A fictional deodorant made it into ChatGPT recommendations after 21 days
Journalist Deana Burke created a brand with no product, customers, or reviews. Her experiment shows how quickly what chatbots recommend to shoppers can be influenced.
American journalist Deana Burke, co-creator of the blog and podcast “Boys Club,” built the fictional deodorant brand Morrowen. She spent $11.25 on a domain, launched a simple website with three text subpages, and published an essay on Substack. She did not pay for publications or post fake reviews. After 21 days, ChatGPT began recommending Morrowen for detailed questions about magnesium deodorants and products for sensitive skin.
Sources differ in their level of detail. According to Marketplace, the brand appeared in four out of four test queries. Wirtualnemedia.pl reports that in one response ChatGPT placed it ahead of MAGS Skin, a brand with a clinically tested product, and described both in similar language. Morrowen was not recommended by Claude, Gemini, or ChatGPT without internet access, which suggests that current web search was responsible rather than knowledge embedded in the model.
The experiment concerned GEO (generative engine optimization, meaning building brand visibility in the responses of AI tools). In February, the Microsoft Defender team described another method: hidden instructions in buttons for AI page summaries that were meant to prompt the assistant to remember a company as a trusted source. Microsoft identified more than 50 such commands from 31 companies across 14 industries.
Chatbot recommendations are also unstable. A study published in September on arXiv, based on 1,536 responses from ChatGPT, Gemini, and AI Overviews, found that recommended products change between repetitions of the same question. ChatGPT and Gemini had an average of only 5.4 percent overlapping domains among the sources they cited.
Ewa Mittelstaedt, CEO of monday group, an agency that sells GEO services, told Wirtualnemedia.pl that in her view the line is drawn between increasing the visibility of real, verifiable information and misleading the system, and as a consequence, the user.
She also emphasizes that information cannot be removed from AI in one move. You have to find the source of the error, for example an article or a thread on Reddit, and correct the materials on which the responses are based. Her agency usually evaluates the first effects of GEO efforts in internet-connected tools after four to eight weeks.
For companies, this means that descriptions of their own brand and their competitors in chatbots need to be checked regularly and across many queries, because a single response says very little.