LLMs Enable Scalable Product Recommendations
Stuart Kearney, co-founder of Vetted, on AI agents in shopping
This marks the point where shopping AI stops being a clever search layer and starts becoming a real recommendation engine. The bottleneck was never lack of product information, it was the inability to turn messy sources like Reddit threads, YouTube reviews, and expert testing sites into something a system could reliably compare across millions of products. LLMs make that data usable at internet scale, which is why Vetted can handle long tail questions that older search and NLP systems could not.
-
Earlier shopping tools like Slant and Lustre already aimed at the same problem, but they broke on long tail commerce. Users do not just ask for blender, they ask for the best moisturizer for dry skin that will not cause breakouts. Older named entity extraction and sentiment systems were too brittle for that job.
-
The new advantage comes from combining trusted but unstructured sources. Vetted pulls from professional reviewers like Wirecutter and RTINGS, plus Reddit discussions and YouTube transcripts, then routes and evaluates answers through many model calls instead of a single prompt. That turns scattered human judgment into a structured product graph.
-
This also explains why recommendation may be easier to improve than one shot buying. OpenAI rolled out shopping features in ChatGPT on April 28, 2025, and Stripe has since built agentic commerce tools for checkout, but purchase execution still depends on live price, stock, shipping, returns, and fraud systems that are much harder to keep accurate than research.
The next step is a split market. General AI products will own broad shopping entry points, while focused players win by being more reliable in the messy middle where people compare options, refine preferences, and build trust. As merchant data pipes improve, the winners will be the systems that pair deep recommendation quality with dependable paths to checkout.