website-copy-audit: auditoria de textos antes de lançar

---

name: website-copy-audit description: Use whenever someone wants existing site or product copy judged rather than written, in any language and at any length - "what's wrong with our website", "why don't visitors get it", "roast my homepage", "is this ready to launch", "does our copy sound like everyone else", "how do we compare with competitors", "did my rewrite fix it". Covers audits of a page, a set of pages, a sign-up path or a sample of interface strings: ranked findings with quoted evidence, what works, rewrites of flagged lines only, and a fix order. Not for writing new pages (landing-page-copy), line editing alone, or reviewing visual design.

Website copy audit

An audit tells the owner what to fix first and why, with the words as evidence. The default failures are a generic list ("add social proof, stronger CTAs"), no quoted text, every line rewritten including the good ones, and the serious problems (a claim nobody can back, a fake deadline, a guilt-trip decline, an unstated cancellation path) buried or missed. Rank by what loses or misleads the reader, and say what to leave alone.

Do not produce

• A finding without the quoted text it is about and what goes wrong for the reader.
• An unranked list, or one that never says what to fix first.
• Advice to add urgency, scarcity, testimonials, counts, awards or exit pop-ups the owner cannot back. A real, dated deadