Idea: A New Way to Shop Online

Background

After our learnings in the wine space, it seems as though a lot of people have interest in getting recommended things. This makes sense, at a high-level: there are a lot of products out there in the world, and even for highly artisanal industries, it’s hard to hand curate every single item.

Now bring in LLMs. They introduce a new concept we haven’t seen before: being able to talk to a computer program in pure natural language and get a coherent response in return. In general, this can unlock a lot of value props, but the ability to search through anything with a natural language query is a pretty powerful capability. When it comes to shopping, I like to think about it like: when you’re in a store, do you just tell people you want a specific item? Or do you typically have a description of what you want, including a lot of detail that may be messy and that you probably couldn’t find online?

Idea

The concept is simple: we build a shopping experience that allows people to find items by describing their search in natural language. We can aggregate over multiple sites, and we can make money by using affiliate links, and thereby taking a certain cut when people buy our product — our validation that our searches are actually good and that people are willing to use our site.

An existing site in Wirecutter already exists — this NYT offering has experts curates lists of items and then also monetizes using affiliate links. The concept is there, but I do think it’s difficult b/c there’s a certain limit to how many items you can review using just human-curated lists. And while you can argue that our idea may potentially take out the human aspect, there are ways to actually make this process more human-centered while sifting through a lot of the data. For instance, upon search results, you could simply also link and highlight the product reviews the algorithm found most relevant to your search query, enabling the user to conduct their own manual investigation.

Another key idea here is that even as it rides the tailwind of LLMs, better models won’t necessarily take away our product. In particular: if we ship fast to build a really strong end-to-end user journey — think aggregated reviews and scores, linking to specific reviews, more features — then better models will only entail better search results. There’s also an opportunity to introduce novel opportunities in the space by combing techniques like retrieval augmented generation (RAG) as well as existing research in recommender systems (Netflix Lagos).

Closing Thoughts

Overall, this is at least a good first step that leverages our learnings and our interests from our initial needfinding (curation and recommendation) while also being an interesting technical build. This is also just good for getting a grade in the class. Some questions exist here around initial traffic necessary and the viability of the monetization strategy, but I think there’s value in shipping something quick and seeing what this hypothesis can tell us.