Referral Loops in Marketplaces

By Brian Tr · Updated Oct 6, 2026

A referral program in a software product has one job: turn a happy user into another user. A referral program in a marketplace has a harder brief, because there are two kinds of user and the person you most want to invite is often on the other side of the market. A host can invite a guest. A driver can invite a rider. A shop owner can invite their regular customers to buy from them on the platform.

That cross-side move is what makes marketplace referrals different, and it's the part generic referral advice skips. If you want the general mechanics of reward design, payout timing and program rules, our guide to referral programs covers them, and the math of virality in a single-sided product sits in viral network effects. This article stays on what changes when there are two sides: who invites whom, what a double-sided reward does to your take rate, how to count loops instead of funnels, and how people cheat.

Loops, not funnels

Start with the framing, because it changes what you measure. In "Growth Loops are the New Funnels" (published July 31, 2018), Brian Balfour, Casey Winters, Kevin Kwok and Andrew Chen argue that a funnel is a one-directional picture. In their words, funnels operate in one direction: you put more in at the top and get more out at the bottom, with no concept of reinvesting what comes out at the bottom to get more at the top. They contrast that with loops, which they describe as closed systems where inputs go through a process that generates more output that can be reinvested in the input.

A referral loop in a marketplace fits that definition well. A completed transaction produces a satisfied participant. That participant sends an invite. The invite produces a new participant, who completes a transaction, and the cycle repeats. The funnel view would count "signups from referral" as a channel and stop. The loop view asks a more useful question: what fraction of completed transactions produce another completed transaction through an invite, and how long does one turn take? Our piece on in-product growth loops walks through loop anatomy in general. The marketplace twist is that the loop can cross the market.

The four referral directions

In a one-sided product, an invite goes from a user to a lookalike user. In a two-sided market there are four possible directions, and each behaves differently.

The two same-side directions are the ones most programs build first, because they're easy to copy from consumer apps. They work, but they only move one side. If you've already established which side is scarce in a given city, as covered in supply-first vs demand-first, a same-side loop that adds the abundant side does nothing for you.

The cross-side directions are the distinctive marketplace loops. When a seller brings their own customers, the platform gets demand it didn't have to pay to acquire, and the seller has a reason to push it: they get paid through the platform and often gain tools such as booking, payments and reviews. This is the same logic as the single-player tool pattern discussed in marketplace supply acquisition, applied in reverse. The seller gets something useful; the marketplace gets its first buyers.

The buyer-brings-seller direction is the least common because the buyer has little to gain. It usually works only where the platform builds a way for a buyer to request a specific provider and the provider is eager to be found. It also feeds directly into the quality controls covered in marketplace supply quality, since an invited seller still has to pass whatever bar you hold.

One useful way to use this table is to ask which of the four directions is physically possible in your market, then ask which one adds the scarce side. Often the answer isn't the one the growth team already built.

A published data point on a supply-side loop

Lenny Rachitsky's series on kick-starting marketplaces is the best public collection of early-growth accounts, and it includes referral data. In Part 3 (published November 25, 2019), a person who worked on Uber's early growth says that on the supply side, referrals were about a third of first trips, and the referred drivers were also the best drivers. Treat it as one recollection by one early insider, not a benchmark for your market. But it does illustrate why seller-invites-seller loops are attractive: the people who arrive through a peer tend to look like the peer who vouched for them.

Double-sided incentives

Most marketplace referral programs reward both people: the one who sends the invite and the one who accepts it. The reason is structural. In a marketplace the invitee is often the scarcer, more skeptical party. They have no reason to try a new platform unless something lowers their risk.

Airbnb's engineering team published a first-hand account of rebuilding its referral program. In that March 27, 2014 post, Jason Bosinoff writes that the program gave both the sender and the recipient $25 of travel credit when the invited user completes their first trip. Three details from the post are worth noting for anyone designing a reward.

• The reward triggers on a completed first transaction, not on signup. The platform pays for outcomes.
• Framing mattered. Airbnb tested an email that stressed earning $25 for inviting a friend against one that stressed sharing $25 with a friend. The post reports that the sharing message performed better globally.
• The program was measured as a funnel of rates: monthly active users sending invites, invitees per inviter, and conversion to new user, new guest and new host. That list shows that Airbnb tracked referrals into the host side as well as the guest side.

