Conversion Rate Optimization Checklist

Export step-by-step conversion rates for the funnel under review (landing → form view → submit → MQL, or PDP → cart → checkout → purchase). Pair GA4 with the source-of-truth CRM or order system — GA4 alone often under-reports due to consent loss and ad-blocker stripping.

Rank steps by drop-off rate and dollar impact. The biggest percentage drop is often not the most valuable to fix — focus where a 10% lift moves the most pipeline or revenue. Segment by device, source, and new-vs-returning before picking a target.

Confirm the conversion event fires on actual submit, not on form-blur or button-click. Cross-check GA4 conversion count against CRM lead count for the same window — discrepancies above 10% usually indicate a duplicate fire or a missing event.

Pull source/medium/campaign for the last 90 days and look for drift — "facebook" vs "Facebook" vs "fb", "cpc" vs "paid". Mixed conventions make experiment segmentation unreliable. Lock the convention in the shared link-builder before launch.

Open the page in incognito with EU geolocation and inspect Network — Meta Pixel, GA4, and any optimization-tool script must not fire before consent. OneTrust, Cookiebot, and Didomi all support GTM consent-mode v2; verify the analytics_storage and ad_storage flags flip correctly on accept.

Pull LCP, INP, and CLS from GSC for the candidate URL. LCP > 2.5s and CLS > 0.1 are the common conversion killers on mobile. Note device breakdown — mobile typically fails Vitals first and converts worst.

Run mobile-throttled Lighthouse three times and take the median. Look for render-blocking scripts, unoptimized images, and unused JavaScript above 200KB. File performance fixes that block the test as separate dev tickets — do not bundle them into the variant.

In Hotjar, FullStory, or Microsoft Clarity, watch 10–15 sessions of users who dropped before the conversion event. Look for rage clicks, dead clicks on non-clickable elements, and form fields where the cursor leaves and never returns. Patterns from 10 sessions usually beat aggregate funnel charts for hypothesis quality.

One variable per test. If the hypothesis is about CTA copy, only the CTA changes; do not also swap the hero image "while we're at it" — that ruins attribution of the lift. Capture the change in a one-line diff so analysis later is unambiguous.

Build in VWO, Optimizely, AB Tasty, or your CMS-native tool (HubSpot, Webflow, Contentful). Set traffic split, audience targeting, and exclusion rules. If the tool flickers on render, switch to server-side or anti-flicker snippet — flicker biases the result against the variant.

BrowserStack or LambdaTest the variant on iOS Safari, Android Chrome, desktop Chrome, Edge, and Firefox. Confirm the conversion event still fires on the variant — broken tracking on the variant is the most common reason a test reads as a loser when it isn't.

Any claim change, pricing reference, or testimonial needs review. Substantiation file required for performance claims under the Lanham Act and FTC Endorsement Guides. Capture approval here so the audit trail beats a Slack DM.

Use the 50/50 split planned in step 4 unless there is a strong revenue-risk reason to skew. Note the launch timestamp — many tools record start time imprecisely. Announce in #growth or #experiments so paid-media and email teams do not push major sends that bias one arm.

Sample Ratio Mismatch (chi-square p < 0.01 between expected and observed split) means the test is broken — usually a redirect or bot-traffic issue. Also watch guardrails: bounce rate, page errors, and revenue per session. Kill early if a guardrail breaks badly; do not peek at the primary KPI for stat-sig.

Read the primary KPI first; segments only after. A "win in mobile only" found post-hoc without correction is exploration, not a result. Confirm the lift exceeds the MDE set in step 4 — a stat-sig 0.5% lift below your MDE means the test is underpowered, not a winner.

Log the hypothesis, variant diff, sample size, primary KPI delta, segment results, and a one-paragraph "what we learned" — losers and inconclusives matter as much as winners. Tag by funnel stage and surface so future tests can search prior art before re-running the same test.

Bake the change into the production CMS rather than leaving it served by the optimization tool — long-running tool overrides accumulate flicker and tech debt. Remove the variant from the optimization tool once the production change is live.