Most of the information that will influence a foodservice purchase is not on the label. It is in the digital record around the product.
By John Wheeler
I recently reviewed a presentation from Mars and K10X called Beyond the Label. It was delivered at a workshop held by W3C and GS1 about e-commerce for people and AI agents.
The presentation made a point that should matter to every food manufacturer: the physical label is no longer enough to represent a product.
A label remains essential. It tells us the product name, ingredients, allergens, nutrition facts, net weight, storage instructions and other required information. But the label can hold only so much. It was designed mainly for identification, safety and compliance. It was not designed to answer every question a buyer, distributor, restaurant operator or AI system may ask.
In foodservice, most of the information that affects a buying decision is not printed on the package. It is off-label data: the larger digital record that surrounds the product.
That may become one of the biggest changes in how foodservice products are found, compared and purchased.
Mars started with a simple product identity problem
Most packaged products have a GTIN, the global number represented by a UPC or another barcode. That number is very useful, but Mars pointed out that it may not always tell us the exact version of the product in front of us.
A product can be reformulated. An ingredient may come from a different source. Packaging may change for a promotion. The same brand may sell different versions in different countries. Old and new inventory can exist in the supply chain at the same time. In some cases, those differences can affect ingredients, allergen warnings, certifications, artwork or other important facts.
Mars and K10X built a proof of concept in which a QR code connects to a product record controlled by the brand. A shopper scanning the code can see a normal webpage. A software application or AI agent can receive the same information as structured data that a computer can read directly.
The most important part of the design is not the QR code. It is what the system does when it cannot identify the exact product version.
It refuses to guess.
If the system knows the GTIN and the exact product variant, it returns one record. If it knows the GTIN but several versions could apply, it presents the possible versions instead of pretending they are all the same. If someone requests a version that does not exist, the system returns no answer rather than selecting the closest match.
That is an important principle for AI: uncertainty should be visible. When an answer could affect allergens, certifications, food safety or customer trust, a confident guess is not good enough.
The label is only the beginning in foodservice
The Mars presentation focused heavily on consumer product information, but the idea may be even more important in foodservice.
A restaurant operator does not choose a product only because the ingredient statement and nutrition panel look acceptable. The operator needs to know whether that product works inside a real kitchen and a real menu.
How many servings come from a case? How much labor does preparation require? Does it need special equipment? How long does it take to prepare? What is the cost per finished serving? How much waste should the operator expect? Can employees prepare it consistently during the busiest part of the day? Does it fit a current menu item, or will the restaurant need to create something new?
The operator may also need to know whether the product is available through the distributor already delivering to that location. The distributor may have its own item number. The product may be stocked in one distribution center but not another. Minimum orders, delivery schedules, substitutions and local availability may matter more than the national product description.
Then there are questions about shelf life, case dimensions, freezer or cooler space, preparation instructions, certifications, country of origin, packaging disposal, photographs, menu descriptions and suggested uses.
Very little of that fits on a label.
Today, much of this information is spread across specification sheets, PDFs, spreadsheets, distributor portals, emails, sales presentations and the personal knowledge of brokers and sales representatives. Some of it is current. Some of it is outdated. Some of it is known by one person but never entered into a system.
That may be manageable when a buyer calls an experienced broker who knows the product. It becomes a serious problem when an AI agent is expected to compare hundreds of products in seconds.
AI will buy from the product record it can see
An AI buying agent cannot inspect a pallet in a warehouse or call a broker from memory. It must work from the data it can reach.
Suppose a restaurant asks an AI system to find a product that is peanut-free, can be prepared in under five minutes, produces at least 40 servings per case, works in a convection oven, has a certain shelf life and is available from the restaurant’s current distributor.
The answer requires far more than the label.
The AI needs a trusted product identity. It needs complete allergen information. It needs preparation and yield data. It needs to understand the relationship between the manufacturer’s item number, the GTIN and the distributor’s item number. It may need current inventory and location data. It also needs to know when each fact was approved and whether it applies to the exact version being sold.
If those facts are missing or trapped in a PDF, the AI may ignore the product, exclude it by mistake or recommend a competing product with better structured data.
