Hi everyone, Blake here, writing from the far side of a busy stretch of travel that began at FMI’s GroceryLab, where we had the privilege of co-leading the lab on Intelligent Merchandising and Assortment Planning.

It was one of the better events I've attended this year. The people were serious, the questions were sharp, and the appetite for using AI to build smarter shelves felt real in a way it didn't twelve months ago. But I kept noticing the same reflex: every time the conversation turned to how a grocer should actually know what to put on the shelf, it went straight to POS data. To the register.

I understand why. The register is the most honest number a grocer owns. But it only records what has already happened, and the more I sat with that, the more it struck me that we are looking for tomorrow's demand in the one instrument that can only describe yesterday's.

Stick around for a Featured Insight at the end, where our friends at ProdX make the point underneath everything else in this edition: none of it works if the data isn't honest.

🔭 What the Register Was Never Built to See
💬 Featured Insight: ProdX

001 // Defensive Data

For most of this business's history, the register was the best demand instrument a grocer had, for the simple reason that it was the only one. It is a faithful account of every transaction that closed, and there is real comfort in that. But a closed transaction is the end of a story whose most interesting chapters happened earlier, out of view. The register can't tell you about the shopper who came in looking for the organic line you've never carried and took that demand down the road. It says nothing about the basket someone assembled and abandoned at the final step, much less why. It has no record of the search that returned nothing, or of the substitute a customer accepted grudgingly versus the one they refused outright before walking out. The register sees the sale, and it is blind to all the demand that gathered at its edges and then quietly dispersed.

That blind spot is what I think of as offensive data, as opposed to the defensive kind the register collects. It isn't an account of what you captured but a map of what you haven't captured yet, and the strange part is that most regional grocers are sitting on far more of it than they realize. They generate it every hour they're open, and then, with no place to put it, let it evaporate.

002 // The Demand You Don't Record

The digital store is where that offensive data lives, because nearly every interaction a shopper has with it is a small, legible statement of intent, a sentence spoken directly to you that the POS was never built to hear.

Consider substitution. When a first choice is unavailable and a shopper accepts or rejects what you offer in its place, they are running a loyalty experiment on your behalf and handing you the result for free. The top handful of substitutes for any item absorb most of its transferable demand, so the pattern of accepts and rejects is a map of where demand travels when supply moves beneath it. When a national-brand yogurt runs short and most shoppers reach for the private label, the shelf is telling you to grow those private-label facings and set them close. When a specialty gluten-free pasta runs short and nobody will accept a replacement, the shelf is telling you something more valuable: this is a destination item, to be protected and never quietly swapped, because the shopper who came for it will leave the store before taking a stand-in.

The other signals work the same way once you learn to read them. The moment a shopper abandons a cart is a confession of the criteria they were judging you against: one who leaves after studying a nutrition panel is naming an attribute they won't compromise on, one who leaves after toggling pack sizes is revealing a sensitivity measured in cents per unit, and one who leaves the instant the total crosses a threshold is showing you the budget ceiling for the whole trip. The discipline is telling merchandising friction apart from fulfillment friction, since a shopper who gives up at the delivery-window step is describing an operations problem, not a shelf problem, and the two call for different responses. A search that returns nothing is the most direct request a shopper can make, a demand spelled out in full and attached to nothing you sell; when a term you've never stocked starts spiking across a handful of stores, that isn't a forecast to validate, it's an instruction, localized to the aisles where it matters. And a steady rise in clicks on attributes like organic, local, high-protein, or under a certain price is one of the earliest reads you'll get on how a neighborhood is changing, weeks before any of it reaches the register.

None of this is engagement in the soft sense the word usually carries. Every one of these signals resolves into a physical decision about facings, endcaps, safety stock, and where the private label sits in relation to the brand it means to challenge. The digital store isn't a channel that happens to sit beside the real one. It is one of the most sensitive merchandising instruments a grocer owns, and most of us have been reading it like a billboard when we should be reading it like a seismograph.

003 // Own the Agent, or Fill the Order

Everything so far assumes a human seated at a screen, generating all of this signal by hand, one click at a time. That world is ending, and what's replacing it raises the stakes in both directions at once.

