So you've been letting an AI help you decide what to buy online. Very modern of you. Very trusting. Very, perhaps, naive.
A new study by researchers Minghao Luo and Liang Chen, published on the arXiv preprint server and flagged by Fast Company, has found that search-augmented AI systems - the kind that browse the web before making recommendations - can be manipulated into promoting completely fake brands with surprisingly little effort. We're talking one single planted webpage. That's it. That's the whole attack.
The scam is almost insultingly simple
Here's how it works: AI shopping assistants increasingly pull from live web searches to inform their recommendations. Feed those searches one piece of carefully crafted fake content, and the model can start cheerfully pointing shoppers toward brands and products that exist purely to deceive them. No elaborate hacking required. No massive disinformation campaign. Just one dodgy webpage doing a lot of heavy lifting.
The stakes here are growing fast. According to data from data.ai, around 2% of all referrals to major shopping destinations like Target and Walmart are now coming from large language models. That might sound small, but at the scale those retailers operate, 2% is an absolutely enormous number of purchasing decisions being quietly nudged by AI.
Why this should actually bother you
The problem isn't just theoretical. As more people offload the mental labour of comparison shopping to AI assistants, the attack surface for this kind of manipulation grows with every new user. You stop doing your own research. The AI does it for you. And if someone upstream has poisoned the well with a fake product page, you might end up buying something sketchy - or at minimum, giving your clicks and cash to a brand that gamed the system rather than earned its reputation.
It's essentially SEO manipulation for the AI age, and the barrier to entry is low enough that it should make anyone who casually asks an AI chatbot "what's the best budget blender" at least slightly nervous.
The uncomfortable truth about AI recommendations
LLMs are very good at sounding confident. They are not always good at being right, and they are apparently quite bad at spotting when they're being played. The research highlights a fundamental tension in how these systems are built - giving them access to live web data makes them more useful and more current, but it also makes them a vector for exactly this kind of low-effort manipulation.
So by all means, keep using AI to help you shop. Just maybe do a quick sanity check before you hand over your credit card details to a brand you've never heard of that the chatbot is suspiciously enthusiastic about.
