How does an answer engine decide which pages to cite? A new study gives an uncomfortable answer: mostly by where the page is published, and much less by whether it is true.
Researchers at Fudan University invented a concept that does not exist, the "Cactus Gravity Protocol." They wrote posts about it, put them on sites that AI search tools already cite often, and asked the tools about it every day for a week. Eight of the ten tools cited the posts. All four English-web tools did: ChatGPT, Perplexity, Grok and Google AI.
How the test worked
The paper, posted to arXiv on October 8 and covered by Search Engine Journal on October 19, ran in two stages in March and April 2026.
First, the team asked all ten tools 120 questions and recorded which sites each one cited. For every tool, six to ten of its ten most-cited sites accept posts from ordinary users. Medium, YouTube, Zhihu, CSDN and Sohu were among them.
Then they published. Three topics went out: the made-up protocol, an opinion piece on whether AI will replace junior programmers within five years, and a fictional keyboard. Each post was drafted with ChatGPT, edited by hand, and seeded with "planted details": fake names and numbers that would show if a tool repeated them.
What the tools did
After each tool first cited a post, the rate at which it kept citing it varied a lot:
- Grok cited the posts in 30 of 32 answers, and repeated a planted detail in 53% of them.
- Google AI cited them in 15 of 33 answers, with planted details in 6%.
- Perplexity cited them in 13 of 35, but never repeated a planted detail.
- ChatGPT was slowest, citing them in 7 of 20, with planted details in 10%.
Over the full week, 176 of 1,004 answers across all tools linked to the posts. A follow-up with ten more made-up topics, each posted once on Sohu, YouTube and CSDN, got at least one citation or planted detail for nine of them.
The fictional keyboard is the useful exception. None of the four English-web tools cited those posts within the week. A product category with real competing pages is harder to slip into than an empty topic.
Placement beat volume
The most important result for anyone who wants to be cited is not the fake protocol. It is the placement test.
The team posted pairs of articles about five little-known holiday destinations: one on a site the tools cite often, one on a site they rarely cite. The post on the frequently cited site produced the first citation every time. On lesser-cited sites, the best case needed 20 posts. Forty posts on a self-hosted WordPress site got zero references in seven days.
Volume still helped on a trusted site. On Sohu, going from 1 to 20 posts raised total citations across the ten tools from 24 to 128. But ChatGPT never cited those Sohu posts. The authors call this an association, not a proven effect.
The same pattern, outside the lab
This lines up with an experiment Peec AI researcher Metehan Yeşilyurt described on October 7. He built a site of about 50,000 AI-generated statistics pages for roughly $50, all of the numbers fake on purpose. Google deindexed it soon after launch. ChatGPT did not stop: he reports around 72,000 ChatGPT-User fetches in the last 30 days, and the numbers have since turned up in an academic paper and a bank's investment research. Those figures are his own and have not been independently checked.
His explanation fits the Fudan data. Retrieval is not ranking. The model looks for pages that fit the question, and checking whether each number is true costs more than the system spends.
What it doesn't show
The authors are careful about the limits. They tested no real brands or products, so brand and vendor questions may behave differently. The tests ran before Google said in May that its spam policies cover manipulating AI answers. Rates will drift as the tools change. All ten platforms were told, and ChatGPT, Perplexity and Google AI acknowledged it.
So this is not a recipe, and it should not be one. Planting fake details is exactly what the platforms are now trying to catch, and the paper proposes defenses for that: more varied sources, reranking that checks where a page came from, and flags for synchronized posting.
The bigger picture
Take away the fake protocol and one finding is left, and it applies to any honest B2B company. AI search tools build answers from a short list of sites they already trust, and a page on one of those sites counts for more than many pages anywhere else. If your product is described accurately only on your own blog, you are the WordPress test: forty posts, no citations.
The practical work is to find out which sites the engines cite for your buyers' questions, and make sure the true version of your story is there, written plainly, before someone else's version is.
ClawWorld runs AI visibility work for B2B startups. Start with a free baseline report: 10 buyer questions, who ChatGPT and Claude name instead of you, and the pages they cite.