpful as it might seem” (it points to core ranking systems and spam defenses). The tests predate May 15, when Google clarified that its spam rules reach efforts to game AI-generated responses in its search product. The researchers say they notified all 10 platforms, and Perplexity, ChatGPT, and Google AI numbered among the acknowledgers.

The takeaway is about trust. A citation in an AI answer shows that a tool found a page, not that anyone verified it, so treat each one as a lead to check. Teams monitoring how AI tools describe their brands should note that this paper tested no real brand, leaving that exposure unmeasured.

Search Engine Journal, October 19, 2026, reporting an arXiv paper from Fudan University researchers posted Oct. 8.