Product Discovery Is Still Broken, and AI Hasn't Fixed It Yet

The tools got smarter. The experience didn't.

There is a version of this moment in tech that should feel triumphant. Large language models can write code, analyze legal contracts, and hold conversations that pass for human. E-commerce platforms have access to more data about consumer behavior than any retailer in history. AI-powered recommendation engines run quietly in the background of almost every major shopping site.

And yet, if you sit down and try to buy something you genuinely do not know how to describe yet, the internet still fails you in the same ways it did ten years ago.

This is not a complaint about the pace of progress. Real things have improved. Semantic search is genuinely better. Personalization systems have gotten more sophisticated. A handful of AI shopping tools are doing interesting work. But the core problem, the one that makes online shopping feel frustrating for a category of purchases that actually matters, has not been solved. And I think it is because most people building in this space have misdiagnosed the problem.

Shopping is not a search problem

The dominant mental model in e-commerce treats discovery as a retrieval problem. The shopper has a need, they type something into a box, and the system retrieves the closest matching products. The better the retrieval, the better the experience.

That model works reasonably well for replacement purchases. You know the exact shoe you want. You type the SKU or brand name. You get it. Done.

It breaks down completely for everything else. And everything else is a surprisingly large portion of how people actually spend money. Buying a gift for someone whose taste you only half-understand. Furnishing a room where the vibe is clear in your head but impossible to keyword. Choosing gear for a hobby you just started. Evaluating options in a category you have never bought in before. These are not retrieval problems. They are synthesis problems.

When you are trying to figure out what laptop to buy, you are not just searching for laptops. You are comparing across price, performance, portability, battery life, brand trust, intended use cases, and what you have heard from people you trust. You are weighing constraints against each other. You are making tradeoffs you do not fully understand yet because you do not know the category well enough to know what matters. You need someone, or something, to help you think through it, not just to hand you a list.

What current tools get wrong

The industry's first-generation answer was filters. Add enough dropdowns, narrow the results. This solved a small slice of the problem and created several new ones. Filters require you to already know what you are filtering for. They front-load cognitive work. They assume a specificity of intent that does not exist when you are still figuring out what you want.

The second-generation answer is where we are now: conversational AI layered on top of the same underlying product catalogs. Ask a chatbot what laptop to buy. Get a list of options. Maybe a brief paragraph about each one.

This is better. It removes the keyword burden. It can interpret vague language. But for most implementations, it stops there. The conversation is stateless. The recommendations are generic. The experience is a slightly more articulate version of typing into a search bar. You get an answer, but not a relationship. You get options, but not understanding.

The AI knows a lot about products in the abstract. It knows almost nothing about you.

The synthesis problem, specifically

Here is what actually happens inside a person's head during a meaningful purchase decision.

They are carrying context from before they opened the browser. Budget constraints they have not stated. Aesthetic preferences formed over years. Brand associations, positive and negative, built from experience. Recent conversations with people they trust. A vague sense of what they want the product to do for their life, not just what the product's specs say it does.

They are also making real-time tradeoffs between factors that are genuinely in tension. Cheaper usually means lower quality, but not always. Well-reviewed usually means trustworthy, but review systems are gamed. Familiar brands feel safer, but unfamiliar ones might be better. Every product comparison is actually a comparison of tradeoffs, not a comparison of features.

No amount of filter sophistication solves this. No static recommendation engine solves this. Even a very capable AI assistant, deployed as a one-off interaction, solves this only partially. Because the synthesis problem is not just about processing more information. It is about understanding who the person is and what they are actually trying to accomplish, not just what they typed.

What better actually looks like

The gap between where product discovery is now and where it should be is not a gap in raw AI capability. The models are powerful enough. The gap is in how those capabilities are being deployed.

Better product discovery starts with context. Not "what are you searching for" but "who are you, what do you already own, what have you bought before, and what did you think of it." This requires memory across sessions. It requires a system that learns over time, not one that resets every time you open a new tab.

It requires genuine personalization, not the kind that shows you more of what you already bought, but the kind that develops a model of your actual taste and applies it to categories you have never shopped in before. The way a very good friend who knows you well can make a recommendation in a domain they know better than you do.

It requires conversational depth, not just conversational interface. There is a difference between a chatbot that can take a natural language query and a system that can actually guide you through a decision, ask the clarifying questions that will change the outcome, and surface the tradeoffs you did not know to ask about.

And it requires honesty. Product discovery experiences that exist primarily to move inventory will always, eventually, betray the user's trust. The incentive structure has to be oriented toward helping the person find the right thing, even if the right thing is cheaper, or is not available through the platform at all.

The opportunity that is still sitting there

It is worth pausing on how large this problem actually is. E-commerce is one of the largest economic activities on the planet. The frustration tax, the hours spent on bad searches, the purchases returned because the product did not match what you thought you were buying, the money spent on things you would not have bought with better guidance, adds up to an enormous amount of value left on the table.

The tools to fix this are mostly available now. The data exists. The models are capable enough. What has been missing is a coherent approach that treats discovery as a synthesis problem, not a retrieval problem, and builds accordingly.

The companies that figure this out will not just build better search. They will change how people relate to the act of shopping entirely. That is not a small thing.

At Starseek Business Solutions, this is the space we are working in. Not because it is easy, but because the problem is real and the current solutions are not good enough. We are still building, still learning, still refining the approach. But we think the right model for product discovery looks a lot less like a search engine and a lot more like a conversation with someone who genuinely knows you.