AI Prompt: Find Why Your Store Gets Traffic but Few Sales

ROLE
You are an evidence-led ecommerce conversion analyst. Produce an investigation plan, not a promise of sales lift.
NAMED INPUTS
[SCOPE]: market, device, period, timezone, session-or-user unit, observation window.
[FUNNEL_CONTRACT]: event definitions, closed ordered sequence, identity, exclusions, late-data and consent policies.
[COUNTS]: entry, cart, checkout, purchase unique units.
[ECONOMICS]: currency, contribution per converting unit, cost boundary, expected implementation cost.
[RECOVERY_ASSUMPTION]: fraction of transition drop-offs hypothetically recovered and why chosen.
[OBSERVATIONS]: anonymized customer task obstacles, dates, devices, source and confidence.
[CHANGES]: campaigns, product/stock/pricing/tracking changes in the period.
CONSTRAINTS
Do not request or reproduce names, email addresses, payment data, session IDs or private replay links.
Do not browse unless explicitly enabled; never invent platform documentation, benchmarks or customer motivations.
Use no paid tools. Verify arithmetic with a local calculator/code tool when available; otherwise mark calculations unverified.
Check unit/window/sequence compatibility before subtracting. Monotonic counts are necessary but not sufficient.
Treat missing denominators as unknown. Do not invent a recovery fraction.
Analyze each transition independently. Do not sum scenarios or label them forecasts, causal lift or guaranteed revenue.
OUTPUT SCHEMA (JSON)
{
 "decision": "MEASUREMENT_FIX | MORE_RESEARCH | CONTROLLED_TEST",
 "scope": "...",
 "data_checks": [{"check":"...", "status":"pass|fail|unknown", "evidence":"..."}],
 "transitions": [{"name":"...", "dropped":0, "step_rate":0, "extra_advances":0, "downstream_rate":0, "extra_converting_units":0, "extra_contribution":0}],
 "hypotheses": [{"mechanism":"...", "support":"...", "alternative":"...", "falsification":"..."}],
 "next_test":{"change":"...", "audience":"...", "primary_outcome":"...", "guardrails":[], "owner":"...", "evidence_needed":[]},
 "limitations":[], "questions":[]
}
Use null rather than zero for unknown numeric values.
METHOD
1 Validate the contract; fail or unknown conditions block a causal conclusion.
2 Compute differences and conditional rates. Apply the recovery assumption only if supplied.
3 Link each recommendation to an observation, not just a percentage.
4 Choose one next action. Separate bug repair, measurement repair and randomized testing.
5 Provide uncertainties and an explicit refinement request.
EVALUATION RUBRIC (0 missing/incorrect, 1 partial, 2 explicit and correct)
Population compatibility; arithmetic; evidence provenance; alternative explanations; actionable test; privacy.
Any zero in compatibility/arithmetic/privacy requires revision regardless of total score.
REFINEMENT PASS
Audit your own draft against the rubric. Correct unsupported statements, flag unverified arithmetic and ask for the smallest missing evidence that changes the decision.

Use AI to challenge a diagnosis, not invent one

This prompt is for growth leads and founders who have aggregate funnel counts and want a structured investigation. It can run as text in a general assistant; a local calculation tool improves arithmetic checking but is not required. No browsing or paid API is required. It has not been validated as equally reliable across models. Independently check the output against the companion calculator before using it.

Replace the named inputs with anonymized aggregate evidence. Do not paste customer identities, payment fields, authentication tokens or raw replay URLs. If the counts or definitions are unavailable, keep them explicitly missing. The correct output may be a measurement repair rather than a conversion recommendation.

Synthetic filled input

SCOPE: one store, mobile sessions, one completed 30-day period, one timezone. CONTRACT: entry → cart → checkout → purchase, indirectly followed in the same session, one count per session, production-only, consistent consent policy. COUNTS: 10,000 / 1,000 / 600 / 300. ECONOMICS: EUR30 contribution per converting session after variable costs; test cost unknown. RECOVERY: assume 10% for sensitivity only. OBSERVATIONS: a synthetic moderated task found the delivery charge appears only at checkout; no frequency estimate exists. CHANGES: no verified change log supplied.

Reference answer to check against

The arithmetic is 9,000 / 400 / 300 drop-offs; 10% / 60% / 50% step completion; 270 / 20 / 30 extra purchasing sessions; EUR8,100 / EUR600 / EUR900 scenario contribution. These are separate alternatives. A defensible decision is MORE_RESEARCH: verify how often delivery-price uncertainty occurs, confirm the change log and scope, and cost a clear delivery-price disclosure test. Do not announce that late delivery costs caused all checkout abandonment.

A candidate mechanism is that earlier cost clarity reduces surprises, but it might also move abandonment earlier without increasing purchases. The primary outcome should therefore be purchase conversion for the eligible audience, with contribution and refunds as guardrails—not merely checkout starts. The answer must preserve that alternative explanation.

Evaluate and refine

Score the six criteria in the template from 0 to 2. A fluent answer with a mixed denominator, fabricated benchmark or exposed customer data fails even if the total looks high. Check all numeric fields against the calculator, verify each source claim and ensure null is used for unknown values. Ask a second pass to identify the weakest assumption and the smallest observation that would change the recommended action.

The example is a reference for arithmetic and reasoning, not proof of prompt reliability. AI can overstate evidence, silently change the unit or turn a hypothetical contribution into a revenue claim. Keep approval with a named person, and do not connect this prompt directly to live checkout changes or ad spending.

Sources and limitations

Primary documentation checked October 18, 2026: Shopify behavior reports explains session-based open and closed funnels; Google Analytics funnel exploration explains user-based sequence counting. Publication dates are unknown for these living documents. These systems do not automatically share a denominator. The worksheet and scenario method are original editorial synthesis, not vendor-approved benchmarks. Examples are synthetic, not client results.