Your ad platform says the campaign worked. Your analytics report gives half the credit to something else. Your CRM says the customer came from “Other.” Very helpful.
You still need to decide whether to keep spending.
Proving marketing ROI starts with separating two questions: Which marketing touches received credit for sales? And how many sales happened because of the marketing?
Those questions need different evidence. Treating them as interchangeable is how an impressive dashboard becomes an expensive misunderstanding.
Attribution assigns credit. Incrementality tests impact.
Google Analytics describes attribution models as rules or algorithms that assign credit to touchpoints along a customer’s path. That makes attribution useful for understanding recorded journeys and making operational decisions.
It does not mean every attributed purchase would disappear if you stopped the campaign. Some customers were already going to buy. Some saw several channels. Some returned through a branded search after deciding yesterday.
The missing piece is the counterfactual: What would have happened without this marketing?
You cannot watch the same person simultaneously receive and avoid the same campaign. You can design a comparison that estimates the difference.
Write the proof standard before launching
Before arguing about attribution settings, write down the decision the evidence needs to support. “Should we expand this campaign?” is useful. “Can we find a number that makes this look good?” is also a question, technically.
Define these items before results arrive:
- The outcome: Completed purchases, qualified opportunities, or another specific business result.
- The comparison: Who receives the marketing, who does not, and how assignment happens.
- The meaningful effect: The smallest improvement that would justify the investment.
- The measurement window: Enough time for normal buying delays, refunds, and relevant sales follow-up.
- The decision rule: What result supports expanding, stopping, or continuing to learn.
Check whether your available audience and conversion volume can detect that meaningful effect. A tiny test cannot reliably answer a tiny-difference question just because everyone waits patiently.
Build a comparison you can defend
A randomized holdout assigns eligible people to treatment and control groups. The treatment group is eligible for the campaign; the control group is held out. Measure outcomes for everyone assigned, including people in the treatment group who never actually see an ad.
Google’s Conversion Lift documentation describes comparisons using user groups or geographic groups. Availability and suitability depend on the campaign and measurement setup.
Keep other conditions as consistent as practical. Watch for leakage, such as control participants encountering the campaign through another account or household. Check that assignment counts match the planned split within expected variation. An unexplained sample-ratio mismatch deserves investigation before interpretation.
When individual randomization is impractical, geographic testing may help. Randomizing eligible regions strengthens the design. Matched regions without randomization require stronger assumptions: similar historical trends do not guarantee similar future behavior. Local promotions, competitor activity, seasonality, and spillover across borders can distort the comparison.
Your landing-page testing discipline matters here too: decide what you are testing, keep the comparison interpretable, and resist changing everything halfway through.
Do not turn a small difference into a big claim
Consider a deliberately hypothetical test. Randomly assign 1,000 eligible people to each group. Fifty treatment participants purchase, compared with 40 control participants.
That is a 5% versus 4% conversion rate: a one-percentage-point observed difference, or an estimated 10 additional purchases among 1,000 treatment participants relative to the equal-sized control.
It is an estimate, not proof of a positive effect. With this small sample and these conversion counts, uncertainty is substantial. Chance remains a plausible explanation for the observed difference.
Google’s guidance on certainty of lift explains why conversion volume, control-group contamination, and random variation affect interpretation. Report an appropriate uncertainty interval alongside the estimate. If the range includes losses and worthwhile gains, the decision remains uncertain.
Turn incremental outcomes into incremental economics
Once the effect estimate is credible enough for the decision, translate incremental purchases into contribution before marketing costs. Deduct relevant product and delivery costs, returns, discounts, and other variable costs. Revenue alone does not pay the marketing bill.
Then subtract the incremental marketing cost and compare the remaining contribution with that cost. Include meaningful differences in creative, agency, technology, and execution costs. If the control also incurs marketing costs, compare the cost difference between groups. Do not quietly assume it costs zero.
For a test with unequal group sizes, compare rates and appropriately scaled costs rather than raw purchase totals. The economic comparison should represent the same-sized eligible population and the same time horizon. If the cost difference is zero or negative, a conventional positive-investment ROI ratio is not the right summary. Report the contribution and cost differences directly.
For the arithmetic, see how to calculate marketing ROI. The hard part here is establishing which outcomes belong in that calculation.
When volume is low, triangulate honestly
A low-volume business may not have enough conversions for a useful lift test. Combine CRM records, customer interviews, historical patterns, sales-cycle evidence, and carefully designed comparisons. Look for agreement and investigate contradictions.
Keep the claims separate. Your CRM is a business ledger. It can document customers, revenue, and recorded sources. It is not automatically a causal study. Start with a tracking system you can maintain, and be clear about what it can establish.
Report what happened, what you estimate marketing changed, what remains uncertain, and the next decision. That is a credible ROI case. Precision you cannot defend is just formatting.
Compare the evidence with the return your business needs. Use the marketing resources to strengthen your measurement approach, or work with me to build an ROI case your budget decisions can use.
