How to analyse a conversion funnel in GA4 with an AI agent
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.
IntentName the step where users leave and the segment where it is worst, with measurement verified before conclusions.
- 4
- Steps
- 7
- Tools
- 3
- Rules
- 3
- Failure modes
When to use it
Use it when
- 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.
Do not use it when
- Sites without the funnel events instrumented: run tracking-audit first, this skill assumes the steps exist.
- Same-day reads: GA4 intraday data is incomplete.
What it needs first
Required
Preconditions
- Funnel steps defined as events in the property.
- Enough volume per step for segment cuts to hold.
Inputs
stepsrequired- The funnel steps, as GA4 events.
window- Analysis window.Default: last 28 days.
Procedure
In order. Every tool named is one the gateway ships, and links to its reference.
Verify the step events exist and fire at plausible volumes.
Check ecommerce measurement health where the funnel is a purchase funnel.
Run the funnel over the window.
Cut the worst step by device, channel and market.
Decision rules
The observable condition, what it lets you conclude, and what takes the conclusion back.
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 every conclusion must carry
- Every rate carries both step volumes and the window.
- Segment findings show the same cut on healthy segments as contrast.
Where the agent stops
- Site changes are product work; the skill delivers the located leak with its evidence.
How it goes wrong quietly
The cases where the analysis is wrong and still looks right. Read them before trusting a number.
- 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.
What the answer contains
- 01The funnel
- Volumes and rates per step, the leak named.
- 02Worst segment
- The cut where it is worst, with contrast.
- 03Measurement verdict
- Whether the numbers can carry the conclusion.