Conversion Rate Optimization

A Definition, and Why B2B Breaks the Standard Playbook

Conversion rate is the number of visitors who complete a defined action divided by the number who could have, over a stated window. Conversion rate optimisation is the practice of raising that proportion deliberately, using research and controlled comparison rather than opinion. Three things are load-bearing: the action is defined in advance, the denominator is stated, and the improvement is demonstrated rather than assumed. The discipline is routinely reduced to A/B testing, which is one method among several, and to redesign, which is a project rather than a practice.

The awkward part in B2B is deciding what to count. If the sales cycle runs nine months, the outcome that matters — signed revenue — is unavailable when the decision has to be made. The workable arrangement is a primary conversion at the handover point, a qualified enquiry or a demo request, recorded with enough source detail to tie it to revenue later, with everything upstream treated as diagnostic.

Four structural features then break a playbook written for consumer volumes and same-session purchases.

• The denominators are tiny. A few hundred visitors a month to the page that matters is normal, and machinery designed for tens of thousands does not degrade gracefully at that scale — it produces confident nonsense.
• The outcome arrives quarters later. A change made in March is judged on deals closing in December, by which point a dozen other things have changed.
• The buyer is a committee. Gartner's 2025 survey of 632 B2B buyers found buying teams of five to sixteen people across as many as four functions, so the visitor is not the unit of decision.
• Mix matters more than rate. Twenty submissions from target accounts beat two hundred from companies you cannot serve.

The Research That Should Come First

Most teams skip straight to testing because the tool makes testing feel like the work. Research is cheap, and it is where testable ideas come from.

• Session recordings and heatmaps. Microsoft Clarity and Hotjar show where people stop, hesitate and rage-click. Mask form inputs by default: session replay and pixel tracking are an active and unsettled area of US privacy litigation, and the choice deserves counsel rather than a plugin default.
• Form analytics. Field-level abandonment and error rates identify the specific field that is failing, which no amount of layout theorising will find.
• Exit and on-page surveys. One question, asked at the point of leaving: what were you looking for that you did not find.
• Recorded sales calls. The objections raised in the first ten minutes of a discovery call are the ones your page failed to handle. Tools such as Gong or Chorus make this searchable; lost-deal notes in the CRM do the same job free.

Five moderated sessions with people who hold the actual buying role usually surface more than a quarter of testing. Finding a problem does not require statistical significance; only measuring the size of a fix does.

The Arithmetic of an A/B Test

Randomise visitors rather than sessions, assign persistently, and decide the sample size before building anything.

For a comparison of two proportions at the conventional 5% two-sided significance level with 80% power, the sample needed per variant is approximately 15.7 x p x (1 - p) / d squared, where p is the current conversion rate as a decimal and d is the absolute improvement you want to be able to detect. The 15.7 is 2 x (1.96 + 0.84) squared: the critical value for a 5% two-sided test plus the value corresponding to 80% power, squared and doubled for two arms.

Work it. A page converting at 2%, with a hope of detecting a 20% relative improvement, gives d = 0.004, so n is about 15.7 x 0.0196 / 0.000016 — roughly 19,200 visitors per variant, or 38,400 in total. At 400 visitors a month that is eight years. Relax the requirement to a 50% relative improvement, from 2% to 3%, and it falls to about 3,100 per variant: still more than a year at the same traffic.

The lesson is in the d squared. Sample size scales with the inverse square of the effect, so halving the effect you want to detect quadruples the traffic required. Small tests can only ever detect large effects. Two checks belong here: run in whole weeks, since B2B traffic collapses at weekends, and verify the split — if you asked for 50/50 and got 54/46 on several thousand visitors, a chi-square test will say that is not chance, and a sample ratio mismatch invalidates the result.

Peeking, Stopping Early, and Testing Four Things at Once

Three ways to get a wrong answer.

Peeking. The fixed-horizon test that produces a p-value assumes one look, at a sample size fixed in advance. Checking every morning and stopping when the figure crosses 0.05 is a different procedure with a different error rate: the literature on repeated significance testing, going back to Armitage and colleagues in the late 1960s, shows the false-positive rate rising well above the nominal 5% as the number of looks rises.

Stopping early on a win. Worse, because it biases the estimate as well as the decision: you stopped because the observed difference was extreme, so the reported effect is the high end of a noisy range. Lifts that fail to reproduce are usually this.

Multiple comparisons. Test four variants against a control and you have made four comparisons. If every variant were genuinely identical to the control, the chance of at least one crossing p below 0.05 is roughly 1 minus 0.95 to the fourth power — about 19%. Slicing results by device, channel and industry afterwards multiplies it. The corrections are ordinary statistics — Bonferroni, Holm, or controlling the false discovery rate — and the discipline is cheaper: name one primary metric and one stopping rule before the test starts, and treat everything else as a hypothesis for next time.

What to Do When the Traffic Is Not There

For most B2B sites the arithmetic says no, and the answer is to change method rather than to pretend.

• Test big, not small. Only large effects are detectable at low volume, so change the page concept instead of the button. You lose the ability to say which element worked and gain the ability to detect anything at all.
• Pool the traffic. Test at template level across every campaign page rather than on one page.
• Use sequential comparison honestly. Before-and-after measurement is weak but not worthless, provided you state the confounds — seasonality, campaign changes, a competitor's launch — and track an untouched control metric to catch site-wide shifts.
• Hold out accounts, not visitors. In a named-account market, withholding a change from a matched account list is often more practical than randomising visitors.
• Buy information qualitatively. Moderated sessions and expert review find broken things quickly, which at these volumes is most of the available gain.

Group sequential methods with alpha spending, and the always-valid approaches some vendors offer, legitimately allow repeated looks without inflating error; Bayesian reporting changes what the output means rather than requiring less data. None create information. A small sample remains a small sample.

Micro-Conversions, Guardrails, and the Lift That Costs Money

Micro-conversions — video views, scroll depth, pricing-page visits — are diagnostic, and they are the only signals with enough volume to move measurably. Their misuse is predictable: they get promoted to targets, then to slides, and eventually somebody optimises for scroll depth.

The failure that matters more is subtler. Remove the qualifying fields from a form, or replace a demo request with a document download, and completions rise while the share of completions that ever become opportunities falls further. Form fills go up, pipeline goes down, and the report says the test won. That is the default outcome of optimising a proxy.

Catching it takes three habits: pass the test variant into the CRM with the lead record, so downstream rates can be compared by arm; report pipeline value per thousand visitors rather than conversion rate alone; and set guardrail metrics — lead-to-opportunity rate, share of submissions from target accounts — that can veto a winner even when the primary metric improves. The verdict arrives a quarter or two after the test ends.

One plumbing note. GA4's behavioural modelling for consent mode requires at least 1,000 events a day with analytics storage denied and 1,000 daily consenting users, so most B2B properties never qualify: they take the full data loss from consent refusal with none of the modelled recovery.

Frequently Asked Questions