The post also reports that referrals increased bookings by over 25% in some markets. As with any company's own account of its own program, read that as a result from one company in 2014, not a promise.

How rewards interact with take rate

Here's the part that's specific to marketplaces and easy to get wrong. A marketplace doesn't earn the whole value of a transaction. It earns a share, the take rate, so every credit you hand out is paid out of a thin slice. A subscription product can give away a free month against a high-margin recurring price. A marketplace is giving away money against a percentage of one order.

The arithmetic below is illustrative only, with made-up numbers chosen to show the mechanism.

On these assumptions the first order loses money, and the referral only pays back if the new user transacts again. With a repeat rate and average lifetime revenue per user, you can compute payback. The point is that the right reward size is a function of your take rate and your expected repeat behavior, not of what a consumer app down the street offers. Three consequences follow.

1. Cap the reward by contribution margin, not by revenue. If an order barely covers payment fees and support, the margin left to fund a credit is smaller than the take rate suggests.
2. Pay on the first completed transaction, or a later one. Paying on signup funds people who never transact. If the market is supply constrained, you may even pay more for the scarce side's second transaction than for their first.
3. Know whether the credit comes out of your take or the other side's price. A credit funded by the platform lowers your margin. A credit funded by a seller discount lowers the seller's earnings, which can anger the very side you need to keep. This ties closely to seller onboarding, because a supplier who feels shortchanged in the first week is a churn risk.

Customer acquisition cost is the right yardstick for all of this: compare the fully loaded referral cost per activated user to what other channels cost, using the definition in customer acquisition cost.

The viral coefficient, with a worked example

The viral coefficient, usually called K or the K-factor, measures how many new users each existing user brings. It's the product of two things:

K = invites sent per user x conversion rate of an invite

If K is below 1, each cohort of users produces a smaller next cohort, and referral growth fades unless you keep feeding the loop from other channels. If K is above 1, each cohort produces a bigger one and growth compounds. Sustained K above 1 is hard to reach, so most programs earn their keep by lowering blended acquisition cost rather than by running on their own. The full discussion of benchmarks and cycle time is in viral network effects.

Illustrative example, not real data. Suppose a marketplace's active buyers send an average of 3 invites each, and 10% of invitees complete a first purchase. Then K = 3 x 0.10 = 0.3. A starting cohort of 1,000 buyers produces roughly 300 referred buyers, who in turn produce about 90, then 27, then 8. The total from the loop is about 425 extra buyers on top of the original 1,000, close to the geometric-series limit of 1,000 x 0.3 / (1 - 0.3), or about 429. That's referral working as a discount on acquisition, not as a growth engine.

Why a marketplace needs more than one K

The single-sided formula hides what a marketplace actually cares about. With two sides you have four coefficients, one for each direction in the table above. Say, illustratively, that each seller brings 5 buyers on average and each buyer brings 0.1 sellers. A seller-to-buyer-to-seller loop then yields 5 x 0.1 = 0.5 new sellers per seller per full turn. The cross-side loop is the product of two numbers on different sides, and either side can be the weak link.

Two practical rules follow:

• Measure each direction separately. A healthy blended K can hide a dead cross-side loop.
• Measure in activated users, not signups. Count invitees who complete a first transaction, or a K computed on signups will flatter the program. Airbnb's own metric list, which ends in conversion to new guest and new host, follows this logic.

Also watch cycle time. A loop with K of 0.4 that turns weekly adds more users in a quarter than one that turns every two months. Marketplaces with long purchase cycles, such as home services or B2B procurement, have slower loops by nature.

Fraud and incentive abuse

Any reward attached to a transaction attracts people who want the reward without the behavior. Marketplaces have an extra exposure because the reward is paid on a transaction the abuser can often fake. The common patterns are:

• Self-referral. One person creates a second account to claim both sides of a double-sided credit.
• Collusion. Two real people trade referrals and fake or pre-arranged transactions to collect the rewards. In a ride or delivery market, a driver and a "rider" can arrange a trip that only exists to trigger a bonus.
• Fake supply. A referrer signs up accounts that accept jobs with no intention of completing them, to trigger a per-job payout.
• Referral farming. Posting codes on coupon sites to attract users who only ever use the discount.

