UX/UI design
CRO: diagnose and test without inventing the cause
Short answer
CRO improves decisions before it claims to improve conversion.
Conversion rate optimisation is a decision process: establish an observable problem, separate it from its possible causes, choose evidence proportionate to the risk, protect guardrails and record whether the change is kept, rejected or still unresolved.
CRO does not turn low traffic into strong evidence. When the sample cannot support a controlled comparison, the correct work is to remove verified defects, test critical tasks, inspect qualitative evidence or retain an insufficient-data status.
Editorial boundary
The funnel owns the denominator; CRO owns the hypothesis and decision.
| Object | Funnel article | This CRO article |
|---|---|---|
| Transition | Population, event, source, window and denominator. | Uses the transition without redefining it after seeing the result. |
| Problem | Shows where an observable loss occurs. | Builds competing explanations and chooses the next proof. |
| Change | Measures the same transition consistently. | Defines variation, guardrails and decision rule. |
Evidence chain
Do not skip from a symptom to a redesign.
Observe
Describe the state without explaining it
Fictional example—not Edikka data: 8 of 20 eligible requests stopped before a valid server response. The number would be observable if that transition were instrumented; the reason would not.
Hypothesise
Write competing explanations
Form error, missing reassurance, mismatch of intent, mobile defect and deliberate qualification should not be collapsed into one story.
Choose
Select the lightest evidence that can change the decision
A technical retest, user task, sales review and controlled experiment answer different questions.
Decide
Keep, reject, iterate or leave unresolved
The record preserves the metric, guardrails, limits and date—not only the winning narrative.
Method selection
The strength of the claim determines the method.
| Question | Appropriate evidence | What it cannot prove alone |
|---|---|---|
| Is the interface technically broken? | Reproduction, DOM/network evidence and dated retest. | Business impact. |
| Can a person complete the task? | Moderated or unmoderated task test with context. | Population-wide conversion uplift. |
| Where does the measured flow lose cases? | Funnel counts with stable definitions. | Cause of the loss. |
| Which version changes the primary outcome? | Controlled experiment with a pre-written rule and adequate sample. | Long-term value unless it is observed. |
| Are enquiries better for the business? | Qualification and downstream commercial statuses. | Interface causality without a comparison design. |
Hypothesis contract
A usable hypothesis contains seven decisions.
Observed problem
State the evidence and affected population
Avoid adjectives that already contain a cause.
Proposed change
Name what changes and what stays fixed
The variation must be reversible and inspectable.
Expected observation
Define primary metric, direction and window
Use the funnel’s existing denominator.
Decision rule
Keep, reject, iterate or insufficient data
Write the rule before reading the result.
The other three decisions are the guardrails, known confounders and accountable owner. Without them, a hypothesis is a prediction without a governance mechanism.
Prioritisation
Priority is the cost of being wrong, not a decorative score.
| Question | Why it matters |
|---|---|
| Is a requirement or critical task currently blocked? | Verified accessibility, security or functional defects can require correction without experimentation. |
| How many eligible cases meet the problem? | Reach must use the affected population, not total traffic. |
| What evidence could reverse the proposed decision? | A non-falsifiable idea is not ready to test. |
| What could improve locally while harming the business? | Clicks can rise while lead quality, trust, performance or accessibility falls. |
Low traffic
When traffic is scarce, change the method—not the confidence label.
An underpowered A/B test can produce a winner-shaped number without a useful decision. For a low-volume B2B site, start with deterministic defects, critical-task observation, enquiry quality, sales objections and reversible before/after changes whose attribution limit is explicit.
Use four states: supported, contradicted, insufficient data and out of scope. “No significant result” and “no test was possible” are not the same state.
Guardrails
A local conversion gain is rejected if it damages the service.
- Qualified enquiries and downstream outcome, not only CTA clicks.
- Form completion errors and successful server responses.
- WCAG 2.2 and applicable accessibility obligations.
- Core Web Vitals and transferred page weight when the interface changes.
- Privacy, purpose limitation and absence of unnecessary tracking.
- Support burden, reversibility and consistency between FR and EN.
Non-negotiable
A known inaccessible state is corrected and retested; it is not kept as an experimental variant to “measure its impact”. Compliance evidence and behavioural evidence answer different questions.
Edikka · public case
The current result is insufficient data, and the article keeps it visible.
Edikka now records a valid contact submission after server validation. It does not yet have a complete exposure denominator or a mature public sequence of qualification and contract outcomes. Therefore this edition demonstrates the method but claims no conversion uplift.
This negative result is operational: no A/B test is launched, no winner is selected and no causal statement is published. Verified usability or accessibility defects may still be corrected, dated and retested without pretending that the correction proves a business gain.
Decision record
A CRO deliverable ends with a decision that another person can audit.
| Field | Content |
|---|---|
| Observation | Raw counts, window, population and source. |
| Hypotheses | Competing explanations, including the null explanation. |
| Method | Protocol, variant, sample rule and known limits. |
| Outcome | Primary metric plus every guardrail. |
| Decision | Keep, reject, iterate, insufficient data or out of scope. |
| Traceability | Owner, date, version, evidence and next review. |
Architecture
Continue with the resource that owns the missing evidence.
Primary references
Standards and public methods frame the evidence; they do not certify an uplift.
- GOV.UK Service Manual · Plan user research — relates research questions to user groups and the method most capable of answering them.
- W3C · WCAG 2.2 — supplies testable accessibility requirements used as guardrails.
- web.dev · Web Vitals — defines user-centred performance metrics used as technical guardrails.
- Google Analytics · Recommended events — provides interoperable lead-event names; event naming does not establish causal impact.
Method limit: this is an Edikka decision framework, not a GOV.UK, W3C or Google standard. Version 1.0 was reviewed on 4 September 2026. Next review: 4 December 2026.
Optimisation is not about producing green arrows. It is about reducing the cost of a wrong decision.
A conversion signal starts an investigation; it does not certify its cause. The method must fit the decision while protecting the quality of the service.
Separate the signal from its cause
A drop, an exit or an error is observable. Its explanation remains a hypothesis until an appropriate method tests it.
Match proof to the decision
Technical tests, task observation and controlled experiments answer different questions. Their conclusions must not be merged.
Protect what a click cannot measure
Accessibility, performance, privacy and lead quality remain part of the decision, even when the primary metric improves.
When evidence is insufficient, leaving the decision open is a result—not a failure of the method.
Optimise without manufacturing certainty
Eight answers about diagnosis, hypotheses, methods, guardrails and attribution limits.
