ChatGPT, Perplexity, Grok, and Google AI each cited posts about a made-up concept within seven days of the posts going up, according to a paper from researchers at Fudan University.
The researchers published the posts from ordinary accounts on nine sites the tested tools already cited often. On two other topics, an opinion question with existing discussion and a made-up product, the same four tools picked up little in the same seven days.
The paper uses the label “Google AI” and links to Google’s general AI site, but it doesn’t specify which Google product was tested. The tests were conducted in March and April, and the paper was posted on arXiv on Oct. 8.
How The Test Worked #
In March, the Fudan team ran 120 questions through 10 AI search tools and recorded the sites each tool cited. The researchers classify ChatGPT, Perplexity, Grok, and Google AI as English-web tools; the other six, including Doubao and DeepSeek, are Chinese-web tools.
Across those questions, Reddit was the most-cited site for Google AI and Grok, YouTube for Perplexity, and Wikipedia for ChatGPT. Six to 10 of each tool’s 10 most-cited sites accepted posts from ordinary users.
In April tests of 22 posting platforms, the researchers rated Medium, GitHub, LinkedIn, and Facebook easy to join and post on, while Reddit’s karma requirements and strict deletions made posting harder.
For each tool, the team then picked its most-cited site that a new account could post on, which gave them nine sites, including Medium, YouTube, and Sohu. The research team posted articles about the fictional “Cactus Gravity Protocol,” whether AI would supplant junior programmers inside of five years, and a fictional keyboard. The first drafts of each article were created using ChatGPT, then each was manually revised and carried planted details (e.g., fabricated names and numbers) to allow the authors to identify if the AI tools repeated that information.
The researchers ran checks before posting to verify that none of the tools provided an informative response to either of the fictional topics or repeated the planted information. Following posting, the researchers ran the same five questions for each of the three topics on all tools daily for one week.
In June I covered a Cornell Tech paper on planted text that tested open-source research agents in a simulation. This team posted on the live web and queried the tools through their public interfaces.
What The Four English-Web Tools Cited #
Eight of the 10 tools cited the made-up concept, the Cactus Gravity Protocol, within a week, including all four English-web tools. Counting answers from the first day each tool cited the posts through the seventh day after posting, Grok linked to a post in 30 of 32 answers, Google AI in 15 of 33, Perplexity in 13 of 35, and ChatGPT in 7 of 20.
Fewer responses to questions included a planted detail from the source posts. Of Grok’s responses (n = 32), 53% did; for ChatGPT (n = 20), 10%; for Google AI (n = 33), 6%, and Perplexity (n = 35) none. In total, among all 10 tools and all three topics, there were links from 176 of the 1,004 answers to the posts during the week.
The only English-web tool that cited an “opinion” post was Perplexity, in 11.4% of its responses (n = 35), and none of the four cited keyboard posts within the week. The authors do not isolate which of the topic, competition, and type of question caused the differences across the three topics, since all three changed together.
In a follow-up, the team put one post each on Sohu, YouTube, and CSDN for 10 more made-up topics. Nine drew at least one citation or planted detail across the 10 tools, and Perplexity was one of the three tools that cited the most of them. The paper doesn’t break out the other English-web tools there.
Placement Beat Volume In These Tests #
To assess location, the researchers created paired postings about 5 relatively unknown vacation destinations. One post was located on a source frequently referenced by the tools; the second post was placed on a site that was very rarely, if ever, referenced. In all cases, the post on the frequently referenced source produced the first reference/citation or planted detail.
The best case on the lesser-referenced side required 20 postings and 40 postings on a self-hosted WordPress site received zero references within seven days. These are presented by the authors as results for specific topics, not as general rules.
A separate experiment tested six fictional subjects using varying numbers of Sohu postings (between 1-20). The total number of citations across all ten tools increased from 24 to 128. However, ChatGPT never generated references to those posts and Google AI only did so when citing the three lowest-volume subject topics. Since each of the volumes utilized a unique subject, the researchers referred to this increase as an association rather than an effect of increasing the number of posts.
Lastly, the research team purchased a Chinese web posting package via a GEO vendor for $14. All 13 postings were delivered to the web within approximately 28 minutes after purchase. A Doubao reference to one of the postings including a planted detail occurred within one hour of delivery and is presented by the authors as one confirmed instance of success, not a success rate.
Why This Matters #
The topic every English-web tool cited within a week existed only in the researchers’ posts. Posts on the other two topics, an opinion question and a made-up product, mostly didn’t get through to the same tools.
None of the topics the researchers posted about was a real brand or product, so the tests don’t show whether the same pattern holds for brand or product searches, including ones with little coverage elsewhere.
Google’s guide to its AI features says seeking inauthentic “mentions” “isn’t as helpful as it might seem,” pointing to core ranking systems that focus on high-quality content and other systems that block spam. The paper doesn’t say which Google product it checked.
Looking Ahead #
The researchers’ tests ran in March and April, before Google clarified on May 15 that its spam policies cover attempts to manipulate generative AI responses in Google Search. The authors say exact rates may drift as the tools change.
The researchers say they disclosed what they found to all 10 platforms, and that ChatGPT, Perplexity, and Google AI were among those that acknowledged receipt. I’ll update this story if a revised version of the paper changes the findings.
Featured Image: miss.cabul/Shutterstock