mbfinotti/advertising-skills / skills/paid-landing-page-audit
paid-landing-page-audit
VerifiedAudit a landing page receiving paid traffic against post-click conversion best practice and return a prioritised fix list - message match to the ad, above-the-fold clarity, form friction, trust, speed, mobile, policy risk - with every finding labelled by evidence class and no statistical claims on thin data. Use whenever the user says their ads get clicks but no conversions, asks why a landing page isn't converting, or mentions post-click experience, message match, conversion friction, form friction, or CPA rising after the click - even if they never say 'landing page audit'. Covers B2B and B2C on any platform. Audit only. Do NOT use for pre-click account problems - use mbfinotti/advertising-skills@ad-account-diagnostic instead.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 24.8 KB |
| evals/evals.json | config | 37.4 KB |
| references/evidence-and-benchmarks.md | markdown | 11.2 KB |
| references/page-checks.md | markdown | 10.9 KB |
| references/report-template.md | markdown | 5.9 KB |
references/evidence-and-benchmarks.md
View on GitHub# Evidence and Benchmarks
Read this before citing any number or naming any framework in a report. CRO is full of confidently repeated numbers with broken citation chains.
An audit that cannot tell its evidence tiers apart is how bad advice spreads. Every entry below states its tier and the caveat that must travel with it.
## Verified numbers - safe to state, with attribution
| Figure | Source | Caveat |
| ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------ |
| LCP ≤ 2.5s, INP ≤ 200ms, CLS ≤ 0.1 = "good" | Google Core Web Vitals (web.dev) | Measured at the 75th percentile of field data, mobile and desktop segmented |
| Contrast 4.5:1 normal text, 3:1 large text (AA); target size 24×24 CSS px (AA), 44×44 (AAA) | WCAG 2.2 / 2.1 (W3C) | AAA target size is the practitioner-preferred floor for paid mobile traffic |
| Median landing-page conversion rate 6.6%; industry medians 3.8%-12.3%; paid search ~10.9%, paid social ~12%, email ~19.3% | Unbounce Conversion Benchmark Report 2024 (41,000+ pages, 464M visitors, 57M conversions) | Vendor dataset, self-selected; "conversion" definitions vary by page type; context, never a target |
| Mobile ≈ 83% of landing-page visits but converts ~8% worse than desktop | Unbounce 2024 | Same dataset caveat; check this page's own split before applying |
| Pages at 5th-7th grade reading level convert ~11.1% vs ~7% at 8th-9th grade | Unbounce 2024 | The report itself presents this as correlation, not causation - say so when citing |
| Average checkout: 5.1 steps, 11.3 form fields, vs an achievable ~8 fields; 17% abandon over complexity/length; 42% typed a full name into "First Name"; 30% hesitated at "Address Line 2" | Baymard Institute, June 2024 | **B2C ecommerce checkout data only** - never present as B2B lead-form benchmarks |
| Extra costs revealed late = top stated abandonment reason (~48%); ~43% of abandoners were "just browsing" | Baymard survey data | The browsing share means the actionable abandonment is far smaller than the headline ~70% cart-abandonment figure |
| ~57% of viewing time above the fold, ~74% in the first two screenfuls | NN/g eye-tracking, 2018 | Kills both "people don't scroll" and "the fold is dead" |
| F-shaped scanning: top stripe, shorter second stripe, left vertical stem | NN/g 2006 (232 users), later reconfirmed | The empirical reason to front-load the message - not a layout to design toward; clear hierarchy should disrupt it |
| Visual first impressions form in ~50ms and predict later judgments | Lindgaard et al. 2006 | About visual appeal, not comprehension; do not stretch it into "users decide to buy in 50ms" |
| Only ~10-33% of experiments improve their target metric | Kohavi et al. 2009 (Microsoft/Google); Optimizely 2026 (127,000 experiments: 12% significant improvement) | The reason this skill never promises a lift: most fixes lose when tested, and "winning" lifts shrink on replication (winner's curse) |
| "Purchase Rate = Desire − (Labor + Confusion)"; reviewer checklist: Conversion, Clarity, Expansion, Disbelief, Interest, Brevity | Julian Shapiro's landing-page guide | A lens and a review script, not a formula with units |
## Frameworks - name them only with their caveats attached
Scoring frameworks, ranked - `PXL > ICE == PIE` on evidence bought per unit of scoring effort: all three cost the same half-hour, and only PXL spends it on questions the proposer cannot answer in their own favour. ICE and PIE tie exactly rather than pending a decision: same three-input 1-10 shape, same half-hour, and the same documented weakness (two of three inputs guessed by the person proposing the fix), so nothing separates them but the labels on the inputs. The diagnostic lenses below (LIFT, MECLABS, the objection families) are deliberately left unranked: they answer different questions and are read together, so an ordering between them would be false precision.
