Short answer: We gave personal AI shopping agents real shopping tasks on our demo ski and snowboard stores: find gear that fits a specific person, stay on budget, use any discount, put it in the cart. Shopping on their own, the agents got the cart wrong more often than not: bindings in the wrong size, snowboard bindings for skis, a regular board for feet that need a wide one. Every time an agent asked the store's digital twin, the cart was right. The catch: agents don't go looking for a store's expert. It has to sit where they make the choice. Watch two of the runs.
The setup
More and more people send an AI agent to shop for them. These agents open a store's website and work through it on the shopper's behalf: they read the product pages, pick a size and fill the cart. We wanted to know what happens when the purchase needs advice, the kind a good salesperson gives in the shop.
So we ran shopping tasks where getting it right depends on knowing the gear:
- Bindings for a ski: the binding's brake has to match the ski's width. Too narrow doesn't fit, too wide catches the snow.
- A snowboard for big feet: with US size 11 boots and up, a regular board is too narrow and your toes and heels drag in turns. You need a wide one.
- Sizes that depend on the person: binding size from boot size, ski and board length from height, weight and ability.
- A discount that only applies to certain combinations.
The tasks read like what a shopper would actually tell an agent: "Advanced skier, 5'7", 150 lbs, mostly groomers, under $850. Buy skis and bindings that fit, use any discount." We never told the agents that the store has an expert they could ask.
We tested two kinds of agents, because they work differently. One browses like a person: it looks at the page, clicks what it sees and reads what's on screen. The other reads the page's code directly.
What the agents got wrong on their own
When an agent shopped without asking the store's digital twin, the cart was wrong more often than not. Three examples:
- Bindings in the wrong size. The agent picked medium snowboard bindings for boots that need a small.
- Snowboard bindings for skis. Asked for a ski setup, the agent put a ski in the cart, paired it with snowboard bindings, and missed the store's discount on top.
- A regular board for big feet. For an advanced rider in US size 11 boots, the agent chose a regular-width board. It's a good board, for someone else.
None of these are exotic mistakes. They're the questions a shopper would ask a salesperson, and an agent reading product pages doesn't know to ask them. Each one ends the same way: a return, or a customer who keeps gear that doesn't fit and blames the store.
What happened when they asked the twin
sqrly gives a store a digital twin: an AI sales expert that knows every product, size, stock level and current offer, plus what the merchant knows about fit and what works with what. When an agent asked it, it got the advice a good salesperson would give:
- "Your 11.5 boots need the wide board so your toes and heels don't drag in turns." The twin picked the Ridgeline in 160W, the right length for the rider's weight, and bindings in large.
- For a 76 mm ski, it picked the binding with the matching 80 mm brake and said why: the brake has to match the ski's width.
- It found the store's discount and said which item it applies to.
Every time an agent asked the twin, the cart was right. Both kinds of agents, skis and snowboards, different shoppers. In the snowboard run, the agent asked the twin to put the setup in the cart, and the twin did it, in the sizes it had recommended.
Agents don't go looking for the store's expert
This was the most useful finding. The twin only helps if the agent asks it, and agents don't go looking for help.
The agent that browses like a person never once opened the "Ask our expert" button in the corner of the page. It walked right past it. What changed its behavior was putting the expert in its path: a short box on the product page, between the options and the add-to-cart button, with a heading and a button. With that box on the page, the agent asked the twin every time. A single line of text with a link worked most of the time, but when the agent felt confident, it read past it.
The agent that reads the page's code found the store's "Ask our expert" page on its own, even with the box switched off. It doesn't need to see a button; it needs a page it can read.
What this means for a store
- Put the expert where the choice happens. On product pages, between the options and the buy button. Agents act at the moment of choice, so that's where the help has to be.
- Make it work for both kinds of agents. A visible box for agents that look at the page, a plain answer page on your own domain for agents that read the code.
- A small box is enough. It doesn't have to dominate the page. A heading and a button did the job.
- Watch what agents do. In sqrly's Live view, a merchant sees each agent's visit as it happens: every page, the question it asks the twin, the answer, and what goes into the cart.
Watch it happen
We recorded two runs on our demo store, one with each kind of agent, with the store owner's Live view next to the agent: sqrly.ai/demo.
sqrly is live on Shopify today, with more commerce platforms coming soon. If you run a store where advice sells and want your own digital twin, email us about becoming a pilot shop.