Conversion rate is the number small online retailers check most and understand least. A single percentage hides who visited, what they wanted and where they gave up, and chasing it directly tends to produce random changes to button colours. This guide sets out a more useful sequence: define the metric, find the leaks, segment, use the data you are probably ignoring, and test changes only when you have enough traffic for the result to mean something.
Define conversion before you try to improve it
“Conversion rate” can be calculated in several defensible ways, and different tools default to different ones.
- Per session or per user? Orders divided by sessions rewards persuading people on a single visit. Orders divided by users is closer to “what share of interested people bought”, which matters for considered purchases such as furniture, where shoppers return several times.
- Which orders? Decide whether cancelled, fraudulent, refunded and test orders count. A shop with high return rates may find that “net” conversion tells a different story.
- Which traffic? Bot traffic and internal visits deflate the rate. Traffic from a newsletter will behave differently from a cold social audience, so a blended figure can rise or fall simply because the mix changed.
- Which consent state? In Europe, analytics typically only sees visitors who accept analytics cookies. Google’s consent mode can model some of the gap, but your analytics conversion rate is an estimate, not a census. Your platform’s order count is the ground truth for revenue; our guide to cookie consent and analytics for ecommerce explains why the numbers diverge.
Write your definition down and keep it stable. A metric that changes definition every quarter cannot show improvement.
Funnel analysis: find where people leave
A funnel breaks the journey into steps: product view, add to basket, begin checkout, add delivery details, add payment details, purchase. In GA4 these correspond to the recommended ecommerce events (view_item, add_to_cart, begin_checkout, add_shipping_info, add_payment_info, purchase), which are not sent automatically: your platform integration or developer must implement them. Once they are in place, a funnel exploration shows completion and drop-off between each step.
Illustrative worked example
The following figures are invented round numbers for illustration, not benchmarks.
A hypothetical UK homewares shop has 20,000 sessions in a month:
| Step |
Sessions |
Continue to next step |
| Viewed a product |
8,000 |
20% |
| Added to basket |
1,600 |
56% |
| Began checkout |
900 |
44% |
| Purchased |
400 |
— |
Session conversion rate: 400 ÷ 20,000 = 2%.
Where to look first? The temptation is the product-to-basket step because the percentage is lowest, but that step always has the most browsing intent mixed in. The more telling figure is that fewer than half the people who began checkout finished. Those shoppers had chosen products and committed to buying. If the shop could move checkout completion from 44% to 55% with no other change, purchases would rise from about 400 to about 495: almost a quarter more orders from the same traffic. Improving the earliest step by the same proportion is usually much harder.
The point of the example is method, not numbers: compare the steps, then ask which leak involves the most committed shoppers and the most fixable causes.
Segment: find who is leaving
An average funnel blends very different visitors. Split it by:
- Device. Mobile checkout problems (tiny fields, keyboards covering buttons, payment pop-ups that fail) are a classic hidden leak.
- Traffic source. Paid social traffic converting at a fraction of email traffic may be a targeting issue, not a site issue.
- New versus returning. Low first-visit conversion is normal for considered purchases.
- Country. Cross-border shoppers may abandon on currency, delivery cost or a missing local payment method.
- Landing page. A campaign sending people to the home page rather than the advertised product wastes intent.
Keep segments large enough to be meaningful. A segment of 40 sessions will produce dramatic percentages that mean nothing.
Use your site-search data
Visitors who search are telling you what they want in their own words. GA4’s enhanced measurement records a view_search_results event when it detects a search parameter in the URL (by default q, s, search, query or keyword; others can be added), and many ecommerce platforms log searches too.
Review monthly:
- Zero-result searches. Add synonyms (“sofa” versus “couch”), fix misspellings, or note demand for products you do not stock.
- High-exit searches. Results exist but shoppers leave, which suggests the results are poor or poorly ordered.
- Search conversion. Compare conversion of searchers and non-searchers. If searchers convert well, make search more prominent on mobile.
