What AI Product Discovery Should Actually Look Like

The phrase "AI shopping assistant" has become so overused that it has almost lost meaning. Every major retailer has announced one. Most of them are the same thing: a chatbot layered on top of the same keyword-search infrastructure that existed before, with a natural-language wrapper to make it feel new. You type "blue waterproof jacket" in a chat box instead of a search bar, and the system returns a list of blue waterproof jackets. The AI is doing the same thing the search bar was doing. It is just more conversational about it.

That is not AI product discovery. That is AI-flavored product search. The distinction matters because discovery and search are fundamentally different problems, and conflating them has led the industry to build tools that feel impressive in demos but fail real shoppers in practice.

What discovery actually requires

Search is what you do when you know what you want. Discovery is what you do when you do not. These two modes of shopping require completely different system designs, and most people spend far more time in discovery mode than they realize.

Think about the last time you bought something genuinely new to you. Maybe you were kitting out for a hobby you had just picked up. Maybe you needed to buy a gift for someone with specific tastes. Maybe you were upgrading a category where your knowledge was years out of date. In every one of those situations, you did not walk in with a query. You walked in with a context: a set of circumstances, constraints, preferences, and unknowns that needed to be translated, through conversation, into something concrete.

The gap between that context and a concrete product is exactly what AI product discovery should bridge. To do that, a system needs to handle three things that conventional search simply cannot.

The first is intent inference. When someone says "I need something for running this summer," they are not describing a product. They are describing a situation. A useful AI assistant infers from that situation a set of product attributes: lightweight fabric, moisture management, likely a short-sleeve or singlet, possibly with reflective elements if they run at dawn or dusk. None of those attributes appeared in the query. They were inferred from context. Current AI shopping tools handle surface-level paraphrase well. Genuine intent inference is much harder and much rarer.

The second is tradeoff articulation. Products rarely have a single right answer. They have a set of tradeoffs, and the right choice depends on what the shopper values most. A trail running shoe that excels at stability may sacrifice cushioning. A camera bag that protects gear may not fit carry-on requirements. A noise-canceling headphone that sounds extraordinary may be too heavy for daily commuting. Humans who know a product category deeply understand these tensions intuitively and surface them during conversation. An AI assistant should do the same. Most do not. They return results ranked by relevance and popularity, which tells you nothing about why one option is better for your specific situation than another.

The third is cross-session personalization. A one-time conversation can only get so far. The assistant can learn what you say during that session, but it cannot learn from what you do not say: the products you clicked on and then closed, the options you were almost persuaded by, the purchases you made and later returned. Over time, the pattern of those signals reveals something much richer than any single query: your actual taste, your real price sensitivity, the brands you trust and the ones you do not. An AI assistant that only knows what you told it this session is leaving most of what it could know about you on the table.

Where current experiences fall short

The best way to understand what is missing is to describe what a genuinely good shopping experience looks like in a high-end physical store, and then notice what breaks down online.

Walk into a specialty outdoor gear shop and tell someone you are planning your first backpacking trip. Within a few minutes, they have asked about your fitness level, your route, the time of year, whether you care about ultralight philosophy or just want something reliable, and whether you have ever slept outside before. By the end of that conversation, they can put a specific pack, a specific sleeping bag, and a specific tent in front of you, and they can explain in plain language why each one is right for you specifically and what you are trading away compared to the alternatives.

That experience requires four things: domain knowledge, conversational skill, the ability to hold and synthesize context across an extended exchange, and the credibility to make a confident recommendation. AI systems have gotten dramatically better at conversational skill and context retention. Domain knowledge, in many categories, is now genuinely competitive with a human expert. The piece that still falls down most often is the recommendation itself: the moment of synthesis where all the gathered context resolves into a specific, justified answer.

Most current AI shopping assistants hedge. They present five options and call them all "great choices." They qualify every statement with "it depends on your preferences." They stop just short of the moment where a recommendation becomes useful, because committing to a specific answer requires a level of calibration that is still difficult to achieve reliably. The result is a system that asks good questions and then fails to use the answers.

The other persistent failure mode is single-retailer confinement. When an AI assistant lives inside one retailer's app or website, it can only recommend products from that retailer's catalog. This is fine for replacing a product you already love. It is completely inadequate for discovering the best option across the market. The best trail running shoe for your foot type may not be sold at the retailer whose chatbot you happened to open. A genuinely useful AI shopping assistant needs to be retailer-agnostic, searching across the full product landscape rather than a curated slice of it.

Why personalization is not optional

There is a common misconception that personalization in product discovery is about convenience, like a streaming service that learns your genre preferences so you spend less time browsing. In shopping, personalization does something more fundamental: it changes what the right answer actually is.

Two people asking "what is a good everyday carry bag?" have asked the same question, but the correct answer for a person who commutes by bike in a rainy city is completely different from the correct answer for someone who drives and needs to look professional in client meetings. Neither person may have mentioned any of that context. A personalization layer that knows their browsing history, past purchases, and even their implicit signals around price and brand can make an informed inference. Without it, the system is answering a generic question for a generic person, and generic answers serve nobody particularly well.

This is why personalization needs to accumulate across time, not just within a session. The most valuable signal is not what someone says when they open the app. It is the sum of everything they have done over weeks and months: the categories they return to, the price range they consistently land in, the products they research carefully versus the ones they buy on impulse. Those patterns reveal preferences that most people could not articulate if asked directly. A well-designed AI product discovery system learns those preferences without requiring the user to explain themselves.

What a better system looks like

The standard should be an AI assistant that behaves like a trusted advisor with encyclopedic product knowledge and deep familiarity with your particular taste. It asks questions, but only the ones it does not already know the answer to. It synthesizes your context into a genuine recommendation, not a ranked list with hedging commentary. It explains the tradeoffs of one option versus another in language that maps to what you actually care about. It searches across every relevant retailer, not just the one it happens to work for. And it gets meaningfully better every time you use it, because it is learning who you are.

This is harder to build than a search bar with a chat interface. The retrieval architecture needs to handle semantic search, not just keyword matching. The personalization layer needs real-time ranking that balances relevance against diversity. The recommendation logic needs calibration data from real outcomes, not just from clicks. The underlying product data needs to be clean enough that the system can actually make meaningful comparisons across brands and platforms.

None of these are unsolved problems. They are hard engineering and hard product work, but the path to each of them is clear. The question is whether anyone is willing to prioritize them over the simpler version of "AI shopping" that can be shipped faster and announced with a press release.

At Starseek, we are not interested in the simpler version. The system we are building, including Spark, our AI assistant, and the broader infrastructure around cross-platform search and personalization, is designed around what a genuinely great shopping experience requires, not around what is easiest to demo. That means slower development in some areas and harder tradeoffs in others. It also means that when the product is right, it will actually be useful.

That is the standard the industry should be held to. AI product discovery is not a feature. It is a fundamentally different way of helping people find what they need. The tools that treat it seriously will earn people's trust. The ones that do not will be another chatbot that felt exciting for a week and then got ignored.