Reading time: 11 minutes
To increase conversion rate, follow a CRO cycle. Measure where visitors leave, write a hypothesis, test one change at a time and repeat for each channel. That way, the cycle helps improve conversion rate and increase sales conversion rate, for search and AI traffic alike.
In short
- CRO is a cycle of measuring, diagnosing, testing and learning, not a list of tricks about buttons or colors.
- Diagnose before you test, because the GA4 funnel shows where visitors leave and session recordings show what they were doing there.
- With low traffic, test big changes, measure a step before the purchase and do not stop the test early.
- AI visitors are usually a small share of traffic, so measure them separately and rely on diagnosis more than on A/B tests.
What is CRO in digital marketing, and how does it help increase conversion rate?
CRO, short for conversion rate optimization, is the set of practices that raise a site’s conversions by making better use of the traffic it already has. The definition comes from RD Station, which sells marketing software. So, to increase sales conversion rate, the focus is not to attract more visits. It is to get more visitors to take the desired action.
CRO differs from SEO, because SEO and GEO bring the traffic, while CRO turns part of it into a lead or a sale. To prove the return of the whole, see the guide to SEO and GEO ROI. Before you optimize, however, find out whether your rate is good. The benchmarks for Brazil are in the guide to landing page conversion rate.
How do you increase conversion rate in 5 steps?
To increase conversion rate, follow five steps, because each one answers a question and leaves a record. RD Station sums up its process in three similar steps. That is, find opportunities and prioritize, model and run, and record what you learned.
- Measure: what is the rate of each channel? The step by step is in how to calculate website conversion rate by channel, including AI visits.
- Diagnose: where and why do visitors leave? For this, use the funnel and the recordings.
- Write the hypothesis: what should change, and why?
- Prioritize: which test comes first?
- Test: does the change win, or was the result chance?
How do you diagnose where the conversion rate drops?
To improve conversion rate, start with the GA4 funnel. Then watch the recordings of the steps where most people leave. The funnel shows where visitors leave, but only a recording shows what they were doing there.
In the GA4 funnel exploration, you define the steps, such as product, cart and payment. The funnel can be open or closed, and you can compare up to four segments. So create one segment per channel, because the search funnel can differ from the AI funnel.
Next, open the recordings. Microsoft Clarity rebuilds each session from the HTML and the visitor’s actions. It keeps recordings for 30 days, or up to 9 months if they are favorites. The session card also shows the entry page and the referral URL, which can help you spot visitors who came from AI. That works, however, only when the referrer arrives. The tool describes itself as free, and its documentation also covers masking of sensitive data.
Our criterion: watch 10 recordings of each step with a high exit rate before you write the hypothesis, because a single recording may be a rare case.
How do you write a hypothesis to increase conversion rate?
A testable hypothesis links a finding from the diagnosis to a change and a metric. Use this template, which is ours: “Because we saw [data], we believe [change] will make [segment] [action], and we will measure it by [metric] in [period].”
Here is an illustrative example, with fictional numbers. To increase sales conversion rate in a store, the hypothesis would be this. “Because 60% leave the 8-field form, we believe cutting it to 4 fields will make more people request a quote. We will measure it by the submission rate over four weeks.” Without data from the diagnosis, therefore, the test becomes a guess.
Which tests come first to improve conversion rate?
To increase sales conversion rate, fix first what is broken, without a test. Then test the hypotheses with high impact, low effort and enough traffic. Score each idea from 1 to 5 on those three criteria and start with the highest total. Optimizely uses similar criteria, because it asks for impact estimates, effort estimates and time to significance for each idea.
Our rule for what to do with each idea:
- Technical error, such as a form that fails or a slow page: fix it directly, without a test.
- Clear friction in a recording: fix it or test it, depending on volume.
- Promise, proof, price or form: run an A/B test.
- Whole layout or brand change: test only with high volume.
The scoring is also our criterion. Plan to lose as well. According to Kohavi and colleagues, who ran thousands of experiments at companies such as Amazon, Booking.com, LinkedIn and Microsoft, most of Bing’s tests fail. The ones that win improve key metrics by 0.1% to 1.0%, once diluted to the overall impact. Bing operates at a much larger scale than a typical store. Still, the lesson holds, since every test yields learning, even without a win.
How do you run an A/B test with low traffic?
To increase sales conversion rate with low traffic, test big changes and measure a step before the final conversion. Run the test for full weeks, without stopping early. Optimizely recommends high-impact changes and micro-conversions for sites with few conversions. If the sample still does not fit, swap the test for a qualitative diagnosis.
How many visitors does an A/B test need to measure conversion rate?
Use this rule of thumb: n = 16 × p × (1 − p) ÷ δ², per variant. Here, p is the current rate, and δ is the smallest absolute difference you want to detect. Evan Miller presents the formula as a rule of thumb, not as an exact value. To check the number, use his calculator, which starts from 80% power and 5% significance.
Here is an illustrative example. A page converts 3%, and you want to detect a jump to 3.6%, that is, a 20% relative gain. The rule gives about 12,900 visitors per variant, and Evan Miller’s calculator gives about 13,000. That is close to 26,000 in total, so with 4,000 visits a month the test would run for more than six months.
