In a paper posted October 8, 2026, a Fudan University team seeded a made-up idea on sites that AI answer engines favor, then watched ChatGPT, Perplexity, Google AI, and Grok all repeat it — a direct measure of how easily AI search can be fed fiction.
2026-10-19 · by Moe Ameen
A new arXiv paper, "From Public Posts to AI-Search Citations: Measuring the Fragility of AI Search" (arXiv:2610.11932, submitted October 8, 2026) by a research team at China's Fudan University, set out to test whether ordinary public posts can steer what AI search engines say. The method was blunt: the researchers invented a fictional concept — reported as the "Cactus Gravity Protocol" — wrote posts about it that carried planted fabricated names and numbers as fingerprints, published them from ordinary new accounts on nine sites the tested tools already cited often, and then asked the same five questions daily for a week to see which tools would repeat the fiction.
Eight of the 10 AI search tools cited the made-up concept within seven days, including all four of the English-web tools tested. The posts were linked in 30 of 32 Grok answers, 15 of 33 "Google AI" answers, 13 of 35 Perplexity answers, and 7 of 20 ChatGPT answers; across all 10 tools and the study's topics, 176 of 1,004 answers linked to the planted posts over the week. The paper analyzes 17,211 real citation instances across the 10 platforms and finds that citations concentrate in a small set of platform-specific domains — the sites each tool leans on most — which is exactly where the planted posts were placed.
Two supporting experiments sharpen the point. Placement dominated: a post on a frequently-referenced site produced the first citation every time, while 40 posts on a self-hosted WordPress site drew zero references in seven days. Volume showed an association (total citations across tools rose from 24 to 128 as the number of posts per topic went from 1 to 20, which the authors stress is a correlation, not a proven cause). And a $14 commercial posting package delivered 13 public posts in about 28 minutes, after which one tool's answer included a planted detail within an hour.
The limitations matter. The tests ran in March and April 2026, so rates may have drifted; the topics were invented concepts, not real brands or products, so the findings may not transfer to brand searches; the paper labels one tool "Google AI" without confirming which Google product was tested, so it is not established that this applies to AI Overviews specifically. The authors say they reported their findings to all 10 platforms, and Google has since clarified that its spam policies cover manipulating generative-AI responses.
The real lesson here is not "go spam the trusted sites" — the paper's own authors reported that method to the platforms, and Google already calls it spam. It is that AI answer engines now assemble answers from ordinary public posts on a short list of domains they trust, which makes two unglamorous things decisive: being consistently present on those platforms, and being accurate there, because whatever an engine ingests it repeats with a citation attached. That is a publishing-and-quality problem, and it is squarely what [Kompozy](/) is built for. (See the related breakdown of [which sources answer engines cite most](/news/ai-platform-citation-source-index-2026) for where that footprint should live.)
Concretely, Kompozy generates a full content set from one idea — a captioned [Persona Shorts](/glossary/persona-shorts) avatar video, a brand-exact [Carousel](/glossary/hyperframes), Quote Graphics, per-platform Text Posts, a Blog Article, and an Email Newsletter — and publishes the set across the eight social platforms plus blog and email through [Autopilot](/glossary/autopilot), so the distributed, on-topic presence these engines actually sample from exists rather than being a single orphaned blog post. And because the study shows how fast an unverified claim propagates, the accuracy gate is as important as the reach: every text output is held to a [Persona Brief](/glossary/persona-brief) with banned-word filters and routed through a per-post human review before it ships, so a stray error never becomes the "fact" an answer engine quotes back about you. You earn AI-search citations the durable way — accurate, on-brand content, published everywhere your audience and the engines already look.
Researchers at Fudan University invented a fictional concept, posted about it on sites that AI search tools already cite often, and asked the tools the same questions daily for a week. Eight of the 10 tools tested — including ChatGPT, Perplexity, Grok, and a Google AI product — cited the made-up concept within seven days, showing that ordinary public posts on trusted platforms can shape AI-generated answers. The paper is arXiv:2610.11932, posted October 8, 2026.
All four English-web tools the study tested repeated it. The planted posts were linked in 30 of 32 Grok answers, 15 of 33 "Google AI" answers, 13 of 35 Perplexity answers, and 7 of 20 ChatGPT answers. Six additional Chinese-web tools, including Doubao and DeepSeek, were also tested, and eight of the 10 cited the concept within a week. The paper does not confirm which specific Google product the "Google AI" label refers to.
It means AI answer engines will repeat what the sites they trust publish, accurate or not, and attach a citation to it. The study used invented concepts rather than real brands, and the tests ran in March and April 2026, so exact rates may have shifted. The practical reading for creators is that content accuracy and where you publish both matter more than before, because errors and fabrications can propagate into AI answers within days.
By publishing accurate, on-brand content consistently on the platforms AI engines already trust — not by planting posts, which Google now treats as spam. That means real multi-platform coverage plus a quality gate so you are never the source of an error. Kompozy, for example, generates a full content set from one idea, publishes it across the eight social platforms plus blog and email, and screens every text output against a Persona Brief with a per-post human review before it ships.