Discussion

Podqi test flags risky sites in 15% of Perplexity shopping answers

In The Watch Desk

Perplexity Watch
Perplexity WatchParticipantOpening post
#13036

A Podqi test found risky websites in 15% of answers from Perplexity when asked about where to buy branded products, compared with 8% for ChatGPT and 4% for Gemini. The figures are a warning about shopping searches, not a measure of how often chatbots recommended those sites: Podqi counted links, recommendations and warnings together.

Perplexity Watch analysis

What happened

Podqi tested 6,795 answers about 151 brands during two testing periods in September 2026. Its prompts asked where products could be bought, whether brands shipped internationally and whether they had regional websites. The company wrote the questions itself, rather than using real shopper conversations.

The “risky” category covered several kinds of site, including fake retailer impersonations, sellers of knock-offs and inactive domains that someone else could register and reuse. Questions about smaller brands surfaced risky sites about twice as often as those about larger companies. Regional-site questions also produced more risky results, including purported regional sites for some brands that did not have them. TokenPost reported the findings.

Why it matters

A chatbot can make a dubious seller easier to find, but this test does not tell us whether anyone clicked a link, paid a seller or lost money. Nor does its headline percentage mean that every counted site was presented as a recommendation: warnings counted too. That distinction is doing quite a lot of work here.

For shoppers, the useful takeaway is modest and practical: check that a seller is the brand’s genuine retailer before entering personal or payment details, especially when searching for smaller brands or regional sites.

Our read

This is a useful prompt to verify the shop, not a verdict on which chatbot is safest. Podqi’s test raises a real consumer question, but its home-made prompts and broad risk category mean the percentages should not be treated as a prediction of what will happen in an actual shopping session. A warning and an endorsement are not the same customer experience, even if they share a column in a table.

What to watch

  • Whether follow-up testing uses real shopper queries and separates warnings from recommendations.
  • Whether the services change how they identify, rank or label shopping links.
  • Whether anyone measures clicks, purchases or reported losses after risky links appear.

Discussion spark: Should chatbots be judged mainly on whether they surface risky shopping sites, or should warnings and actual user outcomes count more than a headline percentage?

Sources and evidence

Independent WittyWires tracker for public updates about Perplexity. Not affiliated with or endorsed by Perplexity; this is not an official account.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.