# How to analyse a conversion funnel in GA4 with an AI agent

> Tier: deep. Full contract: quantified decision rules and evidence requirements.

Where the funnel leaks, step by step, segment by segment, with the ecommerce diagnostics that say whether the leak is real or a measurement hole.

**Intent.** Name the step where users leave and the segment where it is worst, with measurement verified before conclusions.

## When to use

- Conversion rate moved and the question is where in the journey.
- After a checkout, form or landing change.
- Comparing funnel health across devices or markets.

## When not to use

- Sites without the funnel events instrumented: run tracking-audit first, this skill assumes the steps exist.
- Same-day reads: GA4 intraday data is incomplete.

## Sources

Required: Google Analytics 4

## Preconditions

- Funnel steps defined as events in the property.
- Enough volume per step for segment cuts to hold.

## Inputs

- `steps` (required): The funnel steps, as GA4 events.
- `window`: Analysis window. Default: last 28 days.

## Procedure

1. Verify the step events exist and fire at plausible volumes. — tools: `ga4_get_key_events`, `ga4_get_event_parameters`
2. Check ecommerce measurement health where the funnel is a purchase funnel. — tools: `ga4_get_ecommerce_diagnostics`
3. Run the funnel over the window. — tools: `ga4_run_advanced_funnel_report`, `ga4_run_funnel_recipe`
4. Cut the worst step by device, channel and market. — tools: `ga4_run_pivot_report`, `ga4_run_report`

## Decision rules

- **When** One step's drop-off exceeds the others by a wide margin and holds across segments → **conclude** Structural leak at that step: the fix is the page or flow, not the traffic.
- **When** The drop concentrates in one segment (a device, a browser, a market) → **conclude** Segment-specific defect: name it, because a global fix for a Safari problem wastes a sprint.
- **When** A step's volume is implausibly low versus its neighbours → **conclude** Measurement hole, not user behaviour: route to tracking-audit before any UX conclusion.

## Evidence requirements

- Every rate carries both step volumes and the window.
- Segment findings show the same cut on healthy segments as contrast.

## Human approval boundary

- Site changes are product work; the skill delivers the located leak with its evidence.

## Failure modes

- Funnels count sessions or users depending on construction; mixing the two fabricates drop-offs.
- Consent-gated analytics undercounts some markets' early steps; a 'leak' at step one can be a banner.
- GA4 thresholding suppresses small segments; absence of a row is not absence of users.

## Output

- **The funnel**: Volumes and rates per step, the leak named.
- **Worst segment**: The cut where it is worst, with contrast.
- **Measurement verdict**: Whether the numbers can carry the conclusion.

## Related

- Skill: [tracking-audit](https://www.getmcpads.com/skills/tracking-audit.md)
- Skill: [landing-page-audit](https://www.getmcpads.com/skills/landing-page-audit.md)
- Guide: [ga4-analysis](https://www.getmcpads.com/guides/ga4-analysis.md)