That creates a new kind of competition. Manufacturers will still compete on taste, quality, service, price and relationships. But they will also compete on how clearly and reliably their products are represented to machines.
The best product will not always be the product an AI recommends. The recommended product may be the one with the most complete, trusted and usable digital record.
Product data is becoming a distribution asset
Manufacturers have often treated product data as an administrative requirement. Someone completes a distributor form, sends a specification sheet or uploads fields to a data service. Once the form is accepted, the job is considered finished.
That view is becoming outdated.
Product data is becoming part of distribution. It determines whether a product can be found, understood, compared, approved and recommended. It may influence whether the product appears in an operator’s search results or an AI-generated menu suggestion. It may also help a broker explain why a product fits a specific customer.
This changes the purpose of a product information management system, or PIM. A PIM should not be only a large table or a file cabinet for product attributes. It should be the governed source of truth for the product.
That source of truth must connect the manufacturer’s item number to consumer, inner-pack, case and pallet identities. It must distinguish formula and label versions. It must track market differences, effective dates, certifications and customer requirements. It must show where each important fact came from, who approved it and when it was last reviewed.
It must also be able to publish the same trusted facts through several channels. One distributor may require a spreadsheet. Another may use GDSN or Syndigo. A customer may need a portal. An AI agent may need structured data through an application programming interface or a GS1 Digital Link.
The formats may change. The product truth behind them should remain consistent.
Brand authority does not remove the need for governance
The Mars model places the brand owner at the center of the product record. That makes sense because the manufacturer should know its own product better than a retailer, distributor or search engine.
But brand-controlled does not automatically mean correct.
A manufacturer can publish incomplete, outdated or inconsistent information. Someone may update an ingredient statement without updating a sales sheet. A certification can expire. A case configuration can change while an old distributor record remains active. A new formula can enter production while older inventory is still in the channel.
The answer is not simply to publish more data. The answer is governed data.
Every important fact needs an owner. Changes need approval. Effective dates matter. Older versions must remain available when older inventory is still being sold. Systems should identify conflicts instead of silently overwriting them. And when the correct answer cannot be confirmed, the system should stop and ask for review.
That is the same lesson Mars applied to AI product discovery: do not make uncertainty disappear.
What this means for brokers and distributors
Better product data will not make brokers and distributors unnecessary. It will change where they add value.
A broker who understands the operator, menu, territory and distributor network can provide context that a product label never will. But the broker will be more effective when that knowledge is supported by reliable product facts, customer-product matching and measurable results.
Distributors will also need better connections between their item files and the manufacturer’s product record. A distributor’s ordering system may know price and inventory but lack complete preparation, certification or application data. The manufacturer may have those details but lack the distributor’s local availability and customer history.
The valuable answer comes from joining those sources without confusing who owns each fact.
The manufacturer should own the product’s approved identity and specifications. The distributor should own its inventory, pricing and delivery facts. The operator should control its preferences, restrictions and purchasing authority. An AI system can bring those records together, but it should not invent the missing pieces.
The next shelf may be an AI recommendation
For years, manufacturers have worked to win space on the physical shelf, the distributor’s order guide and the restaurant menu. The next important shelf may be the recommendation layer inside an AI system.
Getting onto that shelf will require more than a barcode and a product description.
It will require complete product truth, clear identities, current versions, reliable evidence and data that software can understand. It will also require manufacturers to think about information that has traditionally lived outside the label: operator applications, labor savings, yield, menu fit, channel availability and delivered economics.
Mars’s proof of concept is not yet an adopted industry standard, and its exact technical design may change. But the direction is hard to ignore.
The physical package will remain important. The label will remain important. The GTIN will remain important.
They just will not be enough.
In the future of foodservice buying, the product on the truck and the product in the data system will need to match. If they do not, the manufacturer may lose the sale before a buyer ever sees the case.
Most of the data that will help make that sale is not on the label. It is off-label, in the trusted digital record surrounding the product.
And that record is quickly becoming part of the product itself.
Source: Mars and K10X, “Beyond the Label: Brand-Authoritative Off-Pack Data and Consumer Pack Variants for Accurate AI Product Discovery”, presented at the W3C/GS1 Workshop on E-Commerce for Humans and AI Agents, September 2026.