About conversion rate optimization
The biggest CRO gains usually come from fixing customer uncertainty, product discovery, shipping/returns clarity, checkout friction and site speed — not from changing button colours.
What we optimize
Not just conversion rate. A test that lifts conversion but lowers AOV or margin can be a bad test. We track purchase conversion, add-to-cart, checkout starts, revenue per visitor and AOV, and use the latter two as guardrails.
The process Proofsplit automates
- Measure the funnel
- Find the biggest drop-off
- Investigate why (evidence)
- Form a hypothesis
- Prioritize: reach × impact × confidence ÷ effort
- Run the experiment
- Check data quality (SRM)
- Decide with statistical and business criteria
- Ship the winner
- Monitor, and feed the result into the next analysis
Reading your funnel
| Symptom | Likely investigation |
|---|---|
| High bounce rate | Message mismatch, slow load, weak value proposition |
| Low product-page engagement | Poor navigation, weak product cards, bad traffic quality |
| Good product-page engagement but low add-to-cart | Unclear offer, price, images, variants, trust, shipping |
| Good add-to-cart but low checkout start | Cart friction, surprise costs, weak checkout CTA |
| Good checkout start but low purchase rate | Payment, shipping, fields, errors, trust, technical bugs |
| Good conversion but low profit | Discounting, low AOV, shipping costs, returns |
Rules we enforce
- Do not test without a hypothesis.
- Do not treat statistical significance as business significance.
- Do not stop fixed-horizon tests early.
- Do not optimize only for clicks — the primary metric decides, secondary metrics explain.
- Do not ignore AOV, margins, returns, or refunds.
- Do not let users switch between variants (sticky assignment).
- Do not change traffic allocation mid-test.
- Do not run overlapping experiments on the same surface.
- Do not call an underpowered test a winner.
- Do not forget to ship and monitor the winner.
The simplest good setup
One control, one treatment, 50/50 split, one primary metric, 2–4 guardrails, at least 14 days, pre-calculated sample size, no early stopping, and a decision based on statistical evidence and profit.