Every promotion is a bet that a new customer will stay. A sportsbook gives a free bet, a fintech app pays a referral credit, a store or delivery app gives a first-order coupon. The math works when one person claims it once. It breaks when one person claims it forty times.
Each abusive account looks fine on its own: different name, different email. The abuse only shows when you look at what the accounts have in common. This guide is about finding that, responding with the lightest step that works, and putting a number on it you can defend in front of your finance team.
If you already read our iGaming bonus abuse detection playbook, this one is the operating plan: what to do in which order, and what to measure.
Who this is for
- Sportsbooks and casinos, including regulated LatAm markets. In Brazil, Law 14.790/2023 (art. 29, I) bars licensed operators from granting “adiantamento, antecipação, bonificação ou vantagem prévia” for placing a bet, so the classic welcome bonus is gone there. Promotions for existing players, missions and cashback-style offers still need the same controls. Confirm with counsel what your promotions may include. Our post on Brazil’s regulated market goes deeper.
- Fintech referral programs, where the referrer and the “friend” are often the same person.
- E-commerce first-order coupons and delivery apps, where new accounts keep appearing at the same address.
How abuse shows up in your data
You don’t need a model to start. You need four questions answered per new account.
1. How many accounts does this device already hold? One laptop or phone with three accounts is worth a look. One phone with twelve is not a family. A stable device identifier that survives cleared cookies and private windows is what makes this question answerable.
2. Was it created in a burst? Abuse comes in batches. Accounts that share a device or a network and were created inside a short window are the clearest pattern you will see: “9 accounts from the same place in 20 minutes” is not a coincidence.
3. What network did it come from? Datacenter IPs, VPNs and proxies are context, not proof. Plenty of real people use a VPN. But a batch of new accounts all arriving from datacenter addresses, with automation signals in the browser, is a different story.
4. How does the money move? This is where abuse becomes obvious in hindsight:
- Deposits of exactly the minimum needed to activate the offer.
- Wagering that clears the requirement with the lowest possible risk, then a withdrawal.
- Several accounts withdrawing to the same payout account, wallet or card.
- Referral pairs where the referrer and the referred account share a device.
- First orders with a coupon going to the same address with small variations (“Apt 4”, “Apto. 4”, “#4”).
The playbook
Step 1: Watch for 14 days before you act
Turn on identification on signup and on the promo claim, and change nothing for two weeks. This is the mirror period. At the end you get a list of what you would have stopped, case by case, and you can check it against what actually happened: who withdrew, who charged back, who became a real customer.
This protects you from a rule that looks smart on paper and quietly blocks your best new customers. In Prynt Control, mirror mode does this for the first 14 days, and any new rule can run in evaluate mode before you activate it.
Step 2: Make one decision at the promo moment
Don’t scatter checks across your code. At the moment a bonus, credit or coupon is granted, ask for one decision with three possible outcomes:
| Outcome | When | What you do |
|---|---|---|
| Allow | First account on the device, clean network | Grant the offer |
| Step-up | The device already holds accounts, another account already claimed a bonus here, a VPN or a disposable email, automation | Grant after one extra step |
| Block | Three or more other accounts already claimed a bonus on this device, or a device or account you labeled fraud or put on your deny list | Don’t grant; keep the account if you want |
Prynt’s bonus/promo use-case decision returns exactly this: send useCase: 'bonus' with the identification at the claim. By default a claim on the same device counts as a repeat for 365 days; you can change that window in the console. Prynt never blocks anyone on its own: it returns a decision, and your code decides what happens. Notice that “block” can mean “no bonus” rather than “no account”. A person who opened a second account can still be a paying customer without the welcome offer.
Step 3: Climb the friction ladder one rung at a time
Instead of a yes/no gate, use a ladder. Each rung costs a real customer a few seconds and costs someone running forty accounts a lot more:
- None. Most new accounts should see nothing.
- Captcha. Stops scripts, costs humans almost nothing.
- Verify email. Cheap for a real person, tedious at scale.
- Card or passkey. A card in the player’s own name, or a passkey tied to the device, is expensive to repeat across dozens of accounts.
- Block. Reserved for direct evidence: a device that keeps claiming the bonus for new accounts, and devices or accounts you labeled fraud.
The friction ladder tells you which rung to ask for on each request: an email check for a new account on a device that already has accounts, a card or passkey when another account already claimed the bonus there. Two rules of thumb: never jump from rung 1 to rung 5 on a single weak signal, and give shared households room to pass with one extra step.
Step 4: Review clusters, not accounts
Abusers work in batches, so review in batches. Account clusters, still rolling out in Prynt Control, will group new accounts that share device, network and a time window, so one reviewer can look at a cluster of nine accounts in a minute and decide for all nine, instead of finding them one at a time over a month. Until then, start from the device: accountsOnDevice on each event lists the other accounts seen on it.
Step 5: Label every case
Mark the cases you review fraud or legit, in the console or with POST /v1/outcomes. Use both. Legit labels matter as much as fraud labels: they teach the system which patterns are just a family on one tablet, and a fraud label makes the next decision on that device or account a block. The case queue, still rolling out, will rank cases by suspicion; the Stripe connection, also rolling out, will turn disputes and refunds into fraud labels automatically, so labels arrive from chargebacks without anyone typing.
Step 6: Put a number on it every month
Finance will ask what the promo program actually lost. Give them a number built so it is hard to argue with:
- Count only confirmed cases. Accounts labeled fraud, or with a dispute.
- Use the value that left. Bonus paid out, credit withdrawn, coupon discount applied on a delivered order. Not the face value of offers granted.
- Report a floor, not a forecast. Suspected but unreviewed accounts go on a separate line.
- Show the method next to the number, so anyone can redo it.
Prynt Control’s monthly money statement, still rolling out, is built this way: bonuses that went to abusers, estimated conservatively, with the method shown. It will report what was found, not a promise of savings. Until it reaches your account, the four rules above work in a spreadsheet.
What to measure
Track these monthly, and keep the definitions fixed so the trend means something:
- Prevalence: share of new accounts that look abusive. Is the problem growing or shrinking?
- Proactive rate: share of confirmed cases you caught before the payout, rather than after.
- Step-up pass rate: how many people asked for an extra step complete it. A very low rate on a rung can mean it is too hard for real customers.
- Legit rate among reviewed cases: share of reviewed cases labeled legit. If it climbs, your thresholds are too tight.
- Promo cost per retained customer: total promo spend divided by new customers still active after 30 or 90 days. This is the number the promotion exists to improve.
What a small team can and can’t copy
Large platforms have data science teams and years of labeled history. You don’t, and you don’t need to pretend. What transfers is the method: observe before acting, decide at the moment value moves, prefer friction to bans, review in batches, label everything, and measure in money. The signals underneath need to be accurate, which is why we publish ours: in our lab, datacenter and VPN detection hit 99.8% (89.5% on a blind holdout), and 0 of 240 real humans were flagged as bots. See /accuracy for how we test.
Start with the 14 days of watching. The list at the end will tell you how big your problem is before you spend any effort fixing it. Prynt Control bundles the steps above for iGaming, fintech and e-commerce teams.
Try it free
Prynt is device intelligence with a free plan: 5,000 identifications a month, visitor IDs with confidence and bot detection — no credit card. Upgrade to Pro for VPN, proxy, tampering and behavioral signals.