Ten Ways to See the Shelf, and Still No Definitive Winner

The most consequential operating data in a store is still largely collected by hand. Associates walk the aisles and check stock positions intermittently through the day, and everything downstream, replenishment, ordering, online pick accuracy, promotion execution, runs on however current that last manual pass happens to be.

The goal itself is easy to state. Be on shelf as close to one hundred percent of the time as possible, for every item in every store, and know the stock position of every item in every store with something close to the same certainty. Measured against that standard, intermittent manual checking is not a data collection method. It is a sampling exercise, and the industry has spent the past several years building technology to replace it.

What the industry has produced is options. Fixed camera systems built for shelf monitoring. Existing loss prevention cameras repurposed for inventory vision. Autonomous mobile robots that live in the store, and portable versions that move between them. Smartphone-based capture in the hands of associates and other CPG stakeholders. RFID in fixed infrastructure, in handheld form, and mounted on robots. Camera units riding on shopping carts and stocking carts. Devices worn on associate lanyards. Nearly all of them run on broadly similar computer vision with AI engines underneath, and every one of them can be found working reasonably well somewhere in retail today.

Every Option Works, and Every Option Has its Challenges

What none of them has done is definitively win. After years of deployments and pilots across many formats, no single approach has separated itself from the field with a clearly differentiated value proposition, and the reason is not immaturity or lack of functionality. Each approach fails somewhere on a dimension that matters operationally. Coverage completeness. Capture frequency and recency. The ability to see depth of stock or behind the facing, or height above eye level or top stock. Data quality. Cost to deploy and operate, Complexity of enviironment or operation. Whether associates adopt it or work around it. Whether shoppers accept it or step around it. Whether capture happens at standardized times and locations under store control, or whenever the collection method happens to pass by. Whether the output interoperates well with everything else the retailer runs and data it has available.

The clearest demonstration sits inside a single company. Walmart ended its shelf-scanning robot program in U.S. stores in late 2020, a decision the Wall Street Journal reported came after the retailer found that associates already walking the aisles to pick online orders could produce similar results. Sam’s Club, its club division, went the opposite direction and completed an evalutation of inventory scanning camera towers mounted on its already deployed autonomous floor scrubbers by October 2022, according to the company’s announcements at the time. One parent, two formats, two different answers, and both defensible on their own operating math. The split runs across the field as well: Retail Dive has reported Schnuck Markets and Hy-Vee deploying aisle-scanning robots, while CIO Dive reported Lunds & Byerlys choosing handheld capture in associates’ hands instead. Conversations on Retail covered the newest turn in this story earlier this week, as Instacart placed a version of Walmart’s 2020 bet on the person already in the aisle.

The holes are structural, which is why waiting for a winner is very likely to be a losing strategy. A ceiling camera cannot see behind the front facing or the bottom shelf. A robot sees an aisle only as often as it passes it. RFID performance is inseparable from tagging economics, which differ by category. Phone-based capture is only as complete as the labor plan behind it. These are not engineering gaps a next release will close. They are consequences of physics, store design, and labor models, which means they resolve differently in a supercenter, a small-format grocery store, a club, and a specialty box, and differently again between apparel, center store, fresh, and general merchandise.

Creating a Mosaic Is the Answer Most Evaluations Are Not Built to Find

The uncomfortable conclusion is that the right answer is plural and multidimensional. The deployment that actually reaches the goal is a personalized mosaic of these approaches, assembled against a specific retailer’s formats, departments, stakeholder needs, and desired outcomes. High-value or high-velocity categories may justify fixed infrastructure. Broad general merchandise coverage may be served well by periodic robot passes. Categories where tagging already pays for itself invite RFID. Capture that rides on work already happening, on carts, on lanyards, on the devices associates already carry, can fill gaps the dedicated systems leave. The design question is which combination produces decision-grade visibility where the economics demand it, at a cost the store can adopt.

Most technology evaluations are simply NOT built to reach that conclusion. They are built to compare vendors and technologies against each other and pick one, which quietly assumes the category has a winner. In this category, the comparison that matters is between the retailer’s conditions and the options tested, and that evaluation only works if the retailer has done its design homework first.

Design Questions must Come Before Vendor or Technology Questions

That homework is a short list of hard questions. Which decisions will this visibility actually drive, and how fresh does the data have to be for each of them, because an overnight signal that feeds tomorrow’s replenishment order is a different requirement from a two-hour signal that feeds a picking path. Which departments and formats carry enough economic weight, in sales, shrink, or labor, to justify the denser and more expensive forms of coverage. Who acts on the signal when it arrives, and does that workflow exist yet. How will data from multiple capture methods reconcile into one stock position a buyer or a system can trust. And what can the labor model absorb, because a capture method the store team will not sustain may produce worse data than the manual process it replaced.

The readiness question has flipped. For years the fair question was whether this technology was ready for the store. The technology is ready; the open question is whether a given retailer is ready for it, with the operating design to turn visibility into action. For the CPG teams on the other side of the shelf, the same shift is worth watching, because as retailers assemble these systems, availability conversations with suppliers will increasingly run on store-level, near-real-time data instead of warehouse shipments, POS data or audits.

Finally, shelf visibility will not be the last category that faces this challenge. Opportunities to optimize for RFID, Retail Media and trade promotion execution present similarly shaped sets of problems, and a crowded field of workable solutions currently offering NO single definitive winner. The pathway to excellence for each of these is to broadly pull back the lens and design with the end in mind, leading with the retailers desired outcomes and actions and working backwards to define the appropriate operational plan for your organization and its supporting stakeholders. If done correctly the mosaic of value these capabilities offer retailers and their partners can deliver transformative impact across the value chain for the retailers operating ecosystem!

View Profile

Emil Martinez

Emil Martinez is an advisor on emerging retail technologies and co-host of Retail Ready on Conversations On Retail. His three decades in retail technology span the measurement world of ACNielsen, IRI, NPD, and the standards work at GS1, more than a decade leading Tata Consultancy Services' global retail consulting practice, and most recently serving as Chief Executive Officer of Badger Technologies.

He has sat on the supplier side, the retailer side, and the data side of some of the most demanding relationships in retail, and he has spent years turning AI and computer vision into results on the store floor. His work centers on a single question: whether a technology is real, whether its data is worth a decision, and whether the operation is ready to run it.