Uber's published referral program rules are a useful example of how a platform writes these down. The rules state that referrers cannot request themselves as a rider or Uber Eats user, pre-arrange trips with riders, get someone else to drive or deliver for them, create duplicate or multiple accounts, or accept trips or deliveries without the intention of completing them. They also say that, unless an offer states otherwise, the referee must sign up or activate the offer within 90 days of the invitation being sent, and that Uber reserves the right to end or pause a programme for reasons including suspected fraud, with consequences that can include forfeiture of rewards and account deactivation. The rules are also explicit that rewards depend on the referee completing stated requirements, and that referees for riders may get a first-use discount while new drivers and couriers may get offers tied to the number of trips or deliveries they complete. (These rules vary by country; the page cited is one regional version.)

Reading those rules as a checklist gives you a control design:

1. Pay on a completed, verified transaction. A signup alone is easy to fake. A paid-and-completed order, with payment and fulfillment confirmed, is harder.
2. Add a minimum threshold. Require a minimum order value or a number of completed jobs before the reward unlocks, so a single contrived transaction isn't enough.
3. Link accounts by identity signals. Match device, payment instrument, phone number and address across the referrer and invitee, and hold payouts when they overlap.
4. Hold rewards for a window. A delay before payout lets chargebacks, cancellations and disputes surface first.
5. Cap and expire. Limit rewards per referrer and put an end date on invitations, as Uber's rules do.
6. Reserve the right to claw back. The terms need to say that rewards can be withdrawn for abuse, and support needs a process that applies it consistently.
7. Watch for clusters. Many referees from one referrer, all transacting at one time or place and then going quiet, is the shape of farming.

There's a trade-off here. Every control adds friction for honest users and delays the reward that makes the program feel good. Tune the controls to the size of the reward and the damage a fake can do, rather than applying maximum friction everywhere. A $5 credit on a low-risk market may not justify identity checks; a cash bonus on a high-value supply side probably does.

Where referral loops fit in the growth system

Referral is rarely the first thing to build. A referral loop multiplies what the product already does. If a new buyer's first transaction is disappointing, a reward just pays people to send friends into a bad experience. That's why referrals tend to come after the basics of the chicken-and-egg problem are solved in at least one segment: there has to be enough supply for an invitee to find something, and enough demand for an invited seller to get paid.

A short sequence that holds up in most markets:

1. Get one side to a reliable first-transaction experience.
2. Pick the direction that adds the scarce side and is possible in your market.
3. Set the reward from your contribution margin and pay it on a completed transaction.
4. Instrument each direction separately and count activated users.
5. Add fraud controls proportional to the reward before you scale it.

Key Facts: Referral Loops in Marketplaces

• A marketplace has four possible referral directions (buyer to buyer, seller to seller, seller to buyer, buyer to seller). The cross-side directions are the distinctive marketplace loops.
• Balfour, Winters, Kwok and Chen argue that funnels operate in one direction with no reinvestment, and that loops are closed systems where output is reinvested as input (Reforge, July 31, 2018).
• Airbnb's referral program gave both the sender and recipient $25 of travel credit once the invited user completed a first trip (Airbnb Engineering, March 27, 2014).
• In the same post, Airbnb reports that referrals increased bookings by over 25% in some markets, and that an email emphasizing sharing $25 with a friend beat one emphasizing earning $25.
• An early Uber insider says referrals were about a third of first driver trips and the referred drivers were the best ones (Lenny's Newsletter, November 25, 2019).
• Uber's published rules prohibit self-referral, duplicate accounts and pre-arranged trips, give referees 90 days from the invitation to sign up or activate and allow rewards to be withdrawn for suspected fraud.
• K-factor = invites per user x invite conversion rate. K below 1 lowers acquisition cost; K above 1 compounds.

Related Reading

• The Chicken-and-Egg Problem in Marketplaces
• Supply-First vs Demand-First
• Acquiring Supply for a Marketplace
• Marketplace Seller Onboarding
• Marketplace Supply Quality
• Viral Network Effects
• In-Product Growth Loops
• Referral Programs
• Acquiring Buyers for a Marketplace
• What Is Single-Player Mode?