How to Prevent Off-Platform Transactions in Your Marketplace

Off-platform transactions are best prevented by making the on-platform path cheaper, faster and safer than the alternative, then detecting and pricing the leakage that remains. Policy alone rarely holds: in marketplaces with a take rate above roughly 10%, an estimated 10% to 30% of gross merchandise value (the total value of everything sold through the platform) can migrate to private channels within the first year. The sections below cover what to block, what to measure, and where enforcement should stop.

Why do sellers take deals off-platform?

Sellers move off-platform to avoid the commission, the percentage the marketplace keeps from each sale. On a 2,000 order at a 15% take rate, one message and a bank transfer saves the seller 300.

The take rate is the share of each order the platform keeps, and it is the single number sellers react to. A seller who has already paid 300 in fees has a strong incentive to take the next order direct, especially when the buyer has been found and the trust problem is solved. Platforms often assume the seller drives this. Buyers do it too, usually after a seller offers a discount for paying outside the platform, and once a buyer has paid direct once, they will ask again.

There is a second group that never intends to pay: accounts created to harvest leads. These sellers list real-looking inventory, collect enquiries, and close every deal in a phone call. They cost the platform in support time and refunds, and they are usually the first accounts worth removing.

What signals show a deal is moving off-platform?

The strongest signals are contact details in messages, payment references outside the platform's processor, and listing views that end without a booking. Weaker signals include repeated cancellations and buyers who ask for invoices from a different company name.

Signal collection has to be deliberate, because the useful evidence sits in systems that do not normally talk to each other. The payment processor knows which cards were used. The messaging system holds the text. The listing database knows what was viewed and never bought. If those three do not land in one place, every investigation starts from scratch.

Signals worth logging:

1. Contact details in any form: digits, spelled-out numbers, handles, or email addresses hidden inside images.
2. Payment identifiers: bank account numbers, wallet addresses, or the names of person-to-person transfer apps.
3. Negotiation threads that stop the moment a price is agreed.
4. Buyers with one completed order and no second order, in a category where repeat purchase is normal.
5. Sellers whose review count grows while their completed order count does not.
6. Invoice requests carrying a company name that does not match the account.

Each signal is weak on its own and most have innocent explanations. A courier needs a phone number. A business buyer needs an invoice with a registered company name. Treat signals as a score, not a verdict, or you will suspend sellers for doing normal things. A false positive in a supply-constrained category costs more than the fee it recovers.

Which controls stop off-platform deals before they happen?

Controls that work best remove the reason to leave: mask contact details in early messages, hold funds in escrow (money held by a neutral party until delivery), and cap fees for repeat buyers. Expect friction, and tune by category.

Escrow does most of the work because it moves the trust problem to the platform. A buyer who pays through escrow gets their money back if the item never arrives, and no seller can replicate that protection with a bank transfer. The trade-off is payout speed. Sellers run on cash flow, and holding funds for seven days instead of one will push some of them to test the direct route.

A watermark is a unique identifier hidden in the copy of a listing shown to one account, so a screenshot can be traced back to the account that saw it. Device fingerprinting combines browser and hardware details to link accounts one person controls, which is how a banned seller returns under a new name. Neither has to be perfect. A seller who knows leaks are traceable behaves differently.

How do you detect leakage you could not prevent?

Detection is a weekly review loop: rules flag accounts, a human reviews the evidence, and confirmed cases feed back into the rules. Expect roughly 20% to 30% of early flags to be false.

Two mechanics matter. Automated review of message content works better on patterns than on keywords, because sellers learn the keywords quickly, spell numbers out, or move the conversation to a photo of a business card. Velocity checks catch lead harvesters before they cost anything: an account that sends forty messages on its first day and completes no orders is not behaving like a real seller.

For confirmed cases, an enforcement ladder beats a single penalty:

1. Warning that cites the specific rule.
2. Fee adjustment or a short hold on payouts.
3. Listing demotion, so the account stays active but stops getting traffic.
4. Suspension, with the evidence attached.
5. Permanent removal and a block on linked accounts.

The first two steps recover more revenue than bans do, because most sellers who test a boundary stop once the boundary is visible. Keep the ladder written down and apply it consistently. Inconsistent enforcement teaches sellers that the rules are negotiable.

Where should enforcement stop?

Enforcement should stop where it starts punishing ordinary behaviour. Reading every message to find off-platform deals is surveillance, and buyers who learn that chats are scanned may take their business to a marketplace that does not scan them.

Tell users what is monitored and why. Most buyers accept payment protection and message filtering when the benefit is explained in one sentence at the point of use. What they do not accept is finding out after the fact, through a suspension.

Check the contract before you act on it, too. The clause banning off-platform deals needs to hold up in the jurisdictions you operate in, and it needs to survive a seller who challenges a withheld payout. Have that reviewed once, before you build the penalty system, rather than after the first dispute.

A rollout order that works

Sequence matters more than the size of the build, because each layer produces the data the next one needs.

1. Log the signals first. Two weeks of data shows which categories leak and how.
2. Ship contact masking and the fee cap. Both are cheap, and both reduce intent.
3. Add escrow in the two or three categories with the highest leakage, not everywhere.
4. Review flags by hand for a month before automating any penalty.
5. Publish the rules and the enforcement ladder on one page.
6. Re-measure after 90 days and compare fee revenue against the pre-launch baseline.

Building masking, escrow logic and payout release rules into an existing platform is typically a 6 to 12 week program, and it touches payments, messaging and admin tooling at once. Teams without that capacity often bring in a marketplace development partner rather than stretch an already busy product team.

Expect the result to be partial. A well-built system might cut leakage by half, not to zero, and the sellers who leave are sometimes the ones who were only there for the leads. That is a reasonable outcome: you keep the fee revenue from the sellers who stay, and you stop paying support and refund costs on accounts that never intended to trade honestly.