Plan a Conversion Experiment from Funnel Evidence

Define comparable cohorts, trace a real observation to one change, and prepare metrics, guardrails and a rollback before implementation. Missing counts remain unknown.

Fastest way

Ask your AI to do it

Paste this into the AI agent you already use:

Use Conversion Experiment Lab with my actual page observation and permitted cohort data. Ask for one counting unit, stage definitions and the evaluation window. Leave missing counts unknown, propose one bounded change, and define the metric, guardrails, stopping rule and rollback. Build an inspectable plan into a fresh folder; do not publish, track or charge anything.

Choose one journey and one counting unit

Start with a specific journey such as product page → signup → activation. Use the same eligible cohort, date range and counting unit for every stage. Unique people, sessions, button clicks and orders cannot be mixed into one nested funnel. If a count is missing, leave it unknown instead of substituting zero.

Separate stage rates from overall conversion

In an illustrative exercise only, 1,000 eligible sessions might include 200 signups and 50 activations. Signup rate is 200/1,000 = 20%; activation among signups is 50/200 = 25%; overall activation is 50/1,000 = 5%. These are fictional teaching numbers, not Skillrove measurements. When the previous stage is zero, the conditional rate is undefined, not a measured zero.

Turn an observed obstacle into one testable change

Record the page URL, current copy, action destination and observed behavior. Keep the observation separate from the explanation. For example, an unclear pricing period is observable; the claim that it caused abandonment is a hypothesis. Propose one bounded change and explain what outcome would support or contradict it.

Use Launch Campaign Kit when the change concerns a factual product launch narrative, or Client Proposal Studio when it concerns offer scope and deliverables. Their generated artifacts do not assign experiment groups, collect analytics or prove an uplift.

Write the measurement contract before launch

Name the eligible cohort, assignment rule, exposure event, primary metric and evaluation window. Confirm that event deduplication and actual fulfillment match the business question.
Choose guardrails such as errors, support burden or cancellation, and a rollback condition. Define what you will do if tracking is incomplete or an important guardrail worsens.
Keep current account consent, publication permissions and financial authorization. Installing a planning skill does not authorize new tracking, live purchases or contacting customers.

Build an inspectable plan from the included example

Read Conversion Experiment Lab’s SKILL.md, references/inputs.md and references/experiment-design.md. Python 3.10+ runs the dependency-free example below from the installed folder. Keep its fictional inputs labeled, and use a new output directory.

python3 scripts/build-lab.py --input examples/experiment.json --output ./practice-conversion-plan

Review the plan, then evaluate actual cohorts

Open index.html and review journey, experiment, copy and source views. Check plan.json, experiments.csv, tracking-plan.json, copy-variants.md and handoff.md against the original observation. Implementation and actual data collection are separate work. Use the agreed evaluation method and stopping rule; a prettier page, a before/after traffic change or a small fluctuation is not a demonstrated experiment win.