A retailer who does not exist knocked down our best sales argument

Every small supplier thinks it knows why a retailer should sell its product. So did I.

Rotina Kraft makes 3D printed pieces: keychains, corporate gifts, collectibles and items for optical shops. Part of its sales goes through partner stores, and each one has its own page on the site, with the catalogue, how much the store earns per piece, ordering through WhatsApp and photos to promote the products. For existing partners, it works well.

The problem was the front door. Anyone opening the link for the first time lands on a screen with the logo and a single field: "Enter the password you received". Before seeing a piece, a price or a name, the visitor needs something they do not have yet.

Why rehearse with a customer who does not exist?

Interviewing a retailer is expensive for the retailer. It is time taken away from the counter, in the middle of serving customers, to answer questions that are almost always badly framed the first time.

So I split the roles. An experience designer, which is a Claude session, was in charge of interviewing, mapping the journey, prototyping and testing. The retailer would be another session, with a single rule: answer as a retailer. No touching files, no design suggestions, no helping the designer.

In practice, the app did not let one session open another on its own. The designer then created a helper agent with the same rules to play the retailer. The separation became weaker than I had planned, because both came from the same place. I note this because it changes the weight of what follows.

The retailer in this story is simulated. The goal was not to replace people with AI. It was to rehearse: to reach the conversation with real retailers with better questions and with a prototype that had already taken its first stumbles.

How do you build a simulated retailer?

The persona started from two real Rotina Kraft partners. One is a gift store that displays the pieces on consignment and only settles after selling. The other is an optical shop that buys in volume, resells online and gives pieces away to its own customers.

The designer merged both into one main voice and kept the second as a friend:

• Renata, 46, owner of a gift and promotional products store in Canoas, in southern Brazil. She has one employee and a large share of revenue from corporate gifts. She runs her own Instagram. She has been burned by consignment before: unsold stock, messy settlements, a supplier who vanished for months. She lives on WhatsApp and decides with her wallet.
• Márcia, owner of an optical shop. She buys in volume, resells on Mercado Livre, the largest Latin American marketplace, and gives gifts to customers. Consignment does not work for her.

Canoas was not random. Rotina Kraft's WhatsApp number is from the Porto Alegre area, and the retailer had to be someone who would plausibly receive that link.

What did she say when she opened the link?

In the first round, the designer sent ten questions. How she picks a product for the counter. What she needs to see in the first seconds of a link on her phone. What makes her trust someone she has never heard of. What math she does in her head before saying yes. What would make her give up on the spot.

The answers came with the impatience of someone reading while serving customers (translated from Portuguese):

"If it asks for sign up, password, email before showing me anything, I close it right away."

"How much I pay and how much I sell it for. Both numbers. I don't want to do the math."

"If someone comes with a long speech about 'our story, our mission' and I don't know how much it costs, I lose my patience."

"Computer made photo, all perfect, I get suspicious."

She picks products for three reasons: you want to hold it, it has an impulse price near the register and it fits on the counter. "Every inch has to pay for itself." She also carries a fear specific to 3D printed pieces: looking like a cheap toy.

In the second round, the designer asked her to choose a single question to send to a new supplier. The answer became the title of a section of the page:

"If I take it and don't sell it, what happens? Because the answer tells me everything: whether the supplier picks it up, whether they swap it, whether I take the loss and whether they are serious or will disappear."

Which hypothesis fell?

We believed the strongest argument for consignment was the earnings, shown as a percentage: the store displays, pays nothing and keeps a share of each sale.

The retailer found it too little. In gifts she works with larger margins, and a percentage says nothing to someone who decides with her wallet.

"With [...] I'm practically giving away my shop window."

What mattered was something else: no money at risk and a record both sides can see. Her worst memory of consignment was the count at settlement time.

"It was his notes against mine."

Here came the first surprise. Rotina Kraft already gives partners a dashboard where each sale is logged with date and time, and both sides see the same number. The answer to her trauma already existed. It was hidden behind the password.

The same was true for photos. She wanted good images for her own Instagram and a gift quote she could pass on to her client without the supplier's name. The system already had buttons to download photos and copy descriptions. Nobody outside knew.

In the second round, the percentage hypothesis fell for good. Starting without investment does not make the percentage less annoying. What removes the annoyance is knowing that, if a piece sells well, she can move to buying it with a bigger margin. The page had to show that ladder.

How did the interview become a journey?

The two rounds produced a ten step path. The first steps are the ones that decide:

• Gets the link on WhatsApp, mid service, sometimes sent by the supplier itself, without a referral.
• First ten seconds: what it is, how much she pays and sells for, where it comes from. "If it didn't grab me right away, it's lost."
• Chooses a path: display in the store, buy to resell, or corporate gifts.
• Checks the math on three or four pieces, in reais, with a realistic retail price for Canoas, "not a Porto Alegre mall price".
• Looks for the risk: what if it does not sell, breaks or disappears.
• Looks for proof the person exists: name, city, Instagram, stores already selling.
• Messages on WhatsApp. The first yes happens in a conversation, not on the site.

