Conversion rate optimisation uses reliable measurement and controlled changes to improve useful customer actions while protecting commercial outcomes. Start with a specific barrier, define a meaningful success measure and test plausible changes. This guide provides a diagnostic worksheet and decision rules for improving conversion without ignoring lead quality.
What should a conversion optimisation programme improve?
Start with the action that matters commercially, then define exactly who had the opportunity to take it. A demo request, completed purchase and qualified sales enquiry need different measures. A higher button-click rate is useful only if the next steps also work.
For a lead-generation site, write the metric as: visitors who submit an accepted enquiry divided by eligible visitors exposed to the page. Count each visitor once within your chosen window. For a store, you might use purchasing visitors divided by eligible visitors. Do not switch between users, sessions and events midway through a comparison.
Pair the primary metric with a guardrail: a measure that must not deteriorate beyond an agreed limit. Suitable guardrails include lead acceptance, refund rate, contribution margin or form errors. The limit should reflect your economics and risk tolerance. It is not a universal industry threshold.
| Page | Details |
|---|---|
| Demo landing page | Useful primary outcome: Accepted enquiries per eligible visitor Guardrail to check: Sales acceptance and meeting attendance |
| Product page | Useful primary outcome: Purchasing visitors per eligible visitor Guardrail to check: Returns, margin and support questions |
| Pricing page | Useful primary outcome: Suitable sign-ups or enquiries per visitor Guardrail to check: Early cancellation or qualification mismatch |
| Checkout | Useful primary outcome: Completed purchases per checkout entrant Guardrail to check: Payment errors and duplicate orders |
How do you separate a measurement problem from a page problem?
Reconcile the events with the underlying business records before drawing conclusions from a funnel. A missing purchase event can look like checkout abandonment. A form event that fires on a click can make failed submissions look successful.
- Test a full journey. Check success, validation errors, back navigation and a repeat visit on mobile and desktop.
- Confirm event timing. A lead event should represent the accepted submission you defined, rather than merely pressing the submit button.
- Check duplicate counting. Reloading a confirmation page must not create an extra order or enquiry.
- Reconcile totals. Explain differences between analytics, orders and CRM records. Consent choices and blocked measurement can make analytics incomplete.
- Fix the denominator. Define eligibility, bot and employee exclusions, the attribution window and the reporting timezone before comparing versions.
Google’s GA4 event reference defines recommended events such as generate_lead, begin_checkout and purchase. These additional events require configuration. Their existence in the documentation does not confirm your implementation is correct.
Which page changes deserve investigation first?
Investigate a specific mismatch between the visitor’s question and the page’s answer. Use analytics to locate a pattern, then use task observation, support questions or sales feedback to develop explanations. A recording of someone hesitating shows behaviour; it does not reveal their motive.
| Observed pattern | Details |
|---|---|
| Ad visitors leave before the form | Question to investigate: Does the page fulfil the ad’s promise for this audience? Candidate change: Align the opening offer and qualification details |
| Product visitors repeatedly open delivery information | Question to investigate: Can they assess delivery and returns before committing? Candidate change: Make verified fulfilment details easier to find |
| Pricing visitors contact sales with basic questions | Question to investigate: Are inclusions and billing conditions clear? Candidate change: Clarify differences without hiding exclusions |
| Mobile form submissions fail | Question to investigate: Is there a reproducible input or validation error? Candidate change: Repair the failure and verify completion |
| Checkouts stall at payment | Question to investigate: Are total cost, payment availability or technical errors responsible? Candidate change: Investigate the specific failure before redesigning |
Fix a confirmed defect without waiting for a persuasive A/B result. Broken validation, inaccessible controls and incorrect information are repair work. Record the change and monitor it. Reserve experimentation for uncertain choices between acceptable experiences.
Performance deserves the same discipline. Review real-user loading and interaction measures by device, then reproduce the problem. A historical speed statistic cannot tell you what a particular performance change will do to your conversion rate. Our guide to Core Web Vitals explains the technical investigation.
If sales and marketing disagree about which enquiries count, use the SaaS campaign responsibility matrix to assign acceptance and measurement owners before testing.
