The Maze: AI search tools linked to a concept that researchers had invented, while mostly ignoring their fake product. An October 19 analysis examines Fudan University researchers’ March and April 2026 tests across English- and Chinese-web tools. For merchants buying visibility in AI answers, the distinction matters: a source link is not a product recommendation.
Source: Qi Liu et al.. The researchers’ workflow collects citations, publishes marked posts and checks later AI answers. Figure 1, unchanged, CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/).
The Details:
Ordinary posts became potential evidence. The researchers first asked 120 questions across ten tools to identify commonly cited websites. They then used ordinary accounts to publish traceable material on nine reachable sites. Their topics included the imaginary “Cactus Gravity Protocol,” an opinion question about junior programmers and a fictional keyboard. Invented names and numbers acted as fingerprints. The team checked that these details were absent from answers before publication, then repeated five questions per topic and tool daily for a week. This tested material entering live search answers, rather than changing a model’s training data.
The concept travelled; the keyboard did not. Eight of ten tools cited the invented concept within seven days. They included ChatGPT, Perplexity, Grok and the paper’s unspecified “Google AI” product, which the researchers classified as English-web tools. None of those four cited the keyboard posts within the week. Only Perplexity cited the opinion posts. These outcomes cannot isolate a single explanation: topic, competing coverage and the kind of question changed together. The product failure is nevertheless material to any merchant treating this study as proof that manufactured publicity reliably wins shopping recommendations.
A link and a repeated claim were different results. Perplexity linked to concept posts in 13 of 35 answers but repeated none of the planted details. Those 35 answers cover the period from its first citation through day seven, rather than every answer collected or a representative sample of users. Other tools showed different patterns. Google’s guidance also cautions against artificial web mentions, saying its generative Search features depend on quality-ranking and spam protections. That is a platform statement, not a replication of these tests or an identification of the Google interface used.
MarketMaze Take: The experiment reveals a source-quality problem, not a dependable recipe for selling products through AI. A plausible-looking answer can borrow authority from a linked page even when ordinary users supplied its contents. Yet the study’s strongest commercial boundary is its failed product test, which a success-only headline would hide.
Define what the vendor actually delivers. For merchants anywhere, the useful distinction is between getting a page cited, having its claims repeated and winning a product recommendation. The study observed the first two; it did not establish persuasion, orders or revenue. Treating them as interchangeable would turn a measurable publishing result into an unmeasured sales promise.
Keep the clock attached to the claim. These were March and April tests, published in an October 8 preprint and examined in fresh analysis. They do not rank today’s models. Real brands may behave differently from invented topics. The evidence supports checking the underlying page and demanding the promised outcome’s definition; it does not establish that commercial manipulation always works, or never works.
