conversion-audit

The full skill.

Original instructions from the publisher’s SKILL.md

# Conversion Audit

Every page exists to move a specific visitor toward a specific action. Most pages fail not because they're ugly or broken, but because they never decided what that action is — or decided, then buried it under everything else the owner wanted to say. This skill audits a page the way a visitor experiences it: skeptical, distracted, five seconds from leaving.

The output is not a redesign. It is a verdict on **intent clarity** plus a prioritized list of moves — things to cut, move, add, or rewrite — each justified by visitor behavior, not taste.

## The Core Questions

Everything in this audit reduces to answering these, in order. Do not skip ahead; each depends on the previous.

1. **Who lands here?** Not "the target market" — the actual arriving human: what did they click to get here (ad, search result, referral, nav link), what do they already know, what mood are they in?
2. **What is the ONE action this page wants from them?** One. If the honest answer is "three things," that is Finding #1 and everything downstream inherits it.
3. **What must they believe before they'll take it?** Write the belief chain explicitly: e.g. "these people are real → they've done this for businesses like mine → it worked → it won't be a nightmare to work with them → booking a call is low-risk." Actions have prerequisites; pages that skip a link in the chain leak visitors at that link.
4. **What on the page builds each belief — and what builds nothing?** Every section either advances a belief, removes an objection, or competes with the goal. There is no neutral content: attention is finite, so a section that does no work does negative work.
5. **When the belief is built, is the action obvious, easy, and low-fear?** The CTA question — visibility, wording, competition, and what the visitor imagines happens after clicking.

## Phase 0 — Establish the Goal

Before reading a single section, determine the page's intended action:

- If the user stated it, use it.
- Otherwise infer from the page itself: the primary CTA, where CTAs point, the funnel position. State your inference explicitly ("This page appears to exist to get a strategy call booked") so the user can correct it.
- If you genuinely cannot tell — multiple CTAs with equal weight, no CTA, contradictory asks — **stop and report that as the headline finding.** A page with no discernible goal cannot be optimized, only decided.

Also establish the **arrival context**: enumerate the realistic entry paths (nav from home? Google search for what query? paid ad with what promise? backlink?) and what each arriving visitor already believes. A visitor from a "web design Houston" search and one clicking from the pricing nav item need different first sentences; note when the page can only serve one of them.

## Phase 1 — Gather the Evidence

Pick per what's available; combine when possible:

- **Code in repo**: read the page component(s) top-to-bottom in render order. Extract the actual copy — headlines, subheads, CTA labels, section order. Note what's above the likely fold. Follow every link/CTA target to know where it actually goes.
- **Live page**: use browser tools to load it, screenshot the initial viewport (the real five-second view), scroll and capture the full page, click the primary CTA to verify the next step works and matches the promise.
- **Pasted copy/screenshots**: work from what's given; flag what you can't verify (fold position, link targets).

Record the raw material as a **section inventory**: for each section in order — its message in one sentence, the belief it serves (or "none"), and its CTA if any.

## Phase 2 — Cold-Visitor Persona Swarm

Your own read is contaminated: you know what the page is trying to say. Spawn parallel agents (Agent tool, one message, concurrent) who don't. Give each agent ONLY the page content/screenshots and its persona — **do not tell them the page's goal or your hypotheses.**

Default panel (adjust to the site's audience):

| Persona | Brief |
|---|---|
| **5-second skimmer** | You will read only the headline, subhead, and whatever is visually loud, then decide: stay or leave. Report what you think this company does, for whom, and what you'd do next. |
| **Skeptical buyer** | You've been burned by vendors before. Hunt for reasons NOT to trust this page: vague claims, missing proof, stock-photo energy, anything that smells templated. Report every objection the page raises and whether it answers them. |
| **Ready-to-act visitor** | You already want this service; you arrived to take action. Time how long (in scrolls/clicks) until you can. Report every obstacle between arrival and done. |
| **Wrong-fit visitor** | You're adjacent to the target market but not it. Report whether the page tells you it's not for you (good) or wastes your time / lets you book a bad-fit call (bad). |
| **Comparison shopper** | You have three tabs open with competitors. Report what, if anything, you'd remember about this one in ten minutes, and which tab you'd go back to. |

Require structured returns from each: `first_impression` (what page is, one sentence), `would_do_next` (the honest action, including "close tab"), `bail_points` (where and why), `unanswered_questions`, `strongest_moment`, `weakest_moment`.

Where persona reports agree, that's signal. Where they contradict your Phase 3 analysis, the personas win — they're the visitor.

## Phase 3 — Conversion-Path Analysis

Now analyze deliberately, dimension by dimension:

**Five-second test.** From the initial viewport alone: can a stranger say what's offered, for whom, and why care? Headlines that are clever-but-opaque fail here; personality is allowed only after clarity is achieved.

**CTA inventory.** Count every clickable ask on the page. Classify: primary action, secondary action, exit ramp (nav links, social icons, outbound links). Compute the competition: how many things fight the primary CTA for the click? Check CTA copy — does it say what happens next ("Book a strategy call") or is it vapor ("Get started", "Learn more")? Check repetition: is the primary CTA re-offered after each belief-building section, or does the visitor have to scroll back up?

**Belief-chain coverage.** Take the chain from Core Question 3 and map sections onto it. Mark every link with no supporting content (leak), and every section supporting no link (ballast). Order matters too: proof before claims land weaker than claims then proof; pricing before value framing raises objections early.

**Objection ledger.** List the top 5 objections this specific visitor has (too expensive, too small a shop, will this take forever, what if it doesn't work, why you and not the other tab). For each: where does the page answer it, or does it not?

**Momentum and friction.** Where does the page's energy die — a wall of text, a section that changes topic, a detour to something the owner cares about but the visitor doesn't? Where does it ask for effort (long form, unclear next step, forced choice) before earning it?

**Post-click integrity.** Follow the primary CTA. Does the next page/step deliver what the button promised? A great page with a broken or bait-and-switch next step converts at zero.

## Phase 4 — Synthesize the Report

Structure, in order:

1. **The one thing** — "This page wants the visitor to ___." (Or: "This page has not decided; that is the finding.")
2. **Verdict** — 2–3 sentences: does the page get a cold visitor there, and where does it mainly leak?
3. **What's working** — genuinely, so it doesn't get broken in revision. Be specific.
4. **Prioritized moves** — each tagged `CUT` / `MOVE` / `ADD` / `REWRITE`, ordered by conversion impact, not ease. Each move states: the element, the change, and the visitor-behavior reason ("the skeptical-buyer persona bailed here because…"). Severity-group them: *leaks* (actively losing people) → *drag* (slowing belief) → *polish* (marginal).
5. **Belief-chain map** — the chain with each link marked covered/weak/missing.
6. **What I couldn't verify** — analytics, real traffic sources, actual conversion data. Never fabricate these; recommend measuring where it would settle a judgment call.

Ground every claim in evidence: quote the actual copy, cite the file/line or screenshot, name the persona whose reaction supports it. No "best practices say" without a mechanism — always the *why* in terms of what a visitor believes, feels, or does.

## Scaling

- **Quick pass** (user asks for a gut-check, single page): skip the swarm, run Phases 0–1 and 3–4 yourself, but still write the belief chain and CTA inventory — those are the skill.
- **Full audit** (default for "audit this page/site"): full persona swarm per page.
- **Multi-page/site-wide**: run Phase 0 across all pages first to build the intent map (each page's one action and how pages hand off to each other), then audit the highest-traffic/highest-stakes pages fully. Cross-page finding to always check: do all roads lead to the same conversion door, or does the site scatter its asks?