To detect a 50% jump, from 3% to 4.5%, the rule gives about 2,100 visitors per variant, and the calculator gives about 2,200. That is close to 4,300 in total, and the test takes about a month with the same 4,000 visits. This is why Optimizely says that bigger changes need less traffic to be detected.
What if the traffic is not enough?
Follow these tips from Optimizely for low-traffic sites:
- Test high-impact changes: a small change, such as the text of a button, needs many more visits than a big change, such as a new offer.
- Measure micro-conversions: clicks on “add to cart” reach the sample before orders do.
- Avoid niche tests: prefer the landing page or a sitewide banner to a product page that gets few visits.
- Avoid multivariate tests: each extra variation needs more traffic.
How do you run the test without distorting the conversion rate?
Change one thing only, set the sample size beforehand and run full weeks. RD Station asks that version B change a single element and that the test last at least a week, because of seasonality. Optimizely recommends at least one business cycle, or seven days.
Do not stop the test when the dashboard shows significance. Evan Miller calculates a worst case. In a test with no real effect, with a 50% rate and up to 150 observations, the result is checked after every observation. So, if you stop at the first 5% significance, the false positive rate reaches 26.1%, not 5%. Likewise, distrust results that look too good, because Kohavi and colleagues recommend looking for a measurement error before celebrating.
Which tool should you use after Google Optimize?
Google shut down Optimize on September 30, 2023. According to Google Analytics Help, the company is collaborating on GA4 integrations with AB Tasty, Optimizely and VWO. To choose, check each tool’s page for the current price and the GA4 integration.
How does AI help improve conversion rate?
AI saves time on analysis, but it does not replace the test. It summarizes recordings, groups patterns and suggests hypotheses. You decide and measure. Microsoft Clarity, for example, has Copilot, which summarizes sessions and heatmaps.
Microsoft itself warns that generative models can misinterpret or produce incorrect information. So treat the summary as triage. Our criteria:
- Use AI to choose what to watch: then watch a sample of the recordings to confirm.
- Treat the AI’s hypothesis like any other: it goes into the scoring and goes to a test only if it wins.
- Do not let AI declare the winner: the decision comes from the planned sample.
- Avoid personal data in external chats: export only numbers and page text.
What changes in conversion rate for visitors who arrive from AI?
The method to improve conversion rate is the same, but the diagnosis must separate the channel. Also, an A/B test with AI visitors alone rarely fills the sample, because their share tends to be small. Segment AI in the funnel and use the recordings to decide what to change. The studies on this visitor disagree, so measure yours.
- Segment the channel: create the AI channel group, as the guide to AI traffic in GA4 shows, and open the funnel with that group only.
- Check the landing page: see in the recordings whether the visitor landed on the page the answer promised. For stores, the guide to GEO for e-commerce shows how products enter the answers.
- Continue the conversation: to increase sales conversion rate for this visitor, repeat the promise, show price and proof, and make comparison easy. For ads, see how to advertise on ChatGPT.
- Test for everyone: run the test on the whole page and then read the result by channel. With few AI visits, treat this reading as a clue, not as a conclusion.
Which common mistakes hurt conversion rate?
The most common mistakes when trying to improve conversion rate are testing everything without a plan and copying what worked on another site. Another mistake is treating low-traffic pages like high-traffic ones. RD Station lists these three, plus testing random items or items with no impact. Other mistakes weigh just as much:
- Measuring the wrong conversion: counting duplicated events, not sessions, inflates the rate, and a single average hides the channel that converts worst.
- Stopping the test early: the dashboard shows significance before its time, as seen above.
- Changing several things at once: that way, nobody knows what worked.
- Ignoring the calendar: holidays and campaigns change the traffic and the type of visitor.
- Not recording what you learned: without a record, the team repeats the same test months later.
Frequently asked questions
Not for diagnosis. Microsoft Clarity describes itself as free forever, and the GA4 funnel already shows where visitors leave. For A/B tests, Google shut down Optimize and points to third-party tools integrated with GA4. The test only pays off with enough traffic, and the current price is on each tool’s page.
It lasts at least one full week and then until the planned sample. Optimizely and RD Station recommend that minimum. In this guide’s illustrative example, with 3% conversion and 4,000 visits a month, detecting a 20% gain would take more than six months.
Yes. To improve conversion rate and increase sales conversion rate with low traffic, fix what the funnel and the recordings show. Then compare before and after. This comparison is weaker than a test, because seasonality and channel change between periods, so record the date of each change.
Want to know how your brand shows up in AI answers today?
NotHuman is opening spots for agencies and brands to track how they appear in AI answers. Join the waitlist and we’ll let you know when a spot opens up for you.
Main sources
- RD Station, CRO: what it is and how to do it, in Portuguese (2022).
- Google Analytics, funnel exploration (2026).
- Microsoft Learn, Clarity recordings overview (2026).
- Google Analytics, Google Optimize shutdown (2023).
- Optimizely, test tips for low-traffic sites (2026).
- Optimizely, how long to run an experiment (2026).
- Evan Miller, how not to run an A/B test (2010).
- Kohavi and colleagues, seven rules of thumb for web site experimenters (2014).
