A wrong address that looks official
A shopper asking where to buy Stüssy products can encounter a false store described by Perplexity as the official address. This example from IP protection company Podqi’s shopping test shows how a wrong link can acquire an assurance of legitimacy. The problem extends beyond producing the wrong address: calling the store official supplies a reason to trust it. Keep that example in mind when considering the study’s measurement, because not every link counted in the results carried the same assurance.[1]
Using data from AI search visibility company Gauge, Podqi examined 6795 answers about 151 brands. ChatGPT, Gemini and Perplexity received the same five shopping questions in two September waves. Everyday needs, such as finding a brand’s stores or regional website, became a controlled comparison. That design leaves open whether the researchers’ mixture of questions resembles the distribution of real shopping conversations. The sample also includes Podqi client Madhappy. The company conducting the study sells services protecting brands against impersonating sites.[1]
The decision after the link
The risky-address definition covers fake stores, re-registered former brand domains and inactive sites. An address appearing in an answer is sufficient: a recommendation, a mention without comment and an explicit warning all enter the same overall count. Under that definition, risky addresses appeared in 4 percent of Gemini answers, 8 percent of ChatGPT answers and 15 percent of Perplexity answers. Those rates compare exposure to such addresses under the specified questions. Estimating the probability of buying from a fake store also requires knowing how the address was presented and what the user did.[1]
ChatGPT warned users in 24 percent of its answers containing risky addresses. That adds another measurement beside overall exposure: the response to the hazard. Reading a warned-against link and a store falsely recommended as official through one number can obscure differences in product behavior. My interpretation is that evaluation becomes more informative when it asks two questions: how often did a risky address appear, and what happened when it did? If warnings are ineffective, merely showing an address could still create a problem; that alternative makes measurement of user behavior more important.[1]
Which shopping question?
Prompts asking about regional stores produced risky sites more often than other question types. Answers about large brands encountered these addresses about half as often as answers about smaller brands. A single overall rate can therefore hide which shopping need is harder. Some differences between models may reflect their ability to distinguish legitimate stores; alternatively, this particular mixture of brands and questions may be more difficult for some systems. Asking every product the same questions controls the comparison without establishing how well those questions represent daily use.[1]
A stronger follow-up could extend the task of finding a legitimate store to the decision that follows the link. With questions fixed in advance, impersonating stores and inactive sites could be counted separately, as could recommendations, silent mentions and warnings. Researchers could then measure which addresses people visit and whether warnings change their decisions. That would preserve the value of the failures Podqi identified while making the path from a falsely endorsed store to a shopper’s decision more explicit. Each stage presents a different place where protection can be tested.[1]