Skip to content

Gets data

Conversion research

Runs a technical, a heuristic and an analytics pass on a funnel and keeps only the issues 2 or more passes found, ranked by estimated loss.

Other tools

Add this skill to Claude

Claude takes skills as a ZIP file. Your download has started; if it didn't, use the button below.

  1. 1In Claude, open Customize, then Skills.
  2. 2Choose +, then Upload a skill, and pick conversion-research.zip.
  3. 3Ask for it by name, or let Claude use it when a task matches.

Skills need code execution turned on in Claude's settings. On a Team or Enterprise plan, an admin can share it with everyone.

Download the folder

Try it now, without installing

See the skill work on a made-up input before you add it. Nothing is installed, and you can swap in your own input.

Open in the Claude app

Opens a new chat in the Claude desktop app with the prompt filled in. Nothing is sent until you press enter.

Includes the whole SKILL.md and the sample input. Paste it into Claude, ChatGPT or any chat.

Influenced by
Conversion research (ResearchXL)
Needs
The funnel's steps, and what counts each one · Step counts for this period and the last, the same length · CRM counts for the same steps
Works with
Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md
Version
1.0.0 · Last checked · License MIT
Maintained by
Sabih AhmedCreative
SkillCreative and conversion

Conversion research

Example output

Find where the funnel leaks
IssuePassesStatus
Cookie banner covers Submit on iPhone3Confirmed
2 fields added to the form2Confirmed
Slow page on phones1Watch list

Runs a technical, a heuristic and an analytics pass on a funnel and keeps only the issues 2 or more passes found, ranked by estimated loss.

The skill

The whole skill, in plain text.

Nothing is hidden. Read it, change it to fit your team, or take the whole folder.

Scripts
Optional Python scripts for the counting. It works without them and says which way it ran.
Network
Opens web pages or your tools if your agent can. Works from pasted data if not.

