Funnel Optimization

• Turn a drop-off diagnosis into a prioritized experiment
• Design an experiment that reads the funnel honestly
• Compound funnel gains and know when to stop optimizing

• Drop-off analysis locates the opportunity; experimentation converts it into verified revenue.
• Single-variable steps beat 'let's improve everything' because attribution stays clean.
• Compounding makes funnel work multiplicative, unlike one-off feature bets.
• A closed loop builds reusable playbooks (form length, trust elements, progress cues) the whole company reuses.

Plugging the same size leak at each of three connecting ditches doesn't save 5% three times — it compounds, because every liter saved at ditch one stays saved through ditches two and three.

• Rank leaks by absolute loss; pick the top one where you have a concrete friction hypothesis.
• Pre-register: hypothesis, change, success metric (step + overall conversion) and expected lift (MDE).
• A/B or sequential-roll the change with segmentation intact; run the full planned window.
• Read result against pre-registered metrics and guardrails (no error-rate or refund regression).
• On win: ship, re-baseline the funnel, and move to the next leak; on null, log the learning and move on cheaply.

• A checkout funnel ran three sequential tests (form reduction, trust badges, payment-logos) and compounded overall conversion +61% over six months.
• A signup funnel tested a single-question gate vs no gate; no-gate won on step conversion but lowered lead quality, so guardrais decided the no-go.
• A SaaS rings the optimization loop on its trial→paid funnel, setting an 'optimization budget' that stops when further tests lack traffic for the MDE.

• Optimize one step per experiment and one variable per test.
• Pre-register the step and overall conversion; both must move for a confident win.
• Guardrail the funnel's quality side (refund rate, support tickets) so conversion gains aren't bad-faith users.
• Retire the loop when traffic can't power the next MDE — better to expand scope than run noise.

• Testing on the funnel average without segment-aware stratification, then shipping a change that helps one platform and hurts another.
• Burning multiple experiments on one step in parallel with overlapping samples.
• Reading 'step conversion' while ignoring overall, so a change that pushes users later in the funnel looks like a win.
• Never logging nulls, so the team re-tests the same dead end in triplicate.

• If the step wins but overall doesn't, the change likely shifted the mix (more entrants of lower quality) — check step-1 composition.
• Re-baseline dashboards after shipping; a stale baseline makes the next experiment's readout lie.
• Track the optimization loop's own funnel: how many experiments shipped per quarter and avg verified lift.
• If a winner doesn't reproduce post-launch, verify exposure — a flag leak can inflate the experiment count.