Prodx · Retail Grocery Technology Issue 04 · September 2026 · What runs beneath everything in grocery retail
The story this month isn’t that another grocer launched an AI assistant. Several did. The story is that agentic surfaces started appearing on retailers’ catalogs whether or not the retailer built them.
Start with the version that went well. Schnucks launched an agentic assistant co-developed with VitalityIP, trained on more than six billion lines of shopping, health, and nutrition data, and grounded in actual Schnucks store inventory. Their COO drew the distinction plainly: “it’s not a search bar.” It reasons over aisles, promotions, dietary preferences, and what’s already in your pantry.
Now the version nobody announced. Shipt went live inside both ChatGPT and Claude this month. Which means any grocer using Shipt for delivery now has a conversational shopping surface running on its catalog — one it didn’t provision, doesn’t control, and can’t test.
The clearest signal of where this goes came from outside grocery entirely. When Square opened its own agentic integrations, the selling point was that eligible sellers went live immediately — automatically enrolled, nothing to build, nothing to configure, no fee. Catalogs entering agentic surfaces by default, as a platform decision rather than a merchant one.
“You can’t QA an agent you don’t own. You can only control what it reads.”
The part worth noticing in the Schnucks announcement is a sentence that reads like legal boilerplate and isn’t. Schnucks stated that it retains ownership of the customer relationship, the shopper data, and the brand experience while using VitalityIP’s AI. That is a company that asked who owns what before it signed. It is not a universal question.
Even the version that went well leaves one question open. The Schnucks announcement is precise about the intelligence layer — the volume of shopping and nutrition data behind it, the grounding in live store inventory — and quiet about the product data underneath. That’s not an oversight, it’s the convention. Assistant launches describe the model, not the catalog. But the teams shipping these will be the first to learn what happens when a shopper asks whether something is gluten-free in a category where that attribute was never required, and the assistant answers anyway. The intelligence layer is new. The attributes it reasons over usually aren’t.
Because the surfaces multiply faster than the governance does. Your assistant. Your delivery partner’s assistant. The one your loyalty vendor ships next quarter. The one a platform enrolls you in automatically. Each reads the same catalog, and none of them will tell you when an attribute is missing — they’ll simply answer the shopper anyway.
Which sorts grocers into two groups over the next several quarters. The ones for whom every new agentic surface is a modest integration on top of a catalog that’s already clean, structured, and governed. And the ones for whom every new surface is a fresh opportunity to discover the same data gaps in public, one shopper at a time.
Supply Chain AI
Target announced Proxima on August 13 — a digital twin of the middle-mile inventory positioning system that moves goods from distribution centers toward stores. The sentence to sit with is Target’s own description: the twin runs on the same data and logic as the live inventory platform. Which means it inherits the live platform’s blind spots exactly. A 63-item fresh food pilot lifted on-shelf availability by 2.5%, and Proxima modeled flow through the new Houston receive center at roughly 98% accuracy. Fresh is the hard case — lead times and expiration dates have to be right before anything downstream can be. Note what’s being asked of product data here: not names and images for a shopper, but weights, dimensions, pack configurations, and dates, accurate enough to simulate the physical world. A twin built on approximate data doesn’t fail loudly. It produces a confident simulation of a warehouse that doesn’t exist.
Leadership Signal
Nate Faust joins Kroger as EVP and Chief eCommerce Officer on September 1 — Jet.com co-founder, later SVP of Walmart US eCommerce Supply Chain. The résumé is the part everyone will cover. The language is the part worth reading twice. Greg Foran described the standard Kroger is holding itself to as “getting exactly what the customer ordered.” Faust’s own first public remark named loyalty data as the asset he’s arriving for. Neither man led with an app, a channel, or a growth number. Both described an eCommerce mandate in terms of item-level accuracy and data. When the CEO and the incoming digital chief independently frame the job that way, the operating priorities tend to follow within two quarters.
Multi-Surface
Hy-Vee confirmed on August 18 that roughly 265,000 products are structured across a stack that now spans three distinct purposes: fulfillment and 30-minute delivery, benefit eligibility resolved at the point of sale, and personalization driving substitutions, recommendations, and predictions. Three systems, three different questions, one catalog underneath. That’s the part worth generalizing beyond Hy-Vee: benefit eligibility is a compliance-adjacent read of product data, fulfillment is an operational read, and personalization is a discovery read — and they don’t fail independently. An attribute gap doesn’t pick one surface to break. It shows up at checkout, in the substitution, and in the delivery window simultaneously.
Disclosure: Prodx is the product data layer in that stack.
Retail Media
Circle K relaunched Full Circle Media across 6,600-plus US stores this month — roughly 2.9 billion annual customer visits, spanning fuel pump screens, in-store displays, checkout, the app, Inner Circle loyalty, and offsite digital. Retail media gets discussed as an advertising business and budgeted as one, which obscures what it mechanically is: a system that serves impressions against product classifications. Every misfiled SKU is an ad served to the wrong shopper in the wrong context, and the advertiser pays for it either way. At c-store scale, across five surfaces, categorization accuracy stops being a merchandising hygiene issue and becomes a margin input on a business line the CFO is watching closely.
Worth Your Time
Three things pointed the same direction this month, and none of them were reported as related. Schnucks’ new assistant offers nutritional guidance to shoppers managing chronic conditions and food allergies. Kroger Health’s president keynoted the inaugural GLP-1 Shopper Summit in Chicago on the theme of food as medicine. And Loblaw launched a virtual weight management program with Launchit Solutions across seven Canadian provinces.
Read together, they describe a category shift that hasn’t been named yet. Dietary flags, allergen data, and nutritional attributes were built as merchandising metadata — useful for filtering, occasionally for a badge on a product page. They are now becoming inputs to health guidance delivered by an AI to a shopper who is managing a real condition. The accuracy bar for those two jobs is not the same bar, and most catalogs were populated for the first one. Nobody has said out loud what the second one requires. It’s worth being early to that conversation rather than late.