Measurement
Measure the incremental value of affiliate partnerships
A practical guide to baselines, holdouts, matched tests and decision rules that separate attributed revenue from value genuinely added by partners.

Before you begin
Attribution answers which interaction received credit under a rule. Incrementality asks a different question: what would probably have happened without the partner or intervention? Both are useful, but confusing them can reward activity that captures existing demand while overlooking partners that create consideration earlier in the journey.
Incrementality is not a single dashboard metric. It is an evidence programme built from a clear decision, a credible counterfactual, stable test conditions and a result large enough to change an action. Begin with a narrow commercial question and choose the least complex design that can answer it responsibly.
Name the decision before choosing the metric
Write the action the result will inform: change a commission, expand a partner type, protect a placement, adjust attribution or redesign a promotion. Then define the population, outcome, time horizon and minimum effect worth acting on. A test without a decision often becomes an interesting report that changes nothing.
Use an outcome close to business value: approved new customers, contribution margin, qualified leads or retained subscribers. Record guardrails such as refund rate, average order value, customer support cost and partner experience so a short-term lift does not hide a larger problem.
Build a baseline and inspect selection bias
Describe normal performance before the intervention: seasonality, campaigns, price changes, stock, product launches, competitor activity and customer mix. Identify why a person, region or publisher enters the exposed group. If high-intent customers are more likely to see the partner, a simple exposed-versus-unexposed comparison will overstate impact.
Use enough history to understand ordinary variation, but do not assume last year is an exact counterfactual. Mark structural changes and keep a written test calendar shared with commercial, analytics and partner teams.
Choose a design proportionate to the question
Randomised holdouts can provide strong evidence when exposure can be assigned fairly. Matched-region or matched-audience tests are useful when randomisation is impractical. Time-based switches can help with stable, repeatable activity but are vulnerable to seasonality and carry-over. Difference-in-differences can compare changes between suitable groups when pre-test trends are credible.
Design checklist
- Unit of assignment and risk of group contamination.
- Primary outcome and approval delay.
- Minimum sample or duration required.
- Concurrent activity that could distort the result.
- Stopping, exclusion and interpretation rules agreed in advance.
Protect the test from operational noise
Confirm tracking, consent behaviour, partner links, promo codes, landing pages, inventory and validation feeds before launch. Freeze avoidable changes and log unavoidable ones. A test can be statistically neat yet commercially misleading if one group saw different prices, broken stock or a delayed payment experience.
Monitor implementation health without repeatedly peeking for a favourable result. Separate quality checks from outcome decisions and retain the original plan when presenting the final analysis.
Estimate lift with uncertainty and economics
Report the baseline, absolute difference, relative lift, uncertainty range and observed sample. Translate the result into approved value and contribution after partner commission, platform fees, discounts, returns and operational cost. A positive revenue lift can still be unprofitable; a modest immediate lift may be valuable when retention is strong.
Segment only where there was a prior reason or enough data. Searching many cuts after the test increases the chance of finding a dramatic but unreliable pattern. Label exploratory findings and validate them separately.
Turn evidence into a repeatable decision
Classify the result as sufficient to act, promising but uncertain, neutral within the detectable range or operationally invalid. Record what will change, who owns it and when the effect will be checked again. Share the method and limitations with the partner where this can improve joint planning.
Maintain a test register containing hypothesis, design, dates, exclusions, result, decision and follow-up. Over time, this becomes more useful than a single attribution model because it shows which partner behaviours add value under which conditions.
Read a conversion gap without overstating the experiment
An entirely illustrative experiment has two comparable groups of 10,000 people. The control has 300 buyers and the exposed group has 330. Conversion rises from 3% to 3.3%: +0.3 percentage points or +10% relatively. Both describe the same gap using different units.
| Measure | Calculation in this example | What it does not establish |
|---|---|---|
| Absolute difference | 3.3% − 3% = 0.3 percentage points | A 30% increase |
| Relative difference | (3.3% / 3%) − 1 = 10% | A guaranteed effect across the audience |
| Estimated additional buyers | 10,000 × 0.003 = 30 | That all 330 purchases were incremental |
| Uncertainty | Approximate gap interval: −0.18 to +0.78 percentage points | An established positive effect at 95% |
The interval uses a 95% normal approximation for two independent proportions: difference ± 1.96 × square root[p1(1−p1)/n1 + p0(1−p0)/n0]. It includes zero, so the observed gap alone is insufficient to conclude. The approximation assumes independent observations; geographic clusters, repeated visits and cross-exposure require a suitable analysis. Set the outcome and experiment duration before reading results.
Practical answers
Questions to settle before signing
Is attributed revenue the same as incremental revenue?
No. Attributed revenue follows a crediting rule; incremental revenue estimates the outcome that would not have occurred without the activity.
Does every partner need a randomised test?
No. Use the strongest proportionate evidence available. Randomisation is valuable but may be impractical, unfair or too small for some relationships.
Can incrementality be measured with low volume?
Sometimes, but the detectable effect will be larger and uncertainty wider. Combine quantitative evidence with a documented qualitative case and avoid false precision.
Should commission fall when measured lift is low?
Not automatically. Check the design, customer role, margins, assisted value and strategic contribution. A commercial change should follow a defined decision process, not one isolated estimate.
Sources and further reading
This guide is designed to help you ask better questions and organise a practical plan. It is educational and does not constitute legal, tax or financial advice. Platform rules, market conditions, technical capabilities and eligibility requirements can change, sometimes with little notice. Confirm material decisions with current official sources and, where the consequences matter, with qualified professionals who understand your market and organisation.