Conversion Rate Optimization skill for Codex

Done well: every recommendation names the funnel step it fixes, the evidence behind it (data, recording, survey quote, heuristic), the hypothesis in a testable form, and whether the site has enough traffic to test it or should simply ship the fix.

Ground rules

• Find where people drop before deciding why. A beautiful hero rewrite does nothing if the leak is at the payment step.
• Separate fixes from experiments: broken things (errors, slow pages, unreadable mobile forms) get fixed now; uncertain improvements get tested or shipped with monitoring.
• No dark patterns: fake scarcity, hidden costs revealed at the end, pre-checked add-ons, confirmshaming, hard-to-cancel flows. They lift short-term numbers and create refunds, chargebacks and legal exposure (consumer-protection rules increasingly target them).
• Codex cannot see real users. Ask for analytics exports, recordings or survey data; without them, label findings as heuristic and lower confidence.

Workflow

- [ ] 1. Define the conversion and the funnel steps
- [ ] 2. Quantify drop-off by step, device, source
- [ ] 3. Heuristic review of the leaking steps
- [ ] 4. Technical friction: errors, speed, mobile
- [ ] 5. Qualitative evidence
- [ ] 6. Write hypotheses and prioritize
- [ ] 7. Decide: ship, test, or research more

1. Define

• Primary conversion (paid order, qualified lead, activated signup) and micro-steps. Optimize for the outcome the business values; more low-quality leads is not a win.
• Read AGENTS.md and find the funnel in code (routes, form components, checkout handlers) so findings point to files.

2. Quantify

Build a step table from analytics (GA4 funnel exploration, PostHog funnel, or SQL):

Step	Users	Step conversion	Mobile	Desktop	Notes
Landing page view	40,000	—	64% of traffic
Pricing view	9,200	23%	18%	31%	Mobile gap
Signup start	3,100	34%	33%	35%
Signup complete	1,240	40%	22%	52%	Largest leak, mobile

• Segment by device, traffic source, new vs returning, country. Leaks are usually concentrated in one segment.
• Compare with the step’s own history, not industry benchmarks; benchmarks vary too much by business to be targets.

3. Heuristic review

Review the leaking step against these questions, citing the element and location:

Factor	Questions
Clarity	Can a first-time visitor say what this is, who it is for and what to do next within 5 seconds? Is the primary action obvious?
Relevance	Does the page match the ad, email or query that brought them (message match)?
Motivation	Are the concrete benefits, outcomes and price visible where the decision is made?
Friction	How many fields, steps, required accounts, unexpected costs? Can anything be removed, deferred or autofilled?
Anxiety	Are trust signals present at the point of risk: security, guarantees, cancellation terms, real reviews, support contact?
Distraction	Competing CTAs, navigation, popups or promos that pull away from the main action?

Forms deserve their own pass:

• Remove fields not needed now (ask for company size after signup).
• Correct input types and autocomplete attributes (email, tel, given-name, postal-code, cc-number via the payment provider’s fields).
• Inline validation on blur with specific messages; never clear the form on error.
• Show password rules up front; allow paste; offer social or passkey sign-in where appropriate.
• Show total cost (tax, shipping) before the final step; offer guest checkout.

4. Technical friction

• JavaScript errors and failed API calls at the step (browser console, error tracking, server logs). A validation bug on one browser can halve conversion silently.
• Speed: slow LCP and INP on the step correlate with abandonment; field data targets are LCP ≤ 2.5 s and INP ≤ 200 ms at p75. Diagnosis belongs to core-web-vitals.
• Mobile: tap targets, keyboard covering the CTA, sticky elements hiding fields, horizontal scroll, modals that cannot be dismissed.
• Test the real flow on a phone-sized viewport with a throttled connection.

5. Qualitative evidence

• Session recordings filtered to users who dropped at the step: look for rage clicks, dead clicks, field re-entry, back-and-forth navigation.
• On-page or exit survey with one question (“What, if anything, stopped you from signing up today?”); 50–200 answers are usually enough to see themes.
• Support tickets and sales call notes for objections.
• Quote evidence verbatim in findings (anonymized).

6. Hypotheses and prioritization

Write each as: Because [evidence], we believe [change] for [segment] will increase [metric] because [mechanism].

Score with a simple model and sort:

Hypothesis	Evidence strength (1–3)	Expected impact (1–3)	Effort (1–3, 3 = easy)	Score
Remove 4 optional fields from mobile signup	3 (recordings + drop data)	3	3	27
Show annual price per month on pricing	2	2	3	12

7. Ship, test or research

Check traffic before proposing an A/B test. Approximate users needed per variant for a two-sided test at 95% confidence and 80% power:

from math import ceil
from statistics import NormalDist

def users_per_variant(baseline: float, relative_lift: float, alpha=0.05, power=0.8) -> int:
    p1, p2 = baseline, baseline * (1 + relative_lift)
    z = NormalDist().inv_cdf(1 - alpha / 2) + NormalDist().inv_cdf(power)
    var = p1 * (1 - p1) + p2 * (1 - p2)
    return ceil(z * z * var / (p2 - p1) ** 2)

print(users_per_variant(0.04, 0.10))  # 4% baseline, +10% relative lift -> about 39,000 per variant

Situation	Decision
Clear bug or accessibility failure	Fix and ship; monitor the step
Enough traffic for the minimum lift you care about within 2–4 weeks	A/B test (design in ab-testing)
Too little traffic	Ship changes backed by strong evidence, monitor before/after with seasonality in mind, or test on a higher-traffic upstream step
Weak evidence and low traffic	Research more (surveys, user tests) before changing

Gotchas

• Optimizing a micro-conversion (clicks on “Start trial”) can lower the real one (activated trials). Always track the downstream metric.
• Peeking at test results daily and stopping on the first significant day inflates false positives.
• Before/after comparisons are confounded by campaigns, seasonality and traffic mix; segment and annotate.
• Copy changes that promise more than the product delivers raise conversion and churn. Landing page copywriting belongs to landing-page-copy.
• Popups that appear immediately on mobile hurt both conversion and Google’s page experience expectations.

Output template

# CRO audit: signup flow (data 2026-09-01..30)
**Conversion:** activated signup (created first project within 24h)
**Biggest leak:** signup form, mobile: 22% complete vs 52% desktop (n=2,050 mobile starts)

## Fix now
1. Mobile keyboard hides "Create account" button (recording ids 4, 9, 17); sticky CTA
   above keyboard. File: components/SignupForm.tsx.
2. Password rule revealed only after submit; show rules inline.

## Test (traffic allows ~10% MDE in 3 weeks)
3. Remove company size + role fields from signup, ask in onboarding. Score 27.

## Ship with monitoring (insufficient traffic to test)
4. Add cancellation terms under pricing CTA (survey: 31% cite "lock-in" worry).

## Research
5. Pricing page desktop vs mobile gap: run 5 remote user tests.