Check Your Website Security Online
Test your own site yourself: certificate, security headers, outdated software, exposed admin areas. Which free tools do what, and what they cannot see.
Read the articleNovember 4, 2026 · 12 min read
Conversion rate optimization means raising the share of visitors who do what your page is there for: buy, enquire or sign up. That one action is the conversion, and the conversion rate is its share: conversions divided by visits, times 100. The process has five steps: set the goal, measure the baseline, write down a hypothesis, test, evaluate. The usual levers are the offer, the first screen, the form, load time and trust. An A/B test shows whether a change works, but only with enough visitors, and you work out how many before you start. On small sites, often only big changes are measurable in reasonable time.
How to build the page is covered in Create a Landing Page: Guide and Tools. This article is about what comes after: measuring whether it works, and changing it for a reason. To have it built instead, see Hiring a Landing Page Designer: Steps, Cost. In my business, Sebastian, my AI employee for SEO, GEO and blogging, does the measuring; more below.
You decide what counts as a conversion. Google's Ads help: "You choose what you identify as valuable, like a purchase, sign-up, or phone call." In Google Analytics the same thing is now called a key event: "an event that measures an action that's particularly important to the success of your business."
Conversion rate = conversions ÷ visits × 100
An example: suppose an offer page gets 1,500 visits a month and 45 enquiries. The conversion rate is 45 ÷ 1,500 = 3%.
Two decisions come first:
Start where your numbers point: drop-outs before the button suggest the offer and first screen, drop-outs in the form suggest the form.
There is no general minimum number of visitors; it depends on your baseline and the difference you want to detect. In their survey of online experiments, Kohavi and colleagues give a formula for the minimum sample per variant at 95% confidence and 80% power:
n = 16 × p × (1 - p) ÷ Δ²
Here, p is the original's conversion rate and Δ the difference to detect, as a fraction (0.6 percentage points = 0.006). Using the example: 3% baseline, 50 visits a day, split evenly.
| Variant goal | Δ | n per variant | Total visits | Run time at 50 a day |
|---|---|---|---|---|
| 3.0% to 3.6% (+20%) | 0.6 points | 16 × 0.03 × 0.97 ÷ 0.006² = 12,933 | 25,868 | about 517 days, almost 74 weeks |
| 3.0% to 4.5% (+50%) | 1.5 points | 2,069 | 4,140 | 83 days, so 12 weeks |
| 3.0% to 6.0% (+100%) | 3.0 points | 517 | 1,036 | 21 days, so 3 weeks |
The formula is an approximation. The full two-proportion formula many calculators use gives 2,514 instead of 2,069 per variant in the middle case, so treat the table as a lower bound.
Run time = total visits ÷ visits per day, rounded up to whole weeks. Kohavi and colleagues recommend at least a week or two, then multiples of a week, because weekend visitors often differ from weekday ones, and a 50/50 split: the smaller the variant's share, the longer the test.
At 50 visits a day, +20% is practically unmeasurable. What still works:
Stopping too early. Keep checking and stop at the first "significant", and you find winners that aren't. Johari and colleagues call the usual p-values "wholly unreliable" then. A simulation for this article shows how much: an A/A test, two identical variants at 3% each, 25 visits a day per variant, 12 weeks, 20,000 runs. Evaluated once at the end, 5.1% of runs declared a winner, as expected at 95% confidence. Stopping at the first hit, it was 20.5% with weekly checks and 32.6% with daily checks. With identical variants, every such winner is chance.
Too many variants. At 95% confidence, each comparison with the original has a 5% false-alarm rate. With three variants, the chance of at least one false winner is 1 - 0.95³ = 14%. Each extra variant also splits your visitors further. With few visitors: one variant against the original.
The wrong metric. More clicks are not automatically more enquiries, nor more enquiries more orders. Removing the price may bring more enquiries, but more that don't fit. Decide beforehand on the metric closest to the goal that is reachable within the run time, and watch the other stages as a check. If the visitors come from ads, revenue per ad dollar belongs in the picture too, see ROAS Formula: Calculation and Worked Examples.
Since 19 September 2026 I measure, cookie-free, which button on which page gets clicked on kevinwelter.com. Every button into my community carries an event named after its placement, and the same name is in the link. That shows me per page whether the button at the top or the one in the closing section produced the click: the intermediate stage from the first section. Where that stage sits on the whole path to the community is shown in Marketing Funnel Examples: Stages in Practice.
Measuring is Sebastian's job in my business, in a fixed chain: measurement, finding, action, remeasurement with the same setup. Every figure comes from a saved tool result in his own project folder. He measures load times and Core Web Vitals with Lighthouse and Unlighthouse and pulls real-world field data through the PageSpeed Insights API. He puts nothing live himself; I approve that. How he works is on AI for SEO and Blogging: How Sebastian Works.
One better than your own previous rate, measured the same way. A general target helps little, because the rate depends on offer, price, traffic source and what you count as a conversion: a first consultation and a newsletter sign-up are not comparable.
Conversions divided by visits in the same period, times 100: 45 enquiries from 1,500 visits are 3%. Numerator and denominator must match: same page, same period, without your own test visits.
Work it out: with n = 16 × p × (1 - p) ÷ Δ², a 3% baseline needs about 520 visits per variant to detect a doubling, almost 13,000 for +20%. Divide the total by your daily visits to see whether the test fits your time frame. How to get more visitors from search in the first place is covered in SEO Basics for Small Businesses.
Only once the calculation shows you have enough visitors. Google Optimize has not been available since 30 September 2023; Google names AB Tasty, Optimizely and VWO as providers with a Google Analytics integration. For ads, Meta has its own A/B test, described in How to Run Ads on Meta: A Guide.
Write three lines for your most important page: the metric, its current value and the smallest difference worth having. Then work out the run time. If it doesn't fit into three months, test a bigger change or measure a nearer step. If you'd rather not build the measuring yourself: there's a finished package for Sebastian in my community AI Employee Framework, with the tool chain and the project folder structure ready to rebuild.
Test your own site yourself: certificate, security headers, outdated software, exposed admin areas. Which free tools do what, and what they cannot see.
Read the articleHow a commissioned security test runs: scope and written authorisation, stages from outside in, how findings are rated, the report, and the retest.
Read the article
Kevin Welter
Developer, IT architect, author of technical books (Kubernetes, cloud infrastructures) and speaker. Runs his business with an AI workforce of fifteen AI employees and shows solo business owners in his community how to hire their first AI employee.
Your first AI employee up and running within an hour
Join the community$5/month founding price, first 50 members