When the shopper becomes an agent, when a person stops tapping through filters and just tells an assistant to find the high-protein, gluten-free snacks under ten dollars that their store carries, the signal doesn't vanish. If anything it grows richer than the digital store could make it on its own. A compound intent you once had to reconstruct from a trail of separate clicks now arrives whole, in a single sentence, with the shopper's priorities already ranked inside it. A substitution decision no longer has to be inferred after the fact, because it comes with its reasoning attached, and "rejected, not like-for-like," "accepted, same brand, larger pack," and "rejected, not eligible for my benefits" are three distinct facts about how demand behaves. Where you once saw only that a swap had failed, you now see why. Inference becomes something closer to a transcript.

That's the encouraging version, and it's better than anything the register era could offer. But it rests on one condition: that the demand keeps passing through you.

The other version is less comfortable. If your store is nothing more than an endpoint that takes orders, the agent will do all of its reasoning somewhere else, on a surface you don't own and can't see into, and hand you back a clean, settled basket with everything interesting about it stripped away. The agent watched the shopper weigh the alternatives, hesitate over price, abandon one option for another, and decide, well upstream of all that, which grocer to trust for the category in the first place. What you got was a transaction. That's the register all over again, rebuilt one level higher, and it's more dangerous than the original, because this time the signal isn't going unrecorded for lack of a place to put it. It's being captured in full, by whoever owns the agent, and quietly kept.

We spent some time in April on owning the agent and renting the surface, on Walmart drawing Sparky's logic back onto its own rails while letting the larger platforms supply nothing but reach. The merchandising consequence is that same principle seen from the shelf rather than the boardroom. Across the register, the digital store, and now the agent, the only question that has ever mattered is who holds the demand signal, and the agentic era is about to answer it in a way that will be hard to walk back. A store that only takes orders was, until recently, leaving money on the table. In the world arriving now, it is handing its demand signal to a competitor and mistaking the transaction for convenience.

004 // It All Rests on the Data

All of this leads to a question less exciting than the ones the headlines prefer, and a lot more important. None of what I've described, not the shelf inference, not the agent reasoning over your assortment, not your ability to be recommended when a shopper asks an answer engine which grocer to trust, is possible unless the data underneath it is structured, complete, and honest. An agent reasoning over a catalog of mislabeled attributes doesn't fail loud enough to draw notice. It just recommends the wrong thing to the wrong shopper, with total confidence, and moves on.

The grocers who still own their demand signal a few years from now won't be the ones with the most ambitious slide about AI. They'll be the ones who did the unglamorous work of getting the foundation right first, which is exactly what our friends at ProdX have been writing about this month.

Walmart's Sparky AI shopping agent is now live at scale — weekly active users up over 100% in a single quarter. The new capabilities sound impressive: personalized replenishment, meal planning, in-store use, Spanish-language support. Headlines called it a watershed moment for AI in retail.

What the headlines didn't say: Sparky's meal planning feature works by matching shoppers to recipes and then surfacing the right ingredients. That matching logic runs on product attributes — ingredients, allergens, dietary flags, nutritional data. If those attributes are wrong, incomplete, or inconsistently structured, Sparky doesn't recommend the wrong product smoothly. It just fails quietly.

This is the part of the AI conversation that doesn't make the press release. Every agentic commerce layer — whether it's Sparky, Loblaw's ChatGPT-powered PC Express, or Kroger's Sage employee assistant — sits on top of a product data foundation that most retailers built in a different era, for a different purpose. That foundation was never designed to power AI. It was designed to power a shelf tag.

Meanwhile, Walmart just confirmed a three-stack tech consolidation: Walmart US, Sam's Club, and Walmart International are merging into one global platform. A single unified product data standard must now work across three separate technology architectures, simultaneously. The stakes for getting the foundation right have never been higher — or more visible to the people in charge.

The pattern repeats everywhere you look this month. Kroger's new Chief Data and AI Officer. Loblaw Digital's five consecutive weeks of AI buildout. Northeast Grocery's CIO presenting at GroceryTech on agentic AI for merchant analysis. All of them betting on a layer of intelligence that can only be as good as the data feeding it.

So here's the question worth sitting with: when your organization demos its AI initiative to the board this quarter, is the confidence in the model — or in the data the model is running on? There's a difference, and it shows up in production.

Read the full brief →