Proof mode
An attributed-revenue number cannot tell you whether those sales would have happened anyway. Proof mode is the holdout test that answers that question instead of skipping it.
The problem it solves
A dashboard that only reports "revenue attributed to Clinch" answers the wrong question. It counts every order that used a Clinch code, but some of those shoppers would have bought anyway, discount or not. Counting them as recovered inflates the number without proving anything about whether the offer caused the sale.
How it works
Clinch holds the offer back from a fixed, random share of leaving shoppers, 10% by default and adjustable in Settings, and shows them nothing at all. Everyone else sees the offer as usual. Because the holdout group is chosen at random, comparing the purchase rate of the two groups isolates the effect of showing the offer from everything else happening on the store that week.
Reading the result
The comparison is a two-proportion z-test at 95% confidence, with a Wilson interval on each individual rate so a small sample never implies a conversion rate below zero. Clinch will not print a lift number until the smaller of the two groups has at least around 200 sessions; below that, small-sample noise can move a rate by whole percentage points on its own. Once it does report, it always shows the gap alongside its margin of error, for example "ordered 2.3 points more often, give or take 1.1", and if the gap sits inside that margin, it says so plainly instead of rounding up to a headline number.
Related
How individual recovered orders are attributed and refunds handled · Why measuring incrementality matters for recovery generally
Common questions
Does proof mode cost me sales?
It withholds the offer from a share of leaving shoppers, 10% by default, so yes, a small number of shoppers who might have used a code do not see one. That is the cost of having a real answer to whether the offer works, rather than a guess.
Why does it say 'not enough data yet'?
Below about 200 sessions in the smaller of the two groups, a conversion-rate difference is too easily explained by chance to be worth reporting. Clinch waits until both groups have enough sessions before printing a number at all.
Can I turn proof mode off?
You can lower the holdout share, including to zero. At zero every shopper sees the offer and nothing is withheld, but Clinch then has nothing to compare against and stops being able to measure lift.
What does the reported lift actually mean?
It is the difference in purchase rate between shoppers who saw the offer and the holdout group who did not, shown with a margin of error at 95% confidence. A result inside its own margin of error is reported as no clear difference rather than as a positive number.
Try it on your own store
Clinch is free to install. The Free plan costs $0 per month and takes 10% of revenue it actually recovers, so if it recovers nothing you pay nothing.
Add to Shopify · Free to install