3 test cases

conversion-research/SKILL.md

In the folder

  • conversion-research/SKILL.md
  • conversion-research/references/passes.md
  • conversion-research/scripts/funnel_math.py
---name: conversion-researchdescription: >-  Researches why a B2B SaaS funnel loses buyers by running 3 passes on it (technical, heuristic and analytics, plus  an optional qualitative pass), merging what they find by location, and keeping only the issues 2 or more passes  found, ranked by an estimated loss with its inputs shown. Checks that analytics agrees with the CRM before  trusting any rate, and sizes each proposed test so nobody runs one the traffic can't finish. Use it when a funnel  step falls or a page underperforms and the cause isn't obvious. Not for reviewing one page on its own (use  landing-page-critic) or fixing tracking (use tracking-audit).license: MITcompatibility: >-  Works in Claude, ChatGPT, Codex, Cursor and any agent that reads SKILL.md. Uses connected analytics and a CRM if  the agent has them, otherwise exports or pasted tables. Python 3 is optional, for leak, segment and sample size  math.metadata:  author: SHAKEUP  version: "1.0.0"  stage: get  plugin: creative-conversion  homepage: https://www.shakeuphq.com/skills/conversion-research--- # Conversion research Finds where and why a funnel loses buyers, and keeps only what more than one kind of evidence agrees on. ## Purpose - One job: one funnel, 3 passes, one ranked list of confirmed issues with the action for each.- Stage: get. It gathers and structures evidence and ends in a research verdict, not a redesign.- It never ranks on data it hasn't checked against the CRM, never reports a heuristic opinion as a confirmed issue, never invents a benchmark or a load time, and labels every loss as an estimate with its inputs. ## Use when / not | Situation | Use || --- | --- || A funnel step fell and nobody knows why | this skill || A page's conversion rate is low and the page review didn't explain it | this skill || Deciding whether a proposed test can finish on your traffic | this skill (step 10), or the script's `sample` command alone || Reviewing one page before launch | `landing-page-critic` || Analytics and the CRM disagree | `tracking-audit` first || Pipeline by channel for a report | `pipeline-report` | ## Inputs | Input | Required | Ways to provide, in order | If missing || --- | --- | --- | --- || The funnel: its steps and what counts each (event or CRM stage) | yes | the reader's note; inferred from the analytics and CRM and confirmed | ask once; with none, `BLOCKED` || Step counts, this period and last (same length) | yes, for the analytics pass | connected analytics and CRM; exports; pasted table | analytics pass `not run` || CRM counts for the same steps and periods | yes, for the health check | connected CRM; an export; pasted counts | health check `not run`; coverage `partial`; say every rate is unverified || Conversion by device and browser | no, for the technical pass | connected analytics; an export; pasted table | browser check `not run` || Test submissions, speed reports, error logs | no, for the technical pass | the agent tests if it can browse; the reader's results; PageSpeed Insights output | those checks `not run` || The funnel's pages | no, for the heuristic pass | live URLs; screenshots; pasted copy | heuristic pass `not run` || Recordings, polls, user tests, sales notes | no | the reader's tools or notes; a `message-mining` handoff | qualitative pass `not run` || Weekly visitors to each page under test | no | analytics | test sizing `not run` | Triangulation needs at least 2 passes. With fewer, the verdict is `BLOCKED`. ## Profile read From `company-profile.md`: `## Company` (business model and sales cycle, which set the funnel steps and how long a period must be), `## Buyers`, `## Targets` (lead to opportunity and opportunity to closed-won rates, used to turn lost leads into lost opportunities), `## Accounts and data` (which tools hold the counts), `## Channels`, and every `## Learnings` entry for `conversion-research` with status `noted`, `fact` or `accepted`. No profile: ask in one message for the funnel steps, the CRM and analytics tools, and the sales cycle length. Say that `company-profile` saves this next time. Without target rates, losses stop at the step where they're measured. ## Steps 1. **Apply learnings.** A learning that changes a threshold (a known tracking gap, a step to exclude) overrides the defaults. Name it when it fires. 2. **Map the funnel.** List each step, what counts it (the analytics event or the CRM stage), and the 2 periods. Each period must be at least as long as the sales cycle for steps that sales owns (a shorter period counts opportunities that haven't had time to happen), and the same length as the other. 3. **Run the health check first.** Use the table in `references/passes.md`. Compare analytics with the CRM for each step both hold, in each period: gap = (analytics minus CRM) divided by CRM, ignoring the sign. A gap over 20% on any step, or more than 1 conversion event per submission, is `TRACKING FIRST`: stop here, and name the causes to check: an event that fires twice, spam or test submissions, a broken form-to-CRM sync, a consent banner change. A gap that appears in one period only points at what changed between them. A step whose definition changed inside a period is excluded from the leak and named. 4. **Analytics pass.** Find the leak:    ```   python3 scripts/funnel_math.py leak --csv funnel.csv   ```    CSV columns: stage, this, last. It gives each step's rate now and before, the relative change, and a p-value; the leak is the biggest fall with p < 0.05, else the biggest fall marked `possible`. Without Python, follow `references/passes.md`. Then cut the leaking step by device, channel, landing page and new against returning visitors, and record each segment that carries the fall as a finding with its location. 5. **Technical pass.** Run the checks in `references/passes.md`: conversion by browser within each device class (`python3 scripts/funnel_math.py segments --csv segments.csv`), the form on the top browser and device pairs, the conversion event, speed by the tool's own labels, console errors, embedded tools. Record each finding with its location. 6. **Heuristic pass.** Run `landing-page-critic` on each page in the funnel and take its `failed` list as this pass's findings. Without it, use the 6 questions in `references/passes.md`. Each finding gets a location. 