Machine Commerce 00The note
Research page
Shopping request · journey ep-db6398bc69f1

Could you find a polo shirt in blue, in size L as a present for my brother? I need it in the next 8 days, so standard shipping works for me.

CartTrace-100k

101,416 trips to the store.

We sent an AI agent shopping, all the way to the till, and wrote down every step. Then we counted where it got lost.

An open dataset from General Machines, explored by Machine Commerce.

01

The shopper is a program. It keeps some of what it wants to itself.

Like a person, it doesn't say everything up front. It answers when the agent asks, and never makes a detail up.

Shopper
Agent
02

1,044 stores that don't exist.

No agent may pay on a real store. So we screened 2,339 real ones, learned how each kind of shop sells, and built stand-ins that sell the same way. Their shelves come from open product catalogs. Prices are made up.

Each square is a stand-in store, coloured by its kind of shop.

03

One trip, from the request to the receipt.

Every search, every page read, every question, the cart, the checkout and the payment, through the protocols agents use to buy today. Test tokens only.

04

Send a hundred. Watch where they fall off.

Every loss happens before the till.

05

The till was never the problem.

On journeys that went on to pay, almost every failed check happened before the cart. Building it, checking out and paying: under one percent each, on every protocol.

06

Asking helps where the shopper held something back.

With two needs held back, the agent ended right 62% of the time when it asked and 12% when it didn't. Below: the same task compared with itself.

+12points of success
−200+20

07

Sometimes it said nothing fits, when something did.

The shopper leaves. The store never learns why.

08

Ask the same thing again.

Of 24,549 tasks run three times or more, almost a third both succeeded and failed. Which store it was explains under 2% of the outcome. One test doesn't tell you much.

29.3%of tasks run 3+ times both succeeded and failed
1.9%of the variance in success is the store
09

A shirt needs a size. A drill needs a voltage. A gift needs a guess.

Twenty-five kinds of shop, about four thousand trips each. The hardest and the easiest:

All 25, with a journey from each →
Priority · machine readableNo. CT-100K
FromGeneral MachinesPublished by the lab, explored by Machine Commerce
ToWhoever builds the next agentOr sells to one.

© 2026 General Machines Pte. Ltd. · The research page · machinecommerce.co