Here's a fraud scenario physical ecommerce rarely has to think about.
Someone buys a game key with a stolen card. The key is revealed instantly. Within minutes it's redeemed, or resold on a grey-market site.
Weeks later the chargeback arrives. You refund the money. You don't get the key back. No package to intercept, no return to process, no address to flag. The product left the moment the page loaded.
That's daily life for Increddy, a license-key storefront, and it's why "Stripe handles fraud" wasn't enough.
Why digital goods are a fraudster's favourite
Payment processors rank digital goods among the highest-risk categories, and the reasons stack up:
- Instant delivery. No shipping delay to catch a problem in.
- No address to verify. Nothing has to match anything.
- Instantly resellable. The value is liquid the second it's issued.
- Nothing to recover. Once the key is revealed, it's gone.
Every risk decision that matters has to happen before checkout completes. Afterwards, you're just counting losses.
What Stripe actually answers
To be clear: Stripe's fraud tooling is good, and it stays in Increddy's stack. It scores the payment (card, device, transaction pattern) against signals from across Stripe's network.
But it answers "is this payment risky?"
It doesn't answer "is this payment risky **given that the product is an instantly resellable key we can never claw back?"
Those are different questions, and a key store needs both answered.
A veto, not a footnote
The design choice that matters isn't any single fraud rule. It's where the risk check sits.
A check that runs after Stripe approves is only reacting to a decision that's already been made. So Increddy's risk scoring is a first-class Medusa module inside the checkout flow with the authority to block checkout completion on its own, whatever Stripe returned.
It gets a vote before the order exists, not a note attached to a transaction that already happened.
What we're not telling you
You might have noticed there's no list of signals or thresholds in this post. That's deliberate.
It's the same reason banks don't publish their fraud rules: a risk system works partly because the people it targets can't read the rulebook. The architecture is worth sharing. The heuristics aren't, because publishing them would make them worse at their job.
The rule
Treat fraud risk as a property of the product category, not only the payment. If what you sell is valuable the instant it's delivered and impossible to take back, give your own risk logic a real seat at checkout, with real veto power.
The fraud layer was one of several custom modules in Increddy's move off Shopify. See the full case study on kryeo.com →
The longer version of this post is on the Kryeo blog.
We're Kryeo. We build checkout and risk logic for headless stores through our Medusa development work.
Top comments (1)
Risk as a property of the product category rather than of the payment is the right call, and the placement is the part worth copying: a check that runs after the processor has answered is reacting to a decision somebody else already made. Cards split it the same way. The issuer answers whether the money is good. Whether to accept the order was never the processor's question.
Two things from the chargeback side.
The asymmetry you are describing has a number attached to it that strengthens your argument. A dispute on a digital good is not a refund. It is the refund, plus a fee, plus a ratio, and the ratio is the expensive part: cross a threshold on disputes as a share of transactions and you are in a card scheme monitoring programme with per-transaction penalties, and at the far end you lose acceptance. So the cost of letting one through is not the price of the key. That is the strongest case for the module holding a veto rather than a vote, and it is not in the post.
The number nobody measures sits on the other side. A veto that blocks checkout produces declined good customers, and they are invisible. They do not complain, they do not show up in a fraud report, they just do not buy. Fraud loss has a line in the accounts and false declines do not, so a risk system drifts tight over time unless someone owns the second number. Issuers report approval rate next to fraud rate for exactly this reason. If Increddy's module can block, give it a weekly count of what it blocked and some route to finding out what share of that was real.
On not publishing the signals. Defensible, and I would name the limit rather than leave it as a principle. Checkout either completes or it does not, so every attempt is a free query against the rulebook, and somebody probing at twenty pounds a go maps the boundary in an afternoon. What secrecy buys is the cost of probing, which is why banks pair it with velocity limits on the probing itself instead of relying on the rules being unknown.