Thought Leadership · Grocery eCommerce
Grocery retail is having a moment of AI confidence. Executives are confirming that AI-generated product descriptions are live and working. AI-powered conversational search is showing up in production, with sponsored placements riding along inside it. Competitors are publishing detailed field guides about the billions of tokens they’ve pushed through internal AI pipelines.
The read on all of this, inside a lot of retail organizations right now, is simple: AI solved the product data problem. If a model can write a compelling description, understand a query, or summarize a catalog, surely it can also organize one.
That’s a category error. Generating language about a product and understanding a product’s place in a catalog are two different problems, and only one of them gets solved by pointing an LLM at your data.
Substitution isn’t semantic similarity. It’s ranked, category-specific judgment. A viable diaper substitute has to match on size first, then brand and sub-brand, then situational factors like day-vs-overnight use. A viable coffee substitute ranks caffeine content first, then format, then roast. An LLM asked to find a substitute reasons from general world knowledge about diapers or coffee. It doesn’t know that this retailer’s shoppers make the decision in this order, because that ordering isn’t written down anywhere the model can read it. It’s a deliberate structure, built category by category, not an emergent property of language.
Quantity equivalence requires structure, not label-reading. Knowing that two half-gallons is a legitimate substitute for one gallon depends on a real Number-plus-Unit model with normalization behind it. An LLM guessing from whatever phrasing happens to be on the package will get it right until the packaging gets weird, “2 Squeeze Tubes” instead of “12 fl oz” and then it won’t, silently, with no error to catch.
Synonyms have a correctness rule that has to be enforced, not inferred. A valid synonym has to apply to every product carrying that attribute. “Non-Dairy” is not a valid synonym for “Lactose Free,” because not every lactose-free product is non-dairy. That’s a narrow, deliberate rule. An LLM will get it right on some products and wrong on others confidently, and inconsistently, at catalog scale.
Category placement is a judgment framework, not a classification task. Deciding that an almond-cashew milk blend deserves its own category, linked to, but distinct from, plain almond milk and plain cashew milk, comes from asking a specific sequence of questions about use case, distinctiveness, and shared attributes. An AI model can approximate this from what it’s read on the internet. It can’t know your substitution philosophy, applied consistently across ten thousand categories, because that philosophy doesn’t exist anywhere in its training data. It exists in a decision framework someone built.
The honest test isn’t “structured data vs. no data.” It’s “AI with a real structural foundation vs. AI without one.” We’ve run that test. In a head-to-head against a leading AI-based recommendations engine, a recommendations product built on standardized, structured product data drove a 63% increase in daily add-to-cart revenue and a 14% lift in add-to-cart conversion. Same AI-era expectations. Different foundation underneath.
That’s the gap most catalogs are sitting in right now. Not a lack of AI, but a lack of the layer AI needs in order to reason correctly about what it’s looking at.
AI is very good at producing language that sounds authoritative about a product. It is not, on its own, good at knowing that this diaper and that diaper are actually interchangeable, or that this synonym silently breaks for a third of the catalog. Those are structural questions, and structural questions get answered by structure, a taxonomy and attribute model built deliberately, maintained deliberately, and gotten right before AI is asked to build anything on top of it.
We built ours on a simple idea: get the product data right, and everything else, including AI, gets easier to build. Not instead of AI. Underneath it.