Funnel Segmentation

• Identify the highest-value segment splits for a funnel
• Detect segment-specific leaks that the aggregate hides
• Use segments to prioritize fixes and validate them

• Aggregate conversion hides which users suffer; segments name them.
• Segment leaks have distinct causes (device bugs, channel expectations) and distinct fixes.
• Cohort segmentation ties funnel health to acquisition quality over time.
• After a fix, the target segment's conversion is the validation metric; the company metric is too noisy to read.

'The avenue screams' says nothing about whether it's buses or bikes struggling at the light. Splitting by vehicle type isolates the real congestion source before you redesign the intersection.

• Choose segments that match plausible causes: platform/device/browser for technical risk, channel for expectations, plan/cohort for product intent.
• Compute the funnel per segment with identical step definitions.
• Compare shapes and step rates; look for segments that differ more than noise (use CIs on small segments).
• Prioritize by absolute affected users, not relative anomaly alone.
• Act: one fix per segment cause, then verify the segment's funnel moves while others stay flat.

• A checkout leak was invisible in aggregate (42%) but 18% on Samsung browsers — an in-app browser bug; reverting a WebView change restored parity.
• A growth team found paid-channel cohorts funneled worse at step 1 (expectation mismatch) and reworked the landing page for those campaigns.
• A B2B SaaS segments by plan; the free-plan funnel leaks at the team-invite step, so the fix targets free-plan onboarding, not the enterprise flow.

• Pre-register which segments you'll inspect; post-hoc segment watching is p-hacking again.
• Use identical step definitions across segments — segment tools that apply different dedupe break comparisons.
• Report segment sizes next to rates so thin segments don't drive priority.
• Verify fixes against the specific segment, then confirm the aggregate moved too.

• Reading a 2% segment with a 10% leak as 'the' problem — check absolute users first.
• Segment-hopping: testing 20 slices until one looks interesting, without correction or pre-registration.
• Comparing segment rates across tools with different 'step reached' semantics.
• Fixing a segment leak blind: instrument the segment's events before changing product code.

• Segment by browser-engine class (WebKit, Blink) rather than vendor string when leaks appear on clones.
• Cohort × platform cross-splits catch version-regression leaks (a break affecting a specific app version).
• If one segment's step conversion shows 100%+ it's a dedupe artifact, fix before interpreting.
• Test segment leaks on a staging monolith with the same user-agent strings to reproduce pre-release.