7. **Qualitative pass (if material exists).** Tag each observation from recordings, polls, user tests or sales notes to a location. 1 observation is an anecdote; record it only when several point the same way. 8. **Merge by location.** Put every finding in one table: location, symptom, and which passes found it, with the evidence from each. Merge findings at the same page and element, or the same funnel step and segment.    | Found by | Status |   | --- | --- |   | 2 or more passes | `confirmed` |   | A reproduced failure (a test submission that fails, an event that doesn't fire) | `confirmed` on its own: a reproduction is proof |   | 1 pass | `watch list`, with the pass that would confirm or clear it | 9. **Estimate the loss** for each confirmed issue, and show the inputs:   - At a leaking step: the end-of-funnel count lost had that step held its old rate (from the `leak` output).   - In a segment: its conversions short of its best peer (`sessions_short` from the `segments` output), carried down the funnel with the reader's own rates from `## Targets` or the step counts.   - With no way to size it: `not available`, and rank it after the sized issues.   Rank by estimated loss; ties by the number of passes, then by the cheaper fix. 10. **Pick the action** for each confirmed issue:     | Issue | Action |    | --- | --- |    | A reproduced bug or a broken event | fix now; no test |    | A change to the offer, message or form | size the test: `python3 scripts/funnel_math.py sample --rate <now> --lift <smallest lift worth it> --weekly <visitors a week, both versions>` |    | The test fits in 8 weeks | test it (past 8 weeks a B2B test measures a market that keeps moving) |    | It doesn't fit | test a bigger change, at least the smallest lift the script says the traffic can detect in 8 weeks; or ship the fix and compare periods with the leak method, labelled "not a controlled test" |    | The cause isn't clear | research more: name the pass that would explain it (a user test, an on-page poll) |     Without Python, the sample size per version is about 16 x p x (1 - p) / (p x lift) squared, for 95% confidence and 80% power, where p is the current rate. Say it's the rough form. ## Check and verdict | Verdict | Condition (first match wins) || --- | --- || `BLOCKED` | No funnel steps, or fewer than 2 passes could run || `TRACKING FIRST` | The health check failed: analytics and the CRM differ by more than 20% on a step, or events double count. Ranks built on broken counts point at the wrong step || `ISSUES FOUND` | 1 or more confirmed issues || `NO CONFIRMED ISSUES` | The passes ran and no issue was found by 2 of them | Before the verdict, check: every rate has its counts beside it; every loss is labelled an estimate with inputs; no heuristic finding is listed as confirmed on its own; every pass that didn't run is named. Named stops: `BLOCKED` asks for the missing funnel or data. `TRACKING FIRST` stops and hands to `tracking-audit`, listing the steps and the gap. No other stops. ## Learn Append to `## Learnings` in the profile: - A fix or a test result comes back (the step's rate before and after, or a test read with `funnel_math.py read`): a `fact` with the rates, the dates and the source.- The reader confirms a known, stable tracking gap ("analytics always counts 15% more submits than the CRM because of spam"): a `fact`, so the health check expects that gap next time.- A watch-list issue later proves real: a `proposal` to treat that pass as stronger for that kind of issue. Example: `- 2026-10-24 | conversion-research | fact | Removing the budget field lifted demo form completion from 3.1% to 4.0% (not a controlled test) | Why: GA4 and HubSpot, Aug against Sep 2026 | Status: noted` ## Output ```Verdict: <ISSUES FOUND | NO CONFIRMED ISSUES | TRACKING FIRST | BLOCKED> · Coverage: <complete | partial (what's missing)>Funnel: <steps> · Periods: <this> against <last> · Passes run: <list> · Not run: <list and why> · Math: <script | by hand>Health check: <pass | fail: step, analytics count, CRM count, gap> · Learnings applied: <list or none> Funnel| Step | This | Last | Rate now | Rate before | Change | p | Real | Confirmed issues (ranked)| Rank | Issue | Location | Found by | Evidence per pass | Estimated loss (inputs) | Action | Watch list| Issue | Location | Found by | Pass that would confirm it | Tests| Change | Rate now | Lift sought | Weekly visitors | Per version | Weeks | Call |``` Then the handoff block. ## Hand off ```handoffskill: conversion-researchversion: 1.0.0verdict: ISSUES FOUNDcoverage: partial (no session recordings)for: <funnel name>outputs:  leak: <step> (<change>, <real | possible>)  confirmed: [<issue at location>, <issue at location>]  fix_now: [<bug>]  tests: [<change>: <weeks> weeks, <call>]  watch_list: <n>next: [landing-page-critic]open: [<pass to add>]``` `landing-page-critic` takes the pages with confirmed message or friction issues for the rewrite. For `TRACKING FIRST`, `next` is `[tracking-audit]` and `outputs` lists the steps with the gap. ## Failure modes | Problem | Do this || --- | --- || Counts are too small for any fall to pass p < 0.05 | Report the biggest fall as `possible`; widen the periods to the length of 2 sales cycles if the data allows; lean on the technical and heuristic passes || Only one period of data | No leak; run the segment cuts within the period and say the analytics pass is partial || The reader wants a winner from a finished test | `funnel_math.py read` with each version's visitors and conversions; call a winner only at p < 0.05, and check downstream (opportunities, not only form fills) || A segment flag on a browser with odd traffic (bots, an office network) | Check its engagement first; bot traffic is a tracking finding, not a conversion issue || The reader wants to test with 2 weeks of traffic | Show the sample size and weeks; never call a test early because it looks ahead || The issue sits after the CRM handoff (no-shows, slow follow-up) | Record it with the CRM evidence; it is a sales process finding, still ranked | ## Influenced by - CXL's conversion research framework (ResearchXL), which gathers technical, heuristic, analytics and qualitative evidence separately and acts on the issues they agree on. We cut it to 3 required passes with a CRM health check in front, a merge rule by location, and a test-size gate.- SHAKEUP's funnel leak finder and A/B test calculator: each step compared with itself last period, the leak sized in lost deals, and every test sized before it starts.

