Search works surprisingly well until customers don’t know what they want.
Type an exact product name into any decent store search and you get the product. Type “a quiet laptop for recording music at home” and you usually get nothing, or a wall of loosely related items. The store stocks the right machine. The customer just described their need precisely. The interface lost the sale.
I have read thousands of conversations shoppers have with our assistants, and this is the pattern underneath most of them: the customer knows their problem long before they know the product. The tools we have given them for thirty years assume the opposite.
Two kinds of finding
Every shopping journey starts in one of two ways.
Sometimes customers know exactly what they want.
“Nike Pegasus 42.”
Other times they only know the problem.
“I’m starting running and my knees hurt.”
Those are fundamentally different problems. Yet we give both customers the same interface.
Information retrieval research has names for the two: lookup and exploration. The research on conversational browsing draws the consequence: a query box assumes the customer can formulate their need upfront, and dialogue becomes valuable exactly when they can’t.
Shopping is mostly the second kind. A customer buying a gift, a first bike, a beginner wing foil, a retinol serum, does not have a query. They have a situation: a budget, a constraint, a person in mind, a vague picture of the outcome. Formulating that as keywords is work we have quietly outsourced to the customer, and the customer is bad at it through no fault of their own.
Even the knee-pain runner who finds the right words hits a second wall. Your catalog says “max-cushion trainer.” Your customer says “soft running shoes.” Neither is wrong, and they will never meet. This is measured, not anecdotal: in 1987, Furnas, Landauer, Gomez, and Dumais published the vocabulary problem, showing that two people spontaneously pick the same word for the same thing less than 20 percent of the time, and that an interface keyed to one word fails 80 to 90 percent of first attempts.
What stores do about it today
The industry’s answers are not stupid. Each one attacks a real piece of the problem.
Search engines add synonyms, typo tolerance, and increasingly semantic matching. Filters expose the attribute space so customers can narrow by size, price, and color. Recommendation carousels surface products the customer did not ask for but might want. Category pages encode a merchandiser’s mental model of the catalog. And chatbots, in their scripted era, tried to bolt a conversational surface onto all of it.
If you run a store, you probably have all five. They are worth having.
Why they only partially work
Each tool assumes something the exploring customer does not have.
Search assumes a formulated query. Filters assume the customer already understands the attribute space: someone who has never bought a trail bike cannot filter by suspension travel, because the filter is the question they came to ask. Carousels assume past behavior predicts this purchase, which fails precisely for the infrequent, considered purchases where help matters most. Category trees assume the merchandiser’s taxonomy matches the customer’s mental model, and the vocabulary problem says it usually doesn’t.
Scripted chatbots failed differently: they put a conversational surface over a fixed decision tree, so they could accept a question in natural language and still not understand it. The surface promised exploration. The machinery could only do lookup.
Notice what none of these tools can do: ask a question back. Every one of them is a single shot. The customer emits a signal, the system returns a result set, and if the result set is wrong there is no repair mechanism except starting over. A good floor salesperson is almost entirely repair mechanism.
A different frame: the discovery loop
“A chatbot that knows my products” is the wrong way to think about an AI shopping assistant. It is an interface that runs the loop a competent salesperson runs. Watch it once and you have the whole idea:
Customer: I need a bike to commute in Lyon. There are real hills.
Search: Hundreds of commuter bikes.
Assistant: How far is the ride?
Customer: Eight kilometres.
Assistant: And do you mind maintaining a chain?
Customer: I’d rather not.
Assistant: Then belt drive is probably a better fit. Here are three options. This one climbs well enough for Croix-Rousse.
Customer: Too expensive.
Assistant: Same idea, lower tier.
Elicit context. Retrieve candidates. Explain the fit. Refine on the response. That loop is the product.
Chat is merely the surface of this loop. The substance is underneath: retrieval that matches meaning rather than letters, grounding that restricts answers to what the catalog and policies actually say, and ranking that reduces twelve plausible products to three explainable ones. I wrote separately about why retrieval is the real bottleneck; an assistant with weak retrieval is a fluent voice reading from the wrong shelf.
This frame also explains why “chatbot” is the wrong word. A chatbot is defined by its surface. An assistant is defined by its loop. You could run the same loop through search suggestions, product page Q&A, or cart comparisons, and Amazon is doing exactly that, with no chat bubble in sight.
The loop does not solve everything. It adds latency where a good search adds none. It can ask one question too many and lose an impatient customer. And a badly grounded assistant fails worse than a search box, because it fails fluently. The loop is a better default for exploring customers, not a replacement for lookup.
What this means in practice
If you accept the frame, a few priorities reorder themselves.
Product data becomes interface. The assistant answers from your product text, so a description that says who the product is for and what it is like to use is now functional, not decorative. “Great bike!” gives the loop nothing to retrieve or explain.
Grounding is non-negotiable. Every claim traceable to catalog or policy, and an honest “I don’t know” for everything else. One invented discount costs more trust than a hundred good answers earn.
Explanations matter as much as picks. “This one, because you mentioned hills” is doing real work: it lets the customer correct the reasoning, not just reject the result. Transparency is what makes the loop converge.
And the conversations themselves become an asset no analytics tool has: customers writing down, in their own words, what they came for and did not find. That is demand data. Exit rates never contained it.
The shift
We spent thirty years teaching customers to compress their needs into two or three keywords, then built increasingly clever machinery to decompress them. The compression step was never necessary. It was a limitation of the interface, mistaken for a fact about shopping.
An AI shopping assistant, properly understood, just removes the compression. The customer states the situation; the system does the work. That is not a chatbot feature. It is the query box admitting it was always a workaround.