Keyword search in commerce has always suffered from a translation problem. Manufacturers describe products in the language of specs and engineering. Brands translated that into marketing and categories. Customers must learn the vocabulary, jargon, attributes, and model numbers of what they’re looking for.
Think about high-information purchases: cars, nutrition supplements, electronics. Customers are expected to understand the domain before they can shop confidently. For hobbyists, that’s part of the thrill. But the market of casuals always exceeds the market of hobbyists, and imposing this burden on every customer leaves real demand untouched.
Goal-based shopping addresses this. Rather than retrieving products that match a customer’s keyword fluency, it uses an LLM’s understanding of natural language to match products to a customer’s actual intent. The key evolution isn’t the conversational interface, but the shift in what product data needs to describe. Instead of tagging products solely for keyword retrieval, retailers can serve customers at the level of what they’re trying to accomplish: a camping trip, a daughter’s fifth birthday, a first Thanksgiving hosting.
Goal-based shopping isn’t a new behavior that AI enabled, but the original way people think of shopping. It was actually the search field that forced consumers to translate their real world goals into database syntax. We got good at it the same way we got good at Google searches, not because it was natural, but because it was the only interface available.
Build — But Know Why You’re Building
Retailers (and really, everyone) are racing to add AI-powered search to their websites. This is a good decision. Experimenting is the fastest way to build the organizational AI capabilities every business will need. Teams must get familiarized with building and managing probabilistic, AI-based systems. Only through building will individuals and organizations understand what’s possible.
But for retail organizations, building better customer experiences is less impact than learning to use proprietary data to unlock insights that no foundation model can replicate. In short: build experiences to learn, build data to win.
What’s Uniquely Yours
Frontier AI companies are shipping at a remarkable rate, setting new standards for interfaces, protocols, and infrastructure. The right question for most companies isn’t how to keep pace with every AI development, but what competitive advantage your organization has that frontier AI companies and competitors do not. For retailers, the answer is customer data.
Years of purchases, co-purchase patterns, subsequent orders, return signals, browsing behavior, loyalty data — these are the raw materials for goal-aware product catalogs. As more purchases flow through goal-based shopping, the moat of this proprietary data deepens.
What Goal-Aware Product Data Actually Looks Like
There’s a version of “improving product data for AI” that’s just better keyword search in disguise. Examples include using AI to catch hyper-specific product attributes that were missed in the original listing: the exact shade of blue, the fabric composition, the closure type. It’s valuable, but it’s still attribute-matching, meaning a customer still needs to know they want “moisture-wicking polyester blend with a quarter-zip” to find it.
Goal-aware product data is fundamentally different. Consider this question:
"I'm 45, I've been feeling exhausted lately, I just started working out again after a few years off, and I'm trying to lose 20 pounds. What should I be taking?"
This is an actual question that was asked by actual person, before they start conducting research. The customer knows their situation, not their solution. Without product name, category, or attribute list, keyword search can’t do anything with it.
A cross-sell engine might do better but misses the point. It sees “weight loss” and surfaces fat burners. It sees “working out” and suggests protein and creatine. It sees “exhausted” and suggests preworkouts. Individually, these recommendations make sense, but do nothing to consider the simultaneous factors.
An LLM can. It understands that this customer has three compounding goals (energy restoration, re-entry recovery, body composition) with conflicting constraints (calorie deficit for weight loss, but muscle preservation for someone rebuilding after a long break). It can reason about the interaction. What an LLM cannot do is tell you what 45-year-old re-entry fitness customers at your store actually buy: what they try first, what they return, what they quietly abandon, and what they add four months in when things are working.
That’s where your data comes in. You know that customers like this typically start with energy support, B12, CoQ10, magnesium, before they’re ready to think about protein. You know they often buy an ambitious pre-workout, then come back six weeks later for a stimulant-free version. You know the ones who stick tend to add vitamin D and omega-3s by month three, often before they can articulate why. You know that how quickly a customer reorders magnesium tells you whether their sleep and recovery are actually improving.
None of that exists in any foundation model’s training data. It exists in your transaction logs.
And here’s what it means for your product catalog. A protein powder today is described as: 30g protein, chocolate, 300 calories, 5g BCAAs. That’s a manufacturer spec. A goal-aware product record also captures: best suited for muscle preservation during a caloric deficit; commonly purchased by re-entry fitness customers in their 40s and 50s; often added in month two, not month one, for this cohort. That second layer doesn’t come from the manufacturer. It comes from watching how your customers actually use the product, and encoding that behavioral knowledge into the data layer your AI can read.
Customers simplify how they think of products. LLMs handle the intent-parsing, and your data gives AI agents new surface area for customers to find your products.
The Work Nobody Wants to Do
All of this is solid in theory. In my experience, however, the data to build goal-aware product catalogs is a mess.
Data lives across ten different systems. Some owned internally, some owned by vendors, all with different structures, different APIs, different update cadences. Getting AI tools to actually consume this data right now involves a lot of copy-pasting, a lot of manual translation between systems, and loads of testing. There’s no OpenClaw for unifying disparate legacy retail systems. There’s no viral demo for reconciling product taxonomies across three platforms.
Think of it as tech debt for the AI era. Any meaningful AI initiative, including goal-based shopping, will eventually force you to confront data fragmentation. You can confront it reactively, scrambling to clean data every time you want to try something new. Or you can confront it proactively, treating data readiness as the investment that makes every future AI initiative possible instead of painful.
The point isn’t to build the best AI experience right now. It’s to prepare your data to make connecting to new AI experiences easy, and focus on the ones that move the needle.
The Real Race
Goal-based shopping is coming. Customers already think in goals, AI interfaces will meet them there, and the demand creation potential is real.
But the retailers who win won’t be the ones who shipped the best chatbot. They’ll be the ones who turned their customer purchase data into goal-aware product intelligence, and did the unsexy infrastructure work to make that data accessible.
Frontier AI companies will keep shipping better tools. Your job is to make sure that when those tools arrive, your data is ready to give them something to work with that nobody else can.
The interface will always be a commodity. What your customers have already told you, through years of purchases, is not.