Journal / Kirill Naumov
AI Product Discovery in Retail Starts with a Trustworthy Catalogue
Before adding conversational shopping, make product attributes, variants, availability and the path to checkout accurate and connected.
By Kirill Naumov ·
A customer asks an AI assistant to find a coat for a wet commute, a charger for a particular device or a sofa that fits a narrow room. The answer may sound polished. The useful test is less glamorous: is the product real, does it meet the stated need, is it available in that location and can the customer actually buy it?
First, make the catalogue say what the product is
Product descriptions are often written for several different jobs: brand storytelling, internal merchandising, marketplace listings and store colleagues answering questions. A conversational interface needs clear, consistent facts: dimensions, materials, compatibility, care, included parts, variant differences, price and current availability. “Premium quality” is an opinion. “Fits a 60 cm cabinet opening” is a fact a customer can act on—if it is accurate and belongs to that exact item.
Google’s [Merchant Center product-data specification](https://support.google.com/merchants/answer/7052112?hl=en) calls for accurate, correctly formatted product information and matching availability across the feed, landing page, checkout and structured data. That guidance is about listings, not a guarantee of inclusion in AI answers; the same operational lesson applies to any shopping experience that retrieves catalogue information.
Connect variants to the right answer
A parent product with six sizes and three colours is not one interchangeable item. Keep the relationship between the product and its variants explicit. If a customer asks for a black, size-eight version, the answer should not borrow the image or stock status of a different variant. Likewise, compatibility claims need their own source: a phone case fitting one generation does not become universal because a language model writes a confident sentence about it.
Start with a category where attributes already exist and customers repeatedly need help comparing options. Agree which fields are authoritative, who owns corrections and how frequently changing facts—especially price and stock—are refreshed. If a required attribute is missing, the assistant should say what it cannot confirm, ask a useful follow-up or hand the question to a colleague. It should not fill the gap with a plausible guess.
Make the next step honest
A recommendation is only as useful as its route to action. Show the right item and variant, the source of any important claim, the relevant delivery or collection information and a working path to the product page. Do not imply that a displayed offer is available in every store or that a product can arrive tomorrow unless the live system confirms it. Keep the ordinary search and browse journey available; customers should not have to chat with a bot to find a basic product fact.
Test the whole shopping task
Build a test set from real customer questions, but remove personal details. Include vague requests, misspellings, competing requirements, variant questions, out-of-stock products and questions the catalogue cannot answer. Have a knowledgeable colleague check whether each response uses the correct item, facts, location and next step. Add a red-team set of tempting but unsupported questions such as “Will this fit my model?” where the answer is not in the approved data.
Track more than how often the system returns a result. Measure correct product-and-variant matches, successful product-page visits, add-to-cart and purchase completion, unhelpful answers, handoffs, corrections and returns linked to misunderstood details. Compare with the existing search or assisted-selling experience. A higher click rate is not a win if it comes with more wrong-item returns or extra work for store teams.
If the test finds gaps, improve the catalogue or retrieval rules before adding another layer of AI. Sometimes the strongest “AI transformation” is a tidy product record, a current stock feed and a clear ownership process. Less cinematic than a talking shopping assistant; much more likely to help the customer.
For practical applications across retail, selling, staff learning and loyalty, explore the [AI for Retail workshop](https://www.kirillnaumovai.com/workshops/ai-retail-november).