How it thinks

Inputs in, a verdict out, and a stop only where you decide.

What we built

A single source of evidence is an opinion. We check the tracking against the CRM before trusting any rate, keep only issues 2 passes agree on, and size each test so nobody starts one the traffic can't finish.

It reads your company profile if you have one, or asks for only what it needs. Then it works through its steps, checks its own work and ends in a verdict.

The steps

  1. 01

    Check the tracking first

    Form submits in analytics are set against new leads in the CRM. A gap over 20% stops the run, because a ranking built on broken counts points at the wrong step.

  2. 02

    Find the leak

    Each step is compared with itself last period. The biggest fall that clears a significance test is the leak, and the cost is the deals lost had that step held.

  3. 03

    Run the technical pass

    Browsers are compared within each device class, the form is tested on the main browsers, and speed is read from the tool's own labels, never guessed.

  4. 04

    Run the heuristic pass

    Each page in the funnel goes through the landing page critic, or through 6 questions on relevance, clarity, value, friction, distraction and worry.

  5. 05

    Keep what 2 passes agree on

    Findings merge by location. An issue found by 2 passes, or a bug reproduced in a test, is confirmed. Single-pass findings go on a watch list.

  6. 06

    Size the test before running it

    Its script works out the visitors a test needs. Past 8 weeks, it says to test a bigger change or ship the fix and compare periods.

What you need

  • Required

    The funnel's steps, and what counts each one

    Your note → Inferred from your analytics and CRM, then confirmed with you

  • Required

    Step counts for this period and the last, the same length

    Connected analytics and CRM → Exports → A table you paste

  • Required

    CRM counts for the same steps

    A connected CRM → An export → Counts you paste

  • Optional

    Conversion by device and browser, test submissions and speed reports

    Connected analytics → Your exports and test notes → PageSpeed Insights results

  • Optional

    The funnel's pages

    The live pages, if your agent can browse → Screenshots or pasted copy

  • Optional

    Recordings, on-page polls, user tests or sales notes

    Your tools or notes → A message mining handoff

What it can return

  • ISSUES FOUND

    At least 1 issue is confirmed by 2 passes or by a reproduced bug.

  • NO CONFIRMED ISSUES

    The passes ran and no issue was found by 2 of them. It names the pass to add next.

  • TRACKING FIRST

    Analytics and the CRM disagree by more than 20%, or events count twice. Fix the tracking before ranking anything.

  • BLOCKED

    There are no funnel steps, or fewer than 2 passes can run.

Where it stops for you

  • When analytics and the CRM disagree, it stops and points to the tracking audit
  • When fewer than 2 passes can run

Example output

What you get back.

A run for a made-up company. Your output depends on your inputs.

Here's a run for a made-up B2B SaaS company that sells security questionnaire software. Demo form submits fell this quarter.

Verdict

ISSUES FOUND

Coverage: partial (no recordings)

Tracking check: pass (290 submits in analytics, 281 leads in the CRM)

StepThis quarterLast quarterRate nowLast rateChange
Form starts to submits980 to 290950 to 38030%40%Down 26%, real
Submits to meetings290 to 170380 to 22059%58%Held
RankIssueFound byEstimated lossAction
1The cookie banner covers Submit on iPhoneAnalytics, technical, heuristicAbout 64 submits short, at mobile Chrome's rateFix now, no test
22 fields added to the form in JulyAnalytics, heuristicAbout 57 submits short, at last quarter's desktop rateRemove them and compare periods
On testing the shorter form

at about 750 visitors a week, a 20% lift needs roughly 14,000 visitors per version. That's over 30 weeks, so the skill suggests removing the fields and comparing periods instead.

Watch list

the page is labelled "needs improvement" on phones by PageSpeed Insights, found by 1 pass only.

The handoff block the next skill reads
skill: conversion-research
version: 1.0.0
verdict: ISSUES FOUND
coverage: partial (no recordings)
for: demo request funnel
outputs:
  leak: form starts to submits (down 26%, real)
  confirmed: [cookie banner covers Submit on iPhone, 2 fields added in July]
  fix_now: [cookie banner on iPhone]
  tests: [shorter form: over 30 weeks, not testable as planned]
  watch_list: 1
next: [landing-page-critic]
open: [session recordings for the form step]

What good looks like

  • Analytics checked against the CRM before any rate is trusted
  • Every confirmed issue shows which passes found it
  • Every loss labelled an estimate, with its inputs
  • Bugs fixed now, not tested
  • No test planned that the traffic can't finish in 8 weeks

Install

Use it in the tool you already have.

Pick your tool. Each tab has the shortest path we know of.

  1. 01
  2. 02
    In Claude, open Customize, then Skills.
  3. 03
    Choose +, then Upload a skill, and pick the ZIP. Code execution must be on in your settings.
  4. 04
    Ask for it by name, or let Claude use it when a task matches.

Questions

Before you add it.

Why check analytics against the CRM first?

If analytics counts 400 form submits and the CRM holds 190 leads, every rate is wrong. The leak would land on the wrong step, so the run stops until the tracking is fixed.

Why do 2 passes have to agree?

A page review is an informed opinion, and a dip in analytics can be noise. When both point at the same place, or a test reproduces the bug, it's worth fixing first.

Should we A/B test the fix?

It works out the visitors the test needs. If that takes more than 8 weeks, it suggests testing a bigger change, or shipping the fix and comparing periods, labelled as not a controlled test.

What does the script do?

It finds the leak, flags browsers that convert far below their peers, sizes tests and reads finished ones. It's optional, and the skill gives the rough form of each sum without Python.