Structuring Unstructured Shopping Data
Stuart Kearney, co-founder of Vetted, on AI agents in shopping
This is the key technical shift that turned shopping advice from a content problem into a data problem. Before LLMs, sources like Reddit threads, YouTube transcripts, and review articles were readable by humans but too messy for software to reliably pull out the product name, the context of use, and whether people liked it. Vetted’s product works by turning that raw text into structured fields that can be ranked, compared, and refreshed across thousands of shopping queries.
-
In practice, structuring means extracting concrete facts from messy text, like which blender is being discussed, whether praise is about durability or price, who is making the recommendation, and in what scenario. That lets Vetted rank products on thousands of signals instead of just matching keywords.
-
The business implication is coverage. Manual review sites can test a limited set of products a few times a year, but once web scale product chatter becomes machine readable, a recommendation engine can cover far more of the long tail, from bath towels to skincare to niche accessories, while still layering in human review.
-
This also explains the split with competitors. PerfectRec starts with user quizzes and expert ratings, while Perplexity and ChatGPT are adding shopping flows and product cards. Vetted is centered on the harder middle step, turning scattered product discussion into a dependable research graph before pushing users to checkout.
The next phase is that merchants and publishers will increasingly publish cleaner machine readable product data, while shopping assistants compete on who can best combine that structured merchant data with trusted unstructured signals from creators, reviewers, and user communities. The winners will be the products that make messy consumer research feel as consistent as a database query.