How should you prioritise CRO experiments?
Use an evidence-and-decision worksheet instead of assigning precise-looking scores to guesses. Prioritise confirmed harm, then repeated friction with a testable explanation, then exploratory ideas. Record what would change your mind before implementation begins.
| Worksheet field | Illustrative completed entry |
|---|---|
| Observation and evidence | In a hypothetical task review, buyers ask what happens after requesting a demo |
| Alternative explanation | The visitors may be researching, or the traffic may be poorly matched |
| Hypothesis | Explaining the next step beside the form may reduce uncertainty for suitable prospects |
| Single change | Add accurate meeting expectations; leave the offer and qualification fields unchanged |
| Primary metric | Accepted enquiries per randomly assigned eligible visitor |
| Guardrails | Sales acceptance, meeting attendance and form error rate |
| Owner and decision | Named experiment owner; ship, reject or investigate using the pre-agreed analysis |
This is a proposed workflow and teaching example, not a PixelCrayons client result. Replace each entry with your evidence. If a candidate has no observation beyond “competitors do it”, keep it in the research backlog.
What makes an A/B test trustworthy?
A trustworthy test compares comparable groups under a plan written before the result is visible. Choose the randomisation unit, eligibility rules, primary metric, meaningful effect size, sample requirement and stopping rule. Use a statistical method appropriate to the experiment, with qualified analytical review when needed.
For a fixed-horizon test, do not stop early because the dashboard briefly turns positive. A sequential method needs its own valid stopping procedure. Include relevant business cycles and allow time for delayed outcomes such as sales acceptance. Repeatedly examining many segments and selecting the most flattering result creates false confidence.
Check that the observed allocation matches the planned split. A material mismatch can indicate broken assignment or measurement. Microsoft’s experimentation guidance explains why sample ratio mismatch must be investigated before interpreting effects.
Keep assignment stable for returning visitors where your experiment design requires it. Record overlapping tests, traffic-source changes and releases. Review the effect estimate and uncertainty alongside the guardrails. “No clear difference” means the test did not resolve the decision at its planned sensitivity; it does not prove the variants are identical.
Why can a higher conversion rate still be a worse result?
A conversion definition that ignores quality can reward the wrong change. The following invented example demonstrates the arithmetic. It is not a benchmark, an executed experiment or evidence of statistical significance.
| Illustrative measure | Details |
|---|---|
| Eligible visitors | Version A: 1,000 Version B: 1,000 |
| Submitted enquiries | Version A: 40 Version B: 55 |
| Submission rate | Version A: 4.0% Version B: 5.5% |
| Sales-accepted enquiries | Version A: 24 Version B: 22 |
| Accepted enquiries per visitor | Version A: 2.4% Version B: 2.2% |
Version B produces more forms but fewer accepted enquiries in these illustrative counts. That is a reason to examine qualification and downstream outcomes, not declare a winner from the submission rate. The correct decision would also require the planned sample, uncertainty analysis and outcome maturity.
What if you do not have enough traffic for an experiment?
Choose a method that can answer the question with the evidence available. Low traffic is a reason to narrow the question, not a reason to treat a small noisy test as proof.
- Observe representative users attempting a specific task and record where they cannot proceed.
- Review enquiry quality with sales and classify recurring questions.
- Inspect form failures and checkout errors directly.
- Improve inaccurate or inaccessible content, then monitor the release.
- Use before-and-after data as directional evidence while documenting seasonality, campaign changes and other competing explanations.
Keep a release note stating the baseline, change, affected audience and next review condition. Where commercial outcomes take months, monitor implementation quality first and revisit mature outcomes later. Avoid promising a lift by a fixed date.
For a store, the eCommerce revenue leak audit provides a wider checklist of journey problems to investigate. For a SaaS site, how SaaS and eCommerce buying cycles differ helps explain why accepted leads and delayed outcomes need their own measurement window.
What should you do next?
Choose one important page, validate its measurement and complete the worksheet. Bring the evidence to your designer, developer and analyst before choosing the treatment. Our conversion optimisation services and performance marketing work can support the agreed scope. Request a proposal with the page, business outcome and measurement constraints.
