Why AI Doesn’t Know What’s on the Shelf

Ask an AI where to buy something today and it often points you at a store that doesn't have it. The model knows the catalog. It doesn't know the shelf.

That gap is the reason AI local inventory is still a hard problem. It isn't a reasoning problem. It's a data problem, and a bigger model won't fix it.

Two kinds of facts

A catalog fact says a store sells a kind of thing. A hardware store sells batteries. A pharmacy sells bandages. These facts change slowly, and they're everywhere on the public web.

A shelf fact says a store has this exact item, at this price, right now. These facts change by the hour. A pack sells out. A price gets corrected. A delivery lands in the back room.

Shoppers don't care about the first kind. They care about the second. "Do they have it?" is a shelf question.

When an AI answers a shopping question from what it learned in training, it can only give you catalog facts. It says "this chain usually carries that." It can't say "that door has two left."

Why training data can't fix this

A language model learns from text collected at some point in the past. Even a model with live web search sees pages, not shelves. Most stores don't publish their shelf state as text, and what they do publish ages the moment it's written.

Three things make shelf state hard to learn:

It changes constantly. Anything frozen into a model is already old when you ask.
It's local. The answer depends on which store, not which brand.
It has a source. A count from a store's own register and a line on a public page are not equally trustworthy, and a model can't tell them apart from a sentence.

So the model guesses. A good guess sounds confident and reads like an answer. That's the dangerous part. The shopper drives across town, or the agent places an order, and the item isn't there.

What AI local inventory needs

If you want an agent to be right about local stock, the data has to carry more than a yes or no. Each fact needs:

The product, identified the same way no matter who lists it.
The store that has it.
The price and stock at that store.
The source, meaning where we learned it.
When we last saw it, so age is visible rather than hidden.

That last one matters most. A reading with a timestamp is something you can reason about. A bare "in stock" is not. Every stock claim has an age, whether anyone shows it or not.

We don't claim stock is real time. We keep the source and the last-seen time on every store offer, so the age of every reading stays visible.

How JENNi supplies shelf facts

JENNi builds the SKU Graph, one shared map that joins canonical products, anchored on barcodes, to the stores that carry them. Each link is a store offer with price, stock, source and when we last saw it. It also carries how sure we are right now.

Agents reach that graph through one interface, our MCP server at https://shop.jennipro.com/mcp. MCP is the Model Context Protocol, an open way for an agent to call tools. Any agent can connect.

The flow is short:

search_catalog takes a query and a 5-digit US ZIP, and lists local delivery first, then fills with ship-to-home. Search shows in-stock items only.
get_buying_options takes a product_id, a zip, a quantity and a fulfillment_mode. It returns availability, price, delivery fee, delivery promise and the total before tax.
start_checkout hands the buyer a hosted page that re-verifies current price and availability before anything is charged.

The agent doesn't have to guess which door has the item. It asks, and gets an answer built from shelf facts with a source and an age.

When an order is delivered, that's evidence. When one fails, that's evidence too. Both feed back into the graph, so the next answer is a little better than the last.

What this changes for agents

Without shelf facts, an agent shopping locally has two bad options. It can sound sure and be wrong, or it can hedge so much that it's no use. With them, it can do the job: find a product, check what's possible at this ZIP, show the total, and let the shopper approve it.

JENNi is the seller on these orders. We buy the item from the store for the customer, so if something goes wrong, the issue comes to us, not to the agent builder.

Try it

If you build agents, connect to https://shop.jennipro.com/mcp and run a search with a ZIP you know. Look at what comes back, and notice that every answer was built to be checked.

If you run shelves and want them visible to AI agents, get in touch at jennipro.com.