Skip to main content

Referral growth platform

How it relates to loyalty

The reward economics — reward type and value, minimum order, cap, first-order-only, referrer points — live in your loyalty programme configuration, which the live referral engine reads when minting and converting codes.

This hub owns the governance layer on top: approval to launch, the kill switch, budget, fraud policy and copy — and syncs economics through to the loyalty config so there is one management surface and no split-brain between two places that both claim to set the reward.

If you only want the basic referral mechanic, loyalty referrals covers it. This page is what you need when the programme is spending real money.

The eight tabs

TabWhat it is for
OverviewKPIs, leaderboard, acquisition funnel and timeseries
ProgramConfiguration, launch approval, kill switch, budget
CodesEvery minted referral code and its counters
ConversionsThe per-order ledger, including clawbacks
FraudThe review queue
CopyVersioned, localisable, publishable copy
InvitesCampaigns to your existing base
A/B testsCopy experiments

Governance

ControlWhat it does
Launch approvalA programme cannot go live without a second person approving it
Kill switchStops the programme immediately without deleting configuration
Rolling budgetA cap on referral reward spend over a rolling window
Fraud policyThresholds for the detectors below

A referral programme is an open-ended discount commitment. The kill switch exists because "turn it off now" and "unpick the configuration" are different urgencies, and the rolling budget exists because the failure mode is not one bad order, it is a code spreading somewhere you did not intend.

Copy that matches the offer

The copy tab authors the marketing text. The programme configuration holds the actual reward. These can drift, and when they drift customers arrive at checkout expecting something different from what they get.

Typed incentives in free text — $20 off, 15%, free shipping — are parsed into a structured offer and reconciled against the configured discount, with any mismatch surfaced before you publish.

Copy is versioned per locale, with one published row per store and locale. What is published is what the storefront widget renders and what invite campaigns send — so publishing pushes copy live without retyping it into the theme editor.

The conversion ledger

Every referred order becomes a conversion row: which code, which referrer, which referee, the order value and the reward.

Refunds and cancellations reverse the conversion, so the leaderboard and the programme cost reflect kept revenue rather than gross orders. Without clawback, a referral programme's reported performance drifts steadily above reality by exactly your refund rate.

Fraud

Referral fraud is detected heuristically, because the order path has no device fingerprint. The signals are built from what is genuinely available: customer identity, order timing, refund ratios and email quality.

SignalRaisedWhat it catches
Self-referral attemptAt order timeSomeone using their own code. The block happens in the engine; the signal exists so repeat abusers surface
Shared identityAt order timeReferrer and referee resolving to the same person
Disposable emailAt order timeThrowaway addresses used to farm rewards
VelocityDaily sweepAn implausible number of conversions in a window
Refund abuseDaily sweepA referrer whose referred orders are refunded at an abnormal rate

The two aggregate detectors need a cross-order view, so they run as a daily sweep rather than inline.

Invite campaigns

Codes were historically minted only for customers who visited the storefront while logged in — which meant there was no way to push the programme to the base you already have.

An invite campaign blasts your published referral copy to a Customer 360 segment by email and SMS, with each recipient's own code and share URL generated or retrieved. Sends run in the background, one recipient at a time, fail soft, and respect email opt-in.

Copy A/B tests

Each referrer is bucketed to a variant deterministically, and that variant's copy is genuinely served to them — in the storefront widget and in invite sends. Serving records an exposure.

The engine stamps the variant on every conversion, so results compare conversions among the exposed, divided by exposures, per arm.

Why the exposure denominator is the whole design

Without three things — copy that is actually served, copy that actually differs, and a denominator of who saw it — a declared "winner" is noise dressed as a result. Creating an experiment therefore requires per-variant copy; there is no way to run a test where both arms show the same thing.

Analytics

ViewWhat it answers
KPI headerCodes minted, conversions, revenue, programme cost
LeaderboardYour most effective referrers
Acquisition funnelShare → click → order
TimeseriesTrend over time
Code and conversion tablesPaginated detail, with clawbacks visible

Permissions

PermissionGrantsMinimum plan
referral:read:allView the hubFree
referral:copy:generateGenerate, edit and save copyStarter
referral:manageConfigure, manage codes, triage fraud, run invitesStarter
referral:copy:publishPublish copy live to the storefront widgetPro
referral:experiment:manageLaunch, decide and abandon experimentsPro
referral:approveApprove a programme launch (checker)Agency

Plan tiers

CapabilityFreeStarterProAgencyEnterprise
View the hub and analytics
Programme config, codes, conversions
Fraud queue and daily sweep
Copy authoring and versioning
Invite campaigns
Publish copy to the storefront widget
Copy A/B experiments
Launch approval (second checker)

Best practices

  • Set the rolling budget before you launch, not after the first surprise week.
  • Publish copy rather than pasting it into the theme. The published row is what the widget and invites both read, so they cannot drift apart.
  • Run an invite campaign to your existing base. Most referral programmes under-perform because nobody knows they exist, not because the reward is wrong.
  • Check the reconciliation warning before publishing. A copy-versus-config mismatch is a support ticket per order.
  • Let an experiment run to a real sample. Referral conversion rates are low; a two-day test tells you nothing.

See also