Researchers planted a fake concept online, and 8 of 10 AI search tools cited it within a week

Researchers at Fudan University planted a concept that does not exist and watched eight of ten AI search tools repeat it back as fact within a week, a measurement study posted to arXiv on October 8 finds. The paper, titled "From Public Posts to AI-Search Citations: Measuring the Fragility of AI Search," is the first to put a price and a timeline on how cheaply a made-up idea can work its way into what ChatGPT, Perplexity, Grok and Google's AI search tell a user.

The researchers invented a concept they called the "Cactus Gravity Protocol," along with a fictional keyboard product and an opinion question about whether AI would replace junior programmers within five years. They posted about these topics on nine platforms chosen from a pool of 22 candidates, after testing which sites made it easy to create an account and publish without review. Fifteen of the 22 had low or medium barriers on both counts. Then they ran each topic's questions daily across ten AI search platforms for a week and tracked what came back, as Search Engine Journal reported.

The fabricated protocol was cited by all four English-language tools the team tested. Grok linked to the planted posts in 30 of 32 answers and repeated a planted detail in 53% of them. Google's AI search linked in 15 of 33 answers, Perplexity in 13 of 35, and ChatGPT in 7 of 20. Across every tool and topic combined, 176 of 1,004 answers over the week linked back to content the researchers had written themselves. A separate run with Chinese platforms, including Sohu and a GEO vendor's $14 posting package, produced a citation within an hour on one tool.

What the study actually shows

The paper's broader finding is less about any single fake fact and more about concentration. Across 17,211 citation instances spanning 6,356 source domains, the top 20 domains on each platform accounted for between 20.5% and 70.8% of all citations. One article placed on a platform an AI tool already favored beat more than 20 matched posts spread across lower-preference sites. The posting barrier mattered more than the posting volume: getting onto the right site counted for more than writing a lot of content on the wrong ones.

That matches something practitioners have suspected without a controlled test behind it. Google's own guidance already warns that chasing inauthentic "mentions" "isn't as helpful as it might seem", and the company clarified on May 15 that its spam policies extend to attempts at manipulating generative AI responses in Search. The researchers' tests predate that clarification, so it is not yet clear how much of what they measured still works today.

Who this affects

Anyone running a GEO or AI-visibility program now has a documented reason to ask what their vendor is actually buying them. If a $14 package of posts on easy-access platforms can land a citation inside an hour, the tactic that gets a brand mentioned is not necessarily the tactic that gets it mentioned accurately, or by a source anyone would trust. Marketers who treat an AI citation as validation, the way a backlink or a review once was, are relying on a signal the researchers just showed can be bought cheaply and does not require the underlying claim to be true.

It cuts the other way too. Anyone who assumes AI search answers are a neutral read of what's actually been written about their product should treat that assumption as unproven. A competitor, or an unhappy former customer, has the same $14 option.

What a software marketer should do differently

• Treat an AI citation of your brand the same way you'd treat an anonymous review: worth checking, not worth repeating in a pitch deck as proof of anything.
• Ask any GEO vendor which specific platforms they publish to, and whether those platforms are the "high-preference" sources an AI tool already favors or just high-volume ones.
• Audit what AI tools currently say about your own product for claims that did not come from you, since the same mechanism that plants a fake protocol can plant a fake feature or a fake complaint.
• Do not assume this is settled: the researchers' tests ran before Google's May spam-policy clarification, and before competing AI search tools and platforms adjust in response to studies like this one.

The researchers disclosed their findings to all ten platforms before publishing; ChatGPT, Perplexity and Google's AI search acknowledged receipt. None of the fabricated topics involved real brands or products, so the paper's authors caution that citation rates for an invented concept may not transfer directly to how these tools handle claims about an existing company. The mechanism they documented, however, does not depend on the topic being fictional. It depends on the platform being easy to post to and the tool having already decided to trust it.