Every anti-fraud setup starts the same way: a bad month, a tightening of the rules, and a quiet drop in revenue nobody connects to the tightening. The fraud goes down. So do the paying customers, and they go down further.
Anti-fraud rules fail here for a reason that is arithmetic rather than technological. Fraud is a small share of attempts; good customers are the rest. A filter that removes a percentage of both takes far more money from the second group — and the loss is invisible, because a customer who was blocked does not complain, they just leave.
Contents
- Two numbers, always together
- What a rule should be built on
- The rules that cost more than they save
- Where authentication fits
- How to change a rule without guessing
- What to measure weekly
- When it is not a rules problem
- Who should own this
- In short
Two numbers, always together
An anti-fraud rule can only be judged on both sides at once. Approval rate tells you how much of your traffic turns into money. Dispute ratio tells you how much of it comes back. Move one without watching the other and you are guessing.
The costs are not symmetric either. A chargeback is 100,00 EUR on top of the returned amount and counts against the ratio the card schemes watch; a refund you issue yourself is 50,00 EUR and does not. A blocked good customer costs the whole margin on a sale that never happened — and nothing in your reporting records it. The published lines sit on the pricing page.
What a rule should be built on
- Behaviour, not identity. Ten attempts with different cards from one device is a pattern. One customer from an unusual country is not.
- Your own history. A template ruleset is calibrated on somebody else’s traffic. Your declines and your disputes are the only honest input.
- Velocity, not single events. Most real fraud shows up as speed: many small attempts in a short window, the same card across many accounts, the same account across many cards.
- Consequences that scale. Not everything has to be a block. Ask for authentication, delay settlement, cap the amount, require a second factor — a rule with only one outcome is a blunt rule.
Why the higher-risk categories see more of all this is set out in our piece on what makes a business high-risk.
The rules that cost more than they save
- Blanket blocks on whole countries. They remove real customers by the hundred to stop attempts you could have handled with authentication.
- Blocking on a single mismatched field. A billing address that does not match is a signal, not a verdict — people move, and forms get autofilled badly.
- Hard limits on amount alone. Your best customers are the ones spending the most; a ceiling set for safety is a ceiling set on revenue.
- Rules nobody reviews. The rule that made sense in March is still running in November against a different customer base.
Where authentication fits
The most useful anti-fraud decision is often not to block but to authenticate. When a payment is confirmed by the issuer, liability for a fraud dispute generally shifts away from you, and the customer stays in the flow instead of hitting a wall. Where that step loses conversion, and how to keep it, is unpacked in our piece on why traditional payment gateways do not work for high-risk businesses.
Routing matters too: how attempts are retried and distributed is described under payment processing.
How to change a rule without guessing
Change one thing at a time. Two rules changed in the same week cannot be told apart afterwards.
Run it in observation first. Log what the rule would have blocked without blocking it. A week of that data is worth more than a month of argument.
Look at what it caught. Pull twenty transactions the rule would have stopped and look at them by hand. If most are ordinary customers, the rule is a revenue decision disguised as a security one.
Give every rule an owner and a review date. Rules without either become permanent by accident.
What to measure weekly
- Approval rate — overall and for the segments the rules touch.
- Dispute ratio — split by reason, because «did not recognise» and «fraud» need opposite fixes.
- Blocked-then-returned customers. Anyone who was stopped and came back to pay is proof the rule was wrong.
- Refunds against disputes. Rising disputes with flat refunds means support is losing the race to the bank.
What the thresholds behind the dispute ratio are is set out in our piece on VAMP guidelines for high-risk merchants.
When it is not a rules problem
If approval keeps falling while your own rules stay untouched, the issue is usually upstream: a single acquirer, no fallback, routes that no longer fit the business. That is a different conversation, and the account behind it is described under merchant account.
Who should own this
Anti-fraud sits between three teams and therefore belongs to none of them by default. Finance sees the disputes, support sees the angry customers, and whoever runs the checkout sees the approval rate — and each of the three optimises a different number.
Give the whole trade to one person, with both numbers in front of them and the authority to change a rule. A ruleset that nobody owns drifts towards caution, because tightening never gets blamed for anything while loosening always does.
In short
Anti-fraud is a trade, not a switch. Build rules on behaviour and your own history rather than templates, prefer authentication and delay over blocking, change one thing at a time and observe before enforcing, and judge every rule on approval rate and dispute ratio together. Do that and the money you keep is larger than the fraud you stopped — which is the only measure that matters.

