Skip to content
RAY BEHARRY

Agent 48 of 71 · Conversion Optimization

Funnel Analysis Agent

Locate the leak before anyone touches copy. Rebuild the funnel step by step from exported analytics and CRM data, show every rate with its numerator, denominator and window, segment by source, device and ICP match, and rank the drop-offs by recoverable pipeline rather than by how alarming the percentage looks.

Agent contract

Use to find where and for whom a conversion funnel leaks, from real GA4 and CRM exports with explicit date ranges and denominators. Produces a ranked drop-off diagnosis, never invented rates.

When it runs

  • Monthly funnel review is due.
  • Qualified demo requests fall while traffic holds steady or rises.
  • A new campaign, page or form goes live and needs a pre-change baseline.
  • CRO Performance Analyst needs a baseline conversion rate to size an experiment.

What it needs

InputRequiredFrom
GA4 funnel or landing-page export with date rangeYesGA4 (integration)
HubSpot lifecycle-stage export (lead → MQL → SQL → opportunity)OptionalHubSpot (integration)
Salesforce opportunity export with source fieldsOptionalSalesforce (integration)
ICP definitions for match taggingYesICP Library
Prior funnel observationsOptionalPerformance Observations

How it works

  1. Import each export with the matching `python3 scripts/sns.py import <adapter> <file>` command and record file name, system and date range.
  2. Define the funnel steps explicitly (e.g. landing view → CTA click → form start → form submit → MQL → SQL) and the event or field that marks each.
  3. Compute step conversion as count(step n+1) / count(step n) for one aligned window; show both counts beside every rate.
  4. Segment by traffic source, device, landing page and ICP match; suppress any cut with fewer than 100 entrants or label it low confidence.
  5. Compare with the prior equivalent window and note confounders such as tracking changes, seasonality or a campaign burst.
  6. Rank drop-offs by recoverable volume: entrants at the step × gap to the best comparable segment, carried through to SQLs.
  7. Write the funnel-diagnosis and log each rate as a measured_result observation with source system and window.

What it delivers

Step definition table, step counts and rates with denominators, segment cuts, window comparison, top three ranked leaks with recoverable volume, confounder notes.

Output contract: funnel-diagnosis

Quality bar

  • Every rate shows numerator, denominator, source system and identical date window.
  • Cuts under 100 entrants are suppressed or labelled low confidence.
  • GA4 and CRM counts are never mixed in one rate without a stated join key.
  • Tracking gaps are reported as gaps, not filled with estimates.

Hands off

When a person steps in

  • If tracking broke during the window, stop and publish the gap rather than a corrected guess.
  • Escalate if CRM lifecycle definitions changed mid-window; the comparison is invalid until a human confirms the mapping.

Example

Request

“Why did demo requests drop last month even though traffic is up?”

Result

GA4 1 Sep–30 Sep vs 1 Aug–31 Aug. Sessions +18% (41,200 vs 34,900) but form start → submit fell from 38% (612/1,611) to 24% (418/1,742). Leak concentrated on mobile (n=1,020 starts) after a new required phone field went live on 9 Sep. Paid social share rose, lowering ICP match from 61% to 47%: confounder declared. Two leaks handed on.

Illustrative example from the package documentation, not a client result.