Data-driven conversion optimization, server-side

Three forces make classic client-side testing untenable. First, Safari ITP 2.3 caps client-side cookies at 7 days, a bucketing state that decays after a week breaks every test. Second, ML-based adblockers filter 25 to 40 percent of exposure events, the power analysis runs against invisible sampling. Third, client-side variants force a FOUC before the snippet swaps the variant, measurably damaging CLS and LCP in Core Web Vitals. Server-side assignment at the edge avoids all three: bucketing happens on the server, exposure is logged in the backend, the browser receives the final variant without flicker.

Both, depending on the question. Frequentist (Welch t-test, Mann-Whitney U) yields classical p-values and MDE power analysis, suitable for regulated tests and hard go/no-go decisions before launch. Bayesian (native in GrowthBook or Statsig) gives a direct probability statement ("Variant B is better with 94 percent probability") and allows sequential testing without alpha inflation, useful when the business case is asymmetric and peeking has to be allowed. We pick the approach during the hypothesis kickoff, not afterwards.

Assignment logic runs in Next.js middleware or a Cloudflare Worker. Concretely: the request hits the edge layer, we hash the user ID (anonymous or consent-bound PII) against the test definition, write the bucket cookie, and route to the assigned variant. The exposure event goes to PostHog or GrowthBook in parallel, and to BigQuery for evaluation. Variant code stays with your dev team, we ship the edge handler, flag service and exposure SDK wiring. Works on Vercel, Cloudflare Pages, Netlify Edge and SST.

We prevent client-side flickering through architecture, not trickery. Client-side test frameworks like Google Optimize (sunset 2023), VWO Visual Editor or Optimizely Web inject CSS and DOM mutations after the first paint, producing Flash of Unstyled Content and a measurable CLS jump. With server-side assignment at the edge the browser receives the HTML of the assigned variant directly. No paint cycle, no mutation observer, no layout shift. Before rollout we document the CLS delta between control and variant and stop the test if the variant shows performance regression.

Rule of thumb: at least 1,000 conversions per test variant over 4 weeks. Below that the test runs longer than the hypothesis stays valid, or returns no significance. With lower traffic we use qualitative methods, the obvious wins from the conversion audit and switchback designs at the marketing-channel level. We don't recommend tests that aren't structurally measurable, that's burning budget.

A documented record of every test, won or lost. Each entry contains: hypothesis, test design, MDE, runtime, result with confidence interval, derived learnings. Prevents the same tests running twice, accelerates the next hypothesis by weeks. We host the library in Notion or Linear, ownership stays with you.

At least 2 weeks to cover weekday and weekend effects. Typically 3 to 6 weeks to significance, depending on baseline conversion rate, MDE and traffic volume. We don't stop tests early because an interim result looks good. Peeking without alpha-spending destroys the statistical claim, one of the most common errors in CRO programmes.

Audit Sprint from €2,900 net, 2 weeks delivery. Includes conversion data audit, prioritised funnel leaks, test roadmap. Build Sprint for the edge architecture (middleware handler, flag service, exposure pipeline) at a fixed price set after the audit. Managed Evolution as a monthly retainer, covers hypothesis backlog, test operations and quarterly roll-up.