Hypothesis-led testing, measurement design, and learning velocity.
Experimentation programmes design and run A/B and multivariate tests on websites, landing pages, and product surfaces — but the operational discipline is hypothesis quality, measurement design, and learning capture, not test mechanics. The work covers MDE (minimum detectable effect) sizing to decide whether a test can produce a statistically meaningful answer at current traffic, primary-metric vs guardrail-metric separation, sample-ratio mismatch detection (early-warning that experiment infrastructure is broken), Bayesian vs frequentist analysis tradeoffs, and learning-velocity tracking that measures how fast hypotheses get answered rather than how many tests get run. Tools span Optimizely, AB Tasty, VWO, and Google Optimize successors (GrowthBook, Statsig).
Quick Answer
Hypothesis-led testing, measurement design, and learning velocity.
Conversion rate optimisation (CRO) is the structured improvement of user journeys — pages, forms, messaging, and measurement — to increase the rate at which visitors take valuable actions.
Landing Page Strategy
Message-matched pages that convert search intent into pipeline.
UX Audit
Friction mapping, trust gaps, hierarchy, and mobile review.
Definition
Discipline
The common failure in experimentation is not choosing the wrong variation. It is running a test that could not have detected the effect being looked for, then treating whatever number came out as an answer.
Given the traffic through that page and its current conversion rate, there is a smallest effect the test could reliably distinguish from noise. On a low-traffic page that number is often larger than any realistic improvement — which means the test cannot succeed, whatever the variation does. Establishing this first prevents weeks spent generating an unusable figure.
Pages without the traffic to test are not exempt from improvement; they are exempt from experimentation. Ship the change on the strength of the audit finding and monitor it, or test at a level where volume aggregates. Insisting on a test at every decision is how programmes end up with a backlog of inconclusive results.
One primary metric, chosen before the data arrives, plus a small set of guardrails that must not deteriorate. Stopping when the graph looks favourable, or picking the metric that moved after the fact, converts an experiment into a search for a flattering number — and both are extremely common.
If the two groups did not receive roughly the traffic they were supposed to, something in the assignment or the instrumentation is broken and the comparison is meaningless regardless of how clean the outcome looks. This check takes seconds and invalidates more tests than anyone expects.
Each test closes with what was believed, what was tried, what happened, and what that implies for what to try next — including the losses, which are usually more informative. A programme judged on tests run accumulates activity; one judged on questions answered accumulates knowledge, and the second gets cheaper over time.
Most experimentation programmes generate testing activity but not learning velocity — running tests that lack statistical power, measuring wrong primary metrics, or interpreting results without guardrail context. The discipline that separates compounding experimentation from theatrical experimentation is measurement design, not test creativity.
Search4online system
Delivery model
We map the technical base, demand signals, market position, and AI-search visibility before writing the plan.
We turn the diagnosis into structured content, clean technical foundations, authority campaigns, and conversion paths.
We measure rank, lead quality, AI mentions, and revenue signals so each month improves the last.
FAQ
MDE-dependent: detecting a 10% relative lift on a baseline 3% conversion rate at 95% confidence and 80% power needs roughly 15,000 visitors per variant. Below 5,000 visitors per variant per week, most experiments are underpowered and produce inconclusive results regardless of effect size. Smaller traffic sites should test bigger changes (full-page redesigns vs button colour) where larger MDEs are detectable with limited samples.
Sample-ratio mismatch (SRM) occurs when traffic doesn't split into variants in the proportions the experiment was configured for — e.g., 60/40 actual split on a 50/50 configured test. SRM is a critical early-warning that experiment infrastructure is broken (assignment logic bugs, bot filtering inconsistencies, redirect leaks). Tests with SRM produce uninterpretable results regardless of statistical significance; the test must be rerun after the cause is fixed.
Frequentist (p-values, confidence intervals) is standard, well-understood, and supported by every tool. Bayesian (posterior probabilities, expected loss) handles 'how likely is variant B better' questions more naturally and tolerates early-stopping decisions better. For organisations new to experimentation, frequentist is the safer foundation; Bayesian adds value once teams understand frequentist limitations and want to stop tests early without inflating false-positive rates.
Key takeaways
Related services
Message-matched pages that convert search intent into pipeline.
Friction mapping, trust gaps, hierarchy, and mobile review.
Paid media governance combines structured campaigns, conversion-quality signals, audience modelling, and creative testing so spend follows real sales value.
SEO services improve how a website is crawled, understood, trusted, and converted by combining technical SEO, content strategy, authority signals, and measurement.
Conversion rate optimisation (CRO) is the structured improvement of user journeys — pages, forms, messaging, and measurement — to increase the rate at which visitors take valuable actions.