Remove checkout friction
Checkout problems are usually prosaic. Work through your own checkout on a mid-range phone, on mobile data, as a new customer, and note every moment of doubt:
- forced account creation before purchase;
- delivery costs or dates revealed only at the final step;
- form fields that reject valid input (postcodes with spaces, international phone formats);
- error messages that clear the form;
- missing payment methods your customers expect in their country;
- discount-code fields that send shoppers away to hunt for codes;
- fraud checks or 3-D Secure steps that fail without explanation.
Some friction is deliberate and necessary; strong customer authentication and fraud screening protect you. The aim is to make necessary steps clear, not to remove them. Our guide to payment fraud prevention for smaller businesses covers that balance, and clear delivery and returns information is part of building customer trust.
Page speed and Core Web Vitals
Speed matters most where shoppers are most impatient: product listings and checkout on mobile. Google’s Core Web Vitals give three measurable targets for a “good” experience, assessed at the 75th percentile of page loads, separately for mobile and desktop:
- Largest Contentful Paint (LCP), loading: 2.5 seconds or less.
- Interaction to Next Paint (INP), responsiveness: 200 milliseconds or less. INP replaced First Input Delay as a Core Web Vital in 2024.
- Cumulative Layout Shift (CLS), visual stability: 0.1 or less.
On a typical shop, oversized hero images hurt LCP, heavy third-party scripts (chat widgets, review carousels, multiple tracking tags) hurt INP, and late-loading banners or images without set dimensions cause layout shift, which is especially damaging when an “add to basket” button moves under a thumb.
Google’s Search Central documentation says good Core Web Vitals align with what its ranking systems seek to reward, but it does not promise rankings. Treat them primarily as a user-experience measure. Use field data from real visitors where available, since lab tests on a fast office connection flatter most sites.
A/B testing without fooling yourself
An A/B test randomly splits visitors between the current version and a change, then compares outcomes. It is the right tool for settling a specific question, but small shops routinely run tests that cannot produce a reliable answer.
The sample-size problem
The smaller the effect you want to detect and the lower your baseline conversion rate, the more visitors you need. Illustrative example: with a 2% baseline conversion rate, detecting a 10% relative improvement (2.0% to 2.2%) at conventional settings of 95% confidence and 80% power needs roughly 80,000 visitors per variant, about 160,000 in total. A shop with 20,000 sessions a month would need around eight months. Detecting a 20% improvement (2.0% to 2.4%) needs roughly 21,000 per variant, about two months at the same traffic.
Use a sample-size calculator before you start, and if the required run time is unrealistic, test bolder changes or use other methods.
Other common pitfalls
- Stopping early. Checking daily and stopping when the result “looks significant” inflates false positives dramatically. Fix the sample size or duration in advance.
- Testing too many things at once. With many variants or metrics, something will look significant by chance.
- Ignoring weekly cycles. Run tests in whole weeks so weekday and weekend behaviour are both represented.
- Novelty effects. Returning customers may click a new design simply because it is new.
- Mismatched consent. If consent banners or tracking differ between variants, the comparison is broken.
Qualitative research: learn why
Analytics shows what happened; it rarely shows why. Cheap qualitative methods fill the gap:
- Watch five people shop. Ask friends-of-friends who match your customers to buy something while thinking aloud. Problems that never show up in reports become obvious within minutes.
- One-question surveys. On the order confirmation page: “What nearly stopped you buying today?” On exit intent: “What were you looking for?”
- Read customer-service messages. Repeated pre-purchase questions about sizing or delivery show what product pages fail to answer.
- Session recordings, used sparingly and only with appropriate consent and masking of personal data.
What to do next
- Write down your conversion definition and check your ecommerce events are firing correctly.
- Build a funnel and identify the step where committed shoppers leave.
- Segment that step by device and source.
- Review a month of site-search terms, starting with zero-result searches.
- Check Core Web Vitals field data for product and checkout pages on mobile.
- Watch five people use your checkout.
- Only then decide what is worth an A/B test, and calculate whether you have the traffic for it.
If you are still choosing a measurement tool, our comparison of GA4, Matomo, Plausible and platform dashboards sets out the trade-offs.