53 conversion ideas, sorted by whether you can actually test them

More than most sites have. Across a dataset of 4,200 tests the median requirement was 10,400 visitors per variant server-side and 14,800 client-side. The exact figure depends on your baseline conversion rate, the relative lift you want to detect, and your confidence and power settings. At 95% confidence and 80% power, a page converting at 4% needs roughly 10,300 visitors per variant to detect a 20% relative lift, and about 39,500 to detect a 10% one. Be careful with the simplified rules of thumb circulating online, which often quote figures two to three times lower because they use less conservative assumptions. The calculator above uses the standard two-sided formula, which is why its numbers look larger and why they line up with that 14,800 median.

Around 12% on Optimizely's analysis of more than 127,000 experiments, meaning roughly one tested idea in eight produces a statistically significant improvement. Other platform benchmarks land nearby: VWO around 14%, CXL and Convert around 20%. An independent audit of 2,288 tests found 50.5% beat control on raw numbers but only 19.1% reached statistical significance. Mature programs with disciplined hypothesis generation reach 22 to 30%. If someone quotes you a much higher win rate, ask how they define a win.

At minimum 14 days, covering at least one full business cycle including weekends, and beyond that whatever your pre-calculated sample size requires. Tests under 14 days carry a false positive rate around 61%. The median test in one large agency dataset ran 42 days. Set the duration before you launch and commit to it, because checking daily and stopping when the variant looks good is reportedly done by 47% of teams and is the most reliable way to manufacture a winner that does not exist.

Around 6.1% at the median for tests that win. Lifts above 50% sit at the 95th to 99th percentile, which means they happen occasionally and cannot be planned for. The value in a testing program comes from compounding rather than from any single result: a 6.1% median lift across eight to ten winning tests a year on a checkout funnel produces something in the region of 47 to 61% total improvement. Any agency forecasting case-study lifts on most tests is mis-calibrated or selling.

Most likely the winner's curse. Under-powered tests systematically exaggerate the size of the effects they detect, because only unusually large swings cross the significance threshold when the sample is small, and extreme results drift toward the average over time. In one dataset 41.4% of tests claimed significance with insufficient statistical power. If you tested on 200 conversions rather than 2,000, expect both false positives and inflated lift estimates, and expect implemented winners to underperform their test results.

Then stop trying, which is a legitimate strategy rather than a defeat. Fix the things that are already known to be broken, such as contrast, form labels, page speed, unclear actions and missing trust information, none of which need an experiment to justify. Measure at the business level, where a change from eleven inquiries a month to nineteen is visible without a significance calculation. Use session recordings and support questions to find problems. And run holdouts where you control the audience, such as email flows, which works at far lower volumes than page testing.

Copy and social proof tests show the highest raw win rates, around 57 to 60% in one audit, while form, layout and trust-signal tests win less often. By funnel stage the differences are smaller than people assume: one agency dataset found decision-stage tests on product, cart and checkout pages at 37.5%, consideration-stage at 34.9% and awareness-stage at 35.5%. Lead generation and subscription funnels win more often than mature ecommerce, at roughly 58 to 61% raw against 48 to 50%, mostly because they still carry visible friction that has never been removed.

If you have the traffic, yes, because isolating variables is the entire point. If you do not have the traffic, batching related changes is the more honest approach: you will not be able to attribute the effect to a specific change either way, so shipping a coherent group of fixes and measuring the total beats shipping nothing for eight months. What you should not do is make many changes, run an underpowered test, and then claim to know which one worked.