- **LIFT Model** (Chris Goward, WiderFunnel, 2009). Value Proposition as the central lever; Relevance, Clarity, Anxiety, Distraction, and Urgency as forces acting on it. Use it to organise diagnosis. Mandatory caveat: it quantifies nothing, and a weak value proposition cannot be optimised into a strong one.
- **MECLABS Conversion Sequence Heuristic** (Flint McGlaughlin): C = 4m + 3v + 2(i−f) − 2a. Mandatory caveat, per MECLABS itself: a thought tool, not an equation to solve. The coefficients express a thinking _sequence_ - motivation first, then value, then incentive net of friction, then anxiety - not literal multipliers. Motivation is largely pre-existing and external: a page refers to it. It cannot create it. Any audit computing a "score" from this formula has misread it.
- **PXL** (Peep Laja, CXL). Roughly ten mostly-binary questions summed into a total, built to strip the subjectivity out of ICE and PIE: is the change above the fold, noticeable within five seconds, on a high-traffic page, supported by user testing / qualitative feedback / mouse-tracking / analytics. Ease is bracketed into time ranges; noticeability weighted double. This skill's prioritisation borrows its shape.
- **ICE** (Sean Ellis: Impact, Confidence, Ease) and **PIE** (Chris Goward: Potential, Importance, Ease), each scored 1-10. Shared documented weakness, and why this skill prefers PXL's shape: two of the three inputs are guesses, and the proposer usually scores their own idea.
- **Big-five objection families** (trust, price, fit, timing, effort - practitioner consensus). A compact taxonomy for the trust/objection check; the discipline it enforces is that every named objection gets an evidence-backed counter placed where the objection arises.
## Contested - direction only, never a number
- **Per-form-field conversion cost.** One published source claims a flat ~11% per field; another claims tiered buckets (3 fields baseline, 4-6 costs 10-25%, 7+ costs 25-50%). They disagree in both size and shape. State: every field costs conversion, the cost is non-linear, and no trustworthy universal percentage exists.
- **Single-vendor case-study lifts** (a 66% message-match lift, a ~3x headline-rewrite lift, and every number like them). They demonstrate the mechanism, not a transferable effect size. Cite the mechanism; refuse the number.
- **The "+81% conversions / −38% sales cycle / −28% CAC / +175% referrals" clarity statistic.** Circulates with no traceable source. Do not repeat it, even hedged.
## Folklore - refuse on sight
- The 8-second/goldfish attention span: broken citation chain, the BBC's 2017 investigation found the attribution untraceable.
- "Red beats green" or any universal button colour: contrast matters, hue does not.
- The 3-click rule: no supporting evidence, information scent per click is what matters.
- "People don't scroll" and its overcorrection "the fold doesn't matter": both falsified by the NN/g attention data above.
- "Shorter is always better": length should match offer complexity and traffic temperature, shorter B2B forms can degrade lead quality.
- "Reducing 11 fields to 4 lifted conversions 120%" and its variants: secondary citations with no accessible primary experiment.
- The "$300M button" retold as a button change: the verified lesson is removing forced-registration friction discovered through user research, a single anecdote, not a controlled result.
## Statistical discipline for any claim this audit makes
- Platform-reported and on-site conversions differ by attribution window and view-through; they are different measures of different things. Never mix them in one calculation, and state which one every observation uses.
- A "conversion rate" from under ~1,000 sessions or ~30 conversions is noise wearing a percentage sign - the Volume Floor in SKILL.md governs.
- When a test is recommended in "Test, don't guess", the honest framing is Optimizely's and Kohavi's: most tests lose, a winner's measured lift overstates its true effect, and underpowered "wins" are often wrong in sign or wildly exaggerated in magnitude (Gelman & Carlin's Type S / Type M errors). Recommend a sample-size calculation, full business cycles, and no early stopping.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| rea-analysis-optimizationmorluto/reaOptimize REA providers and shared analysis workflows for agent context, latency, CPU, and memory while preserving correctness and Evidence. Use for analysis bottlenecks and cache or query design; CI/test-suite optimization belongs to rea-ci-tests-audit. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| rea-ci-tests-auditmorluto/reaAudit REA CI and test suites for correctness, fixture drift, false passes, and measured performance; use for suite or workflow investigations and efficiency changes, rather than ordinary feature testing or a single test failure. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| rea-tool-designmorluto/reaAudit, investigate, design, or change REA investigation tools, CLI/MCP contracts, provider capabilities, and Evidence semantics. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-auditDietrichGebert/ponytailQuality audit of the whole repo: bugs, security, real load, missing tests, speed, and what to delete. Most important first. | Stars: 196k | Updated | Claude CodeCodex | Verified | Install |