Skip to content

Part 4 of 7 · Referral payout runner series ~6 min read

How fraud is handled without punishing everyone

Referral fraud is real, mostly small, and much less expensive than the standard response to it. The interesting design question is not how to catch more of it; it is how to catch it without making honest referrers feel accused.

Key takeaways

  • Three patterns are worth catching: self-referral rings, bulk signups, and stolen codes.
  • A false positive costs a referrer permanently. Weight it accordingly.
  • A hold is always shown to the person, with the reason, and with a way to respond.
  • Never suspend an account on an automated signal alone.
  • Track the false-positive rate as a headline number, not the catch rate.

What is worth catching

PatternWhat it looks likeWorth catching?
Self-referralSame person, second accountYes — caught at recording
Referral ringA small group referring each other in a loopYes — clear and rare
Bulk fake signups40 referrals in an hour, no ordersYes — and obvious
Stolen or scraped codesOne code used from many countries at onceYes
Incentivised sharingPosting the link on a deals forumUsually not fraud
Referring familySame surname, same addressAlmost never fraud

The bottom two rows are where programmes go wrong. Same-address referrals are overwhelmingly people recommending things to the people they live with, which is the most natural referral there is, and a rule that blocks them is blocking the intended behaviour of the programme.

The asymmetry

Fraud prevented against referrers lost, at three rule strictness settingsA stacked bar chart with three bars in pounds per quarter. Two series: fraud paid out in red, and the value of referrers who quit after a false hold in orange. Loose rules pay out three hundred and forty pounds of fraud and lose one hundred and twenty pounds of referrers. Balanced pays out one hundred and forty and loses three hundred and eighty. Aggressive pays out forty and loses two thousand one hundred. A note says the aggressive column stops almost all the fraud and costs five times as much.£0£1000£2000£3000£4000~£460Loose rules~£520Balanced~£2140AggressiveFraud paid out, £/quarterValue of referrers who quit after a false holdThe aggressive column stops almost all the fraud and costs five times as much.
Fig 1. The trade at three settings. The right-hand bar is the one that feels prudent in a meeting and is the most expensive option on the chart.

The reason the aggressive setting keeps getting chosen is that the red bar is measured and the orange one is not. Fraud paid out appears in a report; referrers who quietly stopped referring after being told their referral was under review do not appear anywhere at all.

Making the orange bar visible requires only counting held referrals that were later released, and tracking whether those people ever referred again. It is a small piece of instrumentation that changes which setting people choose.

A hold is a conversation

How a suspected fraudulent referral is held and reviewedA vertical chain of five steps entered by a box labelled A signal fires for a ring, bulk signups or a stolen code. Step one holds the payment, not the account. Step two tells the referrer the same day, plainly. Step three says specifically what triggered it. Step four has a person review it within five working days. Step five releases or declines, with a reason either way. A note says holding the payment is proportionate, and suspending the account is not.AWS ACCOUNTA signal firesring, bulk, or codeHold the paymentnot the accountTell the referrersame day, plainlySay what triggered itspecificallyA person reviewswithin 5 working daysRelease or declinewith a reason either wayHolding the payment is proportionate. Suspending the account is not.
Fig 2. What happens after a fraud signal. Every box exists to keep a legitimate referrer from experiencing this as an accusation.
  • App integration
  • Machine learning
  • Management
  • People

Say what triggered it

“Your referral is under review” is the message that makes people angry, because it reads as an accusation with no content. “We hold referrals when several accounts sign up from one address in a short period — this happens legitimately with families and shared houses, and a person will look at it this week” is the same hold and a completely different experience.

The second version is longer, it is a template, and writing it once removes most of the support load that fraud holds generate.

Five working days, and it means it

A review queue with no service level becomes a place referrals go to be forgotten, which is functionally identical to not paying. The deadline needs to be stated to the referrer and monitored, and a queue older than the deadline should be an alert somebody sees.

The rules that quietly kill a programme

Four prudent-sounding rules to avoid

  • “Hold anything from a shared IP address.” That is every office, every student house, and most mobile networks.
  • “Hold same-surname referrals.” Recommending something to your family is the programme working.
  • “Suspend the code after any signal.” The referrer finds out when their next recommendation silently fails to register.
  • “Require the referred customer to confirm.” Adds friction to the one moment the programme depends on, to prevent a rare problem.
  • Each of these gets proposed after a single incident, and each costs more every quarter than the incident cost once.

The pattern is that fraud incidents are memorable and specific, while the cost of over-blocking is diffuse and invisible. Writing the four rules down as known bad ideas, before the incident happens, is a cheap way of having the argument in advance.

Next: what happens when somebody disagrees.

All posts