In this article

A first-order offer produces a busy week. New buyers complete transactions, suppliers get work and the campaign dashboard looks convincing. The next decision is harder: should the marketplace keep paying for those orders?

Answer with a fixed observation window and an explicit cost ledger. Follow everyone eligible for the offer, preserve the customers who never transact, and separate subsidized first orders from later demand. That gives the team an operating result it can assess before attempting a stronger claim about what the incentive caused.

Write down the purchase the incentive is making

A buyer discount may be intended to encourage a first transaction. A supplier guarantee may purchase availability during a difficult time slot. Those are different interventions with different costs and possible benefits. State the intended behavior and the exact condition that makes the incentive payable.

For a buyer offer, record eligibility, exposure, redemption, completed transaction and final incentive cost. Keep the original eligibility time so customers who ignore the offer remain in the denominator. If the reward is paid only after completion, failed or canceled orders should not silently count as successful acquisition.

For supplier incentives, record the availability or completed work purchased and whether it was actually delivered. Do not combine a guaranteed shift payment with a buyer coupon into one unexplained “promotion cost” if the team needs to decide which intervention to continue.

Read an illustrative 60-day ledger

Consider two invented cohorts, each containing 100 eligible people followed for 60 days from eligibility. One receives a $20 subsidy on its first completed order; the comparison cohort receives none. This is a descriptive teaching example, not a randomized experiment or observed customer result.

Assume every completed order contributes $12 after all defined transaction-variable costs but before this incentive. There are no refunds or other promotions. Repeat orders receive no subsidy. Twenty people in each cohort repeat, producing 24 repeat orders in each.

Result within 60 daysOffer cohortComparison cohort
Eligible people100100
First completed orders6040
People completing a repeat order2020
Repeat completed orders2424
Total completed orders8464
Contribution before the incentive$1,008$768
Incentive cost$1,200$0
Contribution after the incentive-$192$768
Contribution per eligible person-$1.92$7.68

The offer cohort has more first orders, but its defined contribution has not covered the incentive within the window. Its 20 repeat buyers are 33.3% of first-order buyers; the comparison cohort’s 20 are 50%. Both have repeat buyers equal to 20% of all eligible people.

Those rates answer different questions. The first describes continuation among people who completed a first order. The second preserves the original opportunity population. Neither establishes an incremental effect because the fictional cohorts were not assigned through a causal design.

Keep the clock from changing the conclusion

The primary window starts at eligibility for everyone. That makes total spending and contribution comparable over the same duration. A person who orders late has less time to repeat inside that window; that is part of the program’s 60-day operating result.

For a separate question about behavior after first purchase, build an additional view with a fixed window from first completion and wait until those customers have enough follow-up. Label that view as conditional on becoming a buyer. Do not replace the eligible-population result with the more flattering one.

The offer cohort would need 16 additional unsubsidized orders at the assumed $12 contribution to cover its $192 deficit. That is arithmetic, not a forecast. It assumes those orders bring no additional incentive or unmodeled cost, and it says nothing about whether they will occur. A longer observation period can answer that question without projecting an attractive lifetime value from one repeat purchase.

Look for movement elsewhere in the market

An incentive can move an existing purchase forward, shift demand between sellers or change which buyers obtain scarce inventory. Inspect market-wide completed work and supplier availability alongside the promoted cohort. Also record whether later orders occur only when another discount appears.

a16z’s marketplace metrics framework discusses incentives within unit economics and recommends examining local markets separately. Use that level of detail when an offer purchases access to a particular supply pool rather than treating every city or category as one interchangeable market.

For an appointment marketplace, a discounted booking that fills an otherwise empty slot and one that displaces a full-price buyer have different implications. The descriptive ledger should preserve the slot, supplier and market so the team can investigate the distinction. It cannot observe both counterfactual outcomes for the same booking.

Be careful with a simple buyer-level A/B test

David Holtz and Sinan Aral’s research on test-control interference explains how treated and control participants in a marketplace can affect one another through shared supply or demand. Their simulation studies found that some clustering approaches reduced bias while increasing uncertainty; the paper also describes limits from its simplified market and network assumptions.

This matters when the business question is the effect of offering the incentive to the whole market. Randomizing individual buyers who compete for the same scarce appointments may not estimate that rollout effect cleanly. The comparison group can experience a changed market even though it receives no coupon.

Before choosing an experiment, map who competes for the same inventory, how far buyers or suppliers cross proposed boundaries, and how long the offer changes behavior. Those facts inform whether a market-level, time-based or other design is credible. No grouping automatically removes spillovers. Obtain design and analysis support appropriate to the market rather than treating a cluster label as a guarantee.

Until there is a defensible design, call the result descriptive. It can still reveal unaffordable spending, weak repeat use or a capacity problem and support a bounded operating decision.

Set the next funding decision in advance

Name the observation window, maximum spend, contribution definition and conditions for continuing before expanding the offer. Include service quality and supplier capacity so a burst of subsidized demand does not receive credit for work the marketplace cannot fulfill.

The example’s result supports reviewing the offer’s economics before adding budget. It does not prove incentives never work or that the comparison cohort is the right causal benchmark. The next step could be a smaller targeted offer, more follow-up time or a properly designed test of a specific market constraint.

Connect the eventual decision to the growth loop the incentive is meant to support. When prioritizing growth work, compare its observed cost and unresolved assumptions with other uses of the budget. First orders are the beginning of that assessment, not the completed business case.

Sources

Something we should clarify?

Suggest a correction