Bringing questions and offers into one private chat

Before the chat, a buyer interested in a bike had to use two tools that did not talk to each other.

Public comments. On web and in the app, a "Questions & Comments" section below each listing was visible to every visitor. Buyers asked about size, condition or service history there, and sellers answered in front of everyone.

A separate offers tool. On the web, buyers made an offer from the listing, but everything after that happened in the user profile, as in the app. Sellers saw amounts and statuses, and could accept, reject or counter an offer, or even mark one as final. Nothing, however, bound the buyer, not even an accepted offer.

Neither tool was the problem on its own. The problem was the gap between them:

• Sellers lost the context. An answer about drivetrain wear stayed in the comments, while the offer that depended on it sat in another section, so the seller responded to an amount without seeing the question behind it.
• Nobody negotiates in public. A seller cannot give one buyer a better price while every other buyer is watching. To take the conversation elsewhere, some users posted their contact details in the comments, often enough that a warning appeared below the input field, and we moderated what we could. We did not know how widespread this was, but we knew that moving from a public exchange to a one-to-one conversation would be an improvement.
• Trust had nowhere to build. Buyers and sellers are strangers, often in different countries and speaking different languages. Trust comes from an ongoing exchange, and a public comment thread works like a shop window, not a conversation.

Users said so themselves:

In February 2024, we replaced both tools with a single private thread between buyer and seller.

The bet was that a private conversation would move negotiations forward better than two separate tools. It led to three principles, each framed as a question:

• Trust over control. What does the interface look like if we trust buyers and sellers to run their own negotiation?
• Restraint over intervention. How much can we remove from the interface while keeping the essential actions within reach?
• Infrastructure over features. How does a pre-sale messaging tool become the foundation of the whole transaction?

These principles became four interaction rules:

• Thumb zone. All primary actions sit at the bottom of the screen.
• A minimal header. The screen is for the conversation, not for transaction details.
• A strictly linear thread. Offers are messages like any other, and platform updates appear in the thread rather than in a modal. Making an offer opens a compact bottom sheet that leaves the conversation visible.
• A fixed action hierarchy. Never more than two actions on screen, and only ever one primary button.

The chat uses no red, and green is reserved for confirmed transactions.

The chat launched in February 2024, as the marketplace was scaling: sales in 2024 were four times those of 2023. A year later, a major update simplified the interface and put the conversation between buyer and seller at the centre, which meant deciding when the platform should step in and when it should stay out. This case shows that version.

After launch, reviews talk about the conversation itself:

In the chat, buycycle is always a third participant, so every design decision raised the same question: should the platform step in here, or leave buyer and seller to it? Over time, a rule emerged. During the negotiation, the platform stays in the background. It steps in only when someone commits: when the seller accepts an offer, or when the buyer pays. The three decisions below follow this rule, and each was reached in a different way: the first by persuading the team, the second by holding firm against internal requests, the third by deciding together.

If the platform steps in when someone commits, it needs a way to do so. For V2, we built on a pattern found on several large marketplaces: a platform message that sits in the thread like any other message. The format already existed. The work was in deciding when the platform speaks and what it says to each side, and it speaks at the moments of commitment. The buyer learns that the seller has accepted their offer and that they can buy at that price, then that the sale is confirmed and that they can track their order. The seller learns that their bike has been bought and that they need to confirm the sale.

The same format then carried three features, without any redesign of the thread:

• Claims. Post-purchase disputes, which support used to handle by email, are now resolved in the thread, with the full history of the negotiation.
• Order tracking. The order and delivery steps appear in the conversation, as described in the shipping case.
• Moderation. An internal task force defined the LLM moderation logic, and I designed how users find out that moderation has taken place.

For claims, I stopped at the model. I defined how the format applied to a dispute, then handed the first version of the full flow over to the team’s UX designer. That was the real test: a format I was not using myself had to be enough for someone else.

From April to September 2026, many more buyers started conversations, but purchases did not follow. Far more buyers opened at least one conversation in a month, and each opened more of them, so the number of conversations almost doubled. Yet purchases that followed a conversation barely moved. The share of these buyers who bought fell by about a quarter, and the share of conversations that ended in a purchase by more than 40%.

My first hypothesis was that buyers were spreading themselves thin: negotiating on more bikes in parallel, and still buying only one in the end. I checked it against the data, and it did not hold.

• Buyers who negotiate a lot still buy. In September, a buyer with a single conversation bought in 18% of cases, and a buyer with ten or more in 41%. The drop affected every group.
• Buyers and sellers did not reply less. The share of conversations in which the seller never writes, or the buyer stops replying, stayed the same.
• Buyers were not taking longer to decide. The time between a buyer’s first conversation and their purchase did not change.
• The drop comes after the price is agreed. First offers stayed at around 76% of the listed price, and sellers accepted them as often as before. But an accepted offer on a bike led to a purchase in 58% of cases in April, and in 39% in September.

So the problem is neither buyers spreading themselves thin nor unanswered messages. It sits between "agreed at this price" and payment, exactly where the second decision leaves the buyer free. From April to September, accepted offers rose by a fifth, while the number of buyers who reached checkout stayed roughly the same, and those who reached it completed it as often as before. The drop-off therefore happens after the seller accepts and before the buyer reaches checkout. What remains is to understand why: a better listing elsewhere, or an offer made without a real intention to buy. I had first located the problem upstream of the chat. It turned out to sit in its last step, and checking rather than assuming led me back to my own decisions.

• Guiding offers without deciding for the buyer. During the negotiation, the platform already steps in a little: buttons for −5, −10 and −15% help buyers propose a price, with −10% recommended. The need is real. From April to September 2026, more than 70% of first offers on bikes were below 85% of the listed price, while an offer within 5% of the listed price converted 3.8 times better. The question is how far to guide buyers without deciding for them.
• Keeping the principles alive. A decision does not stay made on its own. The team changes, and each new person arrives with the instincts everyone had in 2023: a Decline button, binding offers, fees shown everywhere. A large part of my role has been to restate the same principles and explain them again with examples, so that the team can make these calls without me. One question remains: how will these principles hold once no one is there to restate them?