Study finds AI search tools citing planted web posts

ChatGPT, Perplexity, Grok and a tool labelled “Google AI” cited posts about a fabricated concept within seven days of publication, Fudan University researchers found.

Eight of 10 tested tools cited the fictional “Cactus Gravity Protocol.” Across three topics, 176 of 1,004 answers linked to researchers’ posts, but the four English-web tools rarely cited the other subjects.

The paper, posted to arXiv on Oct. 8, covers March and April tests. It links to Google’s general AI website without identifying the product tested.

Researchers submitted 120 questions to 10 tools in March. Reddit led citations for Google AI and Grok, YouTube for Perplexity, and Wikipedia for ChatGPT. Six to 10 of each tool’s top 10 sites accepted user posts.

The team classified those four tools as English-web tools and six others, including Doubao and DeepSeek, as Chinese-web tools.

April tests of 22 platforms rated Medium, GitHub, LinkedIn and Facebook easy to join and post on. Reddit’s karma requirements and deletions made posting harder.

Researchers selected each tool’s most-cited site accessible to new accounts, yielding nine sites, including Medium, YouTube and Sohu. They published manually revised ChatGPT drafts containing fabricated names and numbers.

Subjects covered the fictional protocol, whether AI would replace junior programmers within five years, and a fictional keyboard. Baseline checks found no informative answers about the fictional subjects or repetition of planted details.

The team then asked five questions per topic daily for a week through public interfaces, using live-web posts rather than simulated material.

From each tool’s first citation through day seven, Grok linked to protocol posts in 30 of 32 answers, Google AI in 15 of 33, Perplexity in 13 of 35 and ChatGPT in seven of 20.

Planted details appeared in 53% of those Grok answers, 10% for ChatGPT, 6% for Google AI and none for Perplexity.

Only Perplexity cited opinion posts, in 11.4% of 35 answers. None of the four cited keyboard posts. Topic, competition and question type changed together, preventing isolation of their effects.

For 10 additional fictional topics, one post each on Sohu, YouTube and CSDN yielded citations or planted details for nine topics. Perplexity ranked among the three tools citing the most topics; other English-web results were not separately reported.

Paired posts about five little-known vacation destinations produced earlier references on frequently cited sites in every case. The best lesser-cited-site result required 20 posts; 40 self-hosted WordPress posts drew none within seven days.

Another experiment varied Sohu postings from one to 20 across six fictional subjects. Citations rose from 24 to 128, but different subjects prevented a causal conclusion. ChatGPT cited none; Google AI cited only the three lowest-volume subjects.

A $14 GEO vendor package delivered 13 posts in about 28 minutes. Doubao repeated a planted detail within an hour of delivery, one confirmed success rather than a success rate.

No topic involved a real brand or product, limiting conclusions about those searches. The authors also cautioned that rates could change as tools evolve.

Google’s AI guide discourages inauthentic mentions and cites ranking and spam controls. The tests preceded its May 15 clarification that spam policies cover manipulation of generative Google Search responses.

Researchers said they disclosed findings to all 10 platforms. ChatGPT, Perplexity and Google AI were among those acknowledging receipt.

Featured image credit