How to Measure a Rewards-Based Campaign

Hypothetical illustration. The example figures match the readable tables below; they are not Zenko campaign results. Open the full-size illustration.

Measure four things in order: participation, verified qualified actions, what those people do next, and incremental lift against a randomised holdout group. Count every cost in the numerator. Attributed results flatter reward campaigns, so do not report a cost per result that has no comparison group behind it.

This is the measurement step of rewards-based marketing, and the one that decides whether the rest was worth doing.

Why attributed results mislead

Gordon and colleagues compared observational measurement with randomised experiments across 15 Facebook advertising campaigns, covering around 500 million user-experiment observations. The observational methods often failed to reproduce the experimental results, and they tended to overstate effectiveness (Marketing Science, 2019). Those were ad campaigns rather than reward campaigns, but the logic carries over. People who accept an offer differ from people who do not, and many of them would have bought anyway.

Blake, Nosko and Tadelis found the same pattern in eBay's paid search on brand terms, where much of the traffic would have arrived without the ad (Econometrica, 2015). A reward campaign has the same risk. If you pay people who were already going to act, the campaign looks busy and achieves little.

The metrics

You will see cost per verified qualified action called cost per qualified engagement elsewhere. We use "action" because "engagement" is loose enough to cover almost anything.

How this differs from CPL, CPA, CAC and CPE

• Cost per lead is what you paid for an attributed lead of any quality.
• Cost per acquisition is what you paid for an attributed action.
• Customer acquisition cost is total acquisition spend divided by new customers won.
• Cost per engagement is spend divided by interactions of any quality.
• Cost per verified qualified action is spend divided by actions that were checked and met a stated standard.

Do not present cost per verified qualified action as customer acquisition cost. It becomes comparable only when you restrict it to new customers and adjust for what would have happened anyway. The guide to lead quality explains why a low cost per lead can hide a poor one.

A worked example

These numbers are invented to show the arithmetic. They are not Zenko results.

A brand invites people to complete a three-lesson course. Of 20,000 eligible people, 18,000 are invited and 2,000 are held out at random.

Attribution says £6.25 and the holdout says £37.50. Both calculations are correct. Only the second estimates the extra outcomes above the holdout baseline; uncertainty still matters.

The example also carries a warning. With only 40 purchasers in the holdout, the baseline is noisy, and a real test needs a larger holdout or a longer window. Work out the smallest lift you could detect before you launch.

Pitfalls to name in your report

• Selection bias. People who accept a reward differ from those who do not.
• Deadweight. You paid people who would have acted anyway, and the regulars in the gym evidence are the classic case.
• Cannibalisation and timing. A reward can pull forward a purchase that would have happened next month.
• Novelty. Early enthusiasm fades, so a first-week result is not a steady state.
• Regression to the mean. If you target the customers who look most lapsed, some would have returned by chance.
• Attribution windows. A long window credits the campaign for purchases it did not cause.
• Before-and-after comparisons. They cannot separate the reward from seasonality, other marketing or self-selection.

Check what happens when the reward ends

Measure the invited and holdout groups again in a window with no reward running, for example days 61 to 120. In the largest gym study, effects that were clear during the programmes had mostly gone once they ended (Milkman et al., Nature, 2021). A reward campaign that never looks beyond the payout period has not measured whether it built a habit.

Fix the stop rule before you launch

Decide in advance the smallest lift you would count as a result. If the result does not meet the pre-set threshold, report that plainly. Use "no detectable lift" only when the statistical analysis supports it; a small estimate is not the same as proof of no effect. Deciding afterwards invites you to find the result you wanted.

Fraud losses belong in the cost ledger. How to prevent reward farming without punishing real customers explains how to estimate them. And if you are still choosing the action, start with how to choose the right customer action to reward.

Where Zenko fits

For a campaign with Zenko, start by agreeing the comparison group and the cost ledger, not just the reward. Define the records needed to count participation and verified qualified actions, and include the full cost of verification and delivery. Scope any integrations and reporting requirements before launch rather than assuming they are standard features.

If an action also funds a reward for good, report action completed, funds allocated, partner delivery and independently verified outcome separately. A record at one stage is not evidence of the next.

Cost per lead is only the start. Measure what those leads do next and compare the full cost per verified qualified action with the cost per incremental outcome.

Tell us the action you want to reward and the outcome you need. We will show you how a campaign could work.

Plan a rewards-based campaign with Zenko

Frequently asked questions

What is a good ROI for incentive marketing?

We could not find a credible pooled benchmark, so we do not quote one. Vendor statistics on average incentive ROI rarely say how they were measured. Measure your own campaign against a holdout group and judge it on cost per incremental outcome.

What is cost per qualified engagement?

It is the total cost of a campaign divided by the number of engagements that were verified and met a stated quality standard. We call it cost per verified qualified action. The important part is that unverified or low-quality interactions are excluded from the count.

What is a holdout group?

A holdout group is a random share of your audience that is not invited to the campaign. Comparing what it does with what the invited group does shows how much of the result the campaign caused.

How long should you measure after a reward ends?

Long enough to cover your normal purchase or usage cycle, and in a separate window from the reward period. Sixty and 120 days are illustrative starting points, but the right window depends on how often your customers buy.

Sources

• Gordon et al., A Comparison of Approaches to Advertising Measurement, Marketing Science, 2019
• Blake, Nosko and Tadelis, Consumer Heterogeneity and Paid Search Effectiveness: A Large Scale Field Experiment, Econometrica, 2015
• Milkman et al., Nature, 2021

In this rewards-based marketing series

• Rewards-Based Marketing: How to Reward Actions That Actually Matter
• How to Choose the Right Customer Action to Reward
• How to Prevent Reward Farming Without Punishing Real Customers
• Incentive Marketing Examples That Work Beyond Discounts

Related acquisition measurement

Before comparing channels whose leads are different ages, check that the measurements are fairly comparable.