An unexpected answer: stores in the region already selling the piece build trust, but a neighbouring store selling the same product is a worry. "In the region, but not on my street."

What did the prototype reveal that the interview did not?

The first version replaced the password with an open, mobile first page: a photo of the piece in someone's hand, the three paths, the math in reais, a "What if it doesn't sell?" block, who makes the pieces and WhatsApp always at hand.

The top worked. In ten seconds she understood it was 3D, from southern Brazil and with nothing to pay upfront. About the gift quote without the supplier's brand, she was blunt: "That won me over."

But transparency backfired at one point. For the same piece, the page showed how much she would earn displaying it and how much she would pay buying it. Side by side, the difference was glaring.

"It feels like consignment is the option for suckers."

Small words mattered too. "You only pay for what you sell" scared her: "For me it isn't paying, it's passing on." A rule marked "to be agreed" sounded like "we'll see later, and down the line it turns into a fight". And "every sale you mark is recorded" felt like extra work: "Mark how? Where? In the middle of a rush I'm not doing that."

Márcia, for her part, did the resale math on her own and reached the wrong result. Lesson: show the math already done, the way she thinks, instead of leaving the deduction to her.

The fix that created another problem

The second version swapped "pay" for "pass on our share of what you already sold". It split the paths into tabs, so the same product would not appear with two prices side by side. It put numbers on every rule in "What if it doesn't sell?". It explained that displaying and buying are different because, when displaying, Rotina Kraft covers the pieces, delivery, swaps and breakage.

"Pass on" worked: "I get it right away that the money comes out of what came in, not out of my pocket." The text about the difference did not sound like an excuse, because it was concrete. The ladder convinced her: "That's exactly what I was going to do."

But the tabs hid the corporate gifts, which are a large part of her revenue. "I couldn't skip past that without seeing it." Fixing one problem had created another. In the final version, the three paths are always visible as cards, and only the chosen one expands.

The "Sold" button also raised new questions: can her employee mark it? What if she forgets on a busy Saturday? Two lines went into the page: "on your phone or your employee's" and "forgot to mark it? We count together at settlement, and the count is what counts".

Her verdict on the final version:

"Yes, I'd message them: now I know how much I earn, that I don't put money in and what happens if it doesn't sell, so all that's left is seeing the piece in my hand and the agreement in writing, and I sort that out in a conversation."

Before and after

What can the simulation not do?

A simulated retailer is good at finding the obvious things we were not seeing. It is bad at telling what is true about real people.

Research on the subject names the risk clearly. Language models standing in for participants tend to portray a group the way an outsider imagines it speaks, not the way it actually speaks.

Renata sounds convincing because she was built from details of two real partnerships. But convincing is not the same as true. And in this case, the one asking and the one answering came from the same session. She may be right about the settlement trauma and wrong about how much text a retailer will read on a phone.

That is why the design has two stages. The simulation clears the cheap stumbles. Real partners decide what stays. The second stage does not need many people to pay off.

The simulation takes the part of the curve where problems are easy to find. That way, the real retailer's time goes to what only they can answer.

What is still missing?

I am writing this before the most important stage. Validation with real partners has not happened yet. The questions are ready, short, to be sent on WhatsApp, which is where retailers already are. One of them asks, plainly, whether the consignment percentage feels fair. Another asks the optical shop what it looks at first before buying to resell.

There are also decisions no interview can settle, because they are business rules: how many days until Rotina Kraft picks up what did not sell, what happens to broken or stolen pieces, which day settlement happens, when a piece that sells well moves to the purchase price list. Without those answers, the page would go back to "to be agreed", which was exactly what raised the most distrust.

If real retailers contradict Renata, they win. This text will be updated with what they say.

Customer research and AI, in November 2026

In November 2026, setting up two AIs to talk in separate roles takes an afternoon. What changed is not the ability to generate answers. It is how easy it is to give each side rules: the retailer could not suggest design, the designer could not answer for her.

That separation is what makes the simulation useful. Without it, it becomes an AI agreeing with itself. And the app's limit, which forced the retailer to be born inside the designer's session, shows that the separation still depends on whoever designs the process.

What did not change: deciding what is true about a customer is still work done by people, with people.

One question is worth more than ten hypotheses

Renata does not exist. But her question, "if I take it and don't sell it, what happens?", became the title of a section of the page, and her line about his notes against hers changed the sales argument. The rehearsal did not settle anything for good. It made sure we reach the conversation with real retailers knowing which of our assumptions did not hold up.