8 of 10 AI Search Tools Cited a Made-Up Concept Within a Week of Seeding, Fudan University Study Finds

Ras Al Khaimah, United Arab Emirates, 21 October 2026.

Eight of ten AI search platforms cited a fabricated concept within seven days of researchers posting about it on sites those platforms already favour, according to a paper by six researchers at Fudan University. The study, From Public Posts to AI-Search Citations: Measuring the Fragility of AI Search, was posted to arXiv on 8 October 2026 and carries an NDSS Symposium 2027 identifier. It tested ChatGPT, Perplexity, Grok and Google AI on the English web, and six tools including Doubao and DeepSeek on the Chinese web.

Key findings

• Citations are concentrated. Across 17,211 citation instances covering 6,356 source domains, the top 20 domains took between 20.5% (DeepSeek) and 70.8% (Wenxin) of each platform's citations.
• Many cited sources are easy to post on. Of 22 publication platforms tied to cited domains, 15 had low or medium barriers for both account setup and posting.
• A fictional concept spread fast. 8 of 10 platforms cited the made-up "Cactus Gravity Protocol" within seven days, including all four English-web tools.
• Placement mattered more than volume. In paired tests, one article on a frequently cited platform had more citation impact than over 20 matched posts on rarely cited ones. Forty posts on a self-hosted WordPress site received no citation within a week.
• It can be bought. A $14 purchase from a Chinese GEO vendor produced 13 public posts within 28 minutes, and one platform, Doubao, cited content carrying the researchers' planted markers within an hour.

How the four English-web tools behaved

The researchers measured two rates once a platform had first cited their content: the cite rate (how often an answer linked to a seeded post) and the marker appearance rate (how often an answer repeated a planted fake detail, such as an invented name or number).

Platform	Answers (n)	Cite rate	Planted detail repeated
Grok	32	93.7%	53.3%
Google AI	33	45.5%	6.1%
Perplexity	35	37.1%	0.0%
ChatGPT	20	35.0%	10.0%
All 10 platforms	205	62.9%	32.2%

The paper does not say which Google product it labels "Google AI". Across all three test topics and all ten tools, 176 of 1,004 answers (17.5%) cited the seeded posts during the week, and 75 (7.5%) repeated a planted marker.

The fictional topic was by far the easiest to inject. An opinion topic (whether AI would replace junior programmers within five years) reached an 18.8% cite rate, and Perplexity was the only English-web tool to cite those posts, at 11.4%. A fictional keyboard product reached 36.4% overall but was not cited by any of the four English-web tools within the week.

How the study worked

The team first ran 120 queries, 30 each across product recommendations, news, opinion and technical documentation, through the ten tools in March 2026 to map which domains each one cites. In April they tested 22 posting platforms for how easily a new account could publish. They then picked each tool's most-cited site that a new account could use, wrote articles on three topics, seeded each with marker details, and queried every tool daily for a week through its public interface.

The authors stress limits. None of the topics was a real brand or product, so results may not carry over to commercial queries with more competing content. Topic, competition and question type changed together, so their effects cannot be separated. The GEO vendor test is a single case. The paper says the team completed responsible disclosure to all ten platforms, and that ChatGPT, Perplexity and Google AI, among others, acknowledged receipt.

Context

Google has already addressed this behaviour in policy. Its guide to optimizing for generative AI features says that seeking inauthentic mentions "isn't as helpful as it might seem", and the Search documentation updates log records a May 2026 clarification that Google's spam policies apply to generative AI responses. The tests ran in March and April, before or around that change, and the authors note that rates may drift as platforms update their systems.

What this means for SEO teams

• Map where each AI tool gets its sources. The study shows citation preferences differ sharply by platform, so audit the top cited domains for the queries you care about on each tool separately.
• Put real expertise where AI tools already look. Legitimate contributions on frequently cited platforms had far more effect than volume on a low-visibility blog.
• Monitor your brand for planted claims. If fabricated details can reach AI answers in days, a competitor or bad actor can do the same to your brand. Check AI answers for your brand name regularly and keep primary facts on your own site clear and consistent.
• Do not buy seeded posts. Google treats manipulation of AI responses as spam, and the effect in this study was strongest for topics with no competing information, which is rarely true of a real market.

Conclusion

The Fudan study gives measured numbers to a risk many SEOs suspected: AI search tools lean heavily on a small set of open platforms, and a handful of posts there can shape answers within days when nothing else on the web covers the topic. Grok was the most exposed of the English-web tools tested, Perplexity repeated no planted details, and Google AI and ChatGPT sat in between.

The findings come with clear limits, chiefly that no real brand was tested. Still, the result for SEO teams is practical: know which sources each AI tool cites in your market, and make sure accurate information about your brand is present there before someone else's is.

Questions and answers

Which AI search tools did the Fudan study test?

Ten platforms: ChatGPT, Perplexity, Grok and Google AI on the English web, and six Chinese-web tools including Doubao, DeepSeek, Kimi, Qwen, Wenxin and Yuanbao.

How many AI search tools cited the fabricated concept?

Eight of the ten platforms cited the made-up Cactus Gravity Protocol within seven days of the researchers posting about it, including all four English-web tools.

Did posting on a personal WordPress site work?

No. According to the paper, 40 posts on a self-hosted WordPress site received no citations within one week, while a single post on a frequently cited platform had more impact than over 20 matched posts on rarely cited ones.

Does Google allow seeding content to influence AI answers?

Google's Search documentation updates log records a May 2026 clarification that its spam policies apply to generative AI responses in Google Search, and its AI optimization guide says seeking inauthentic mentions isn't as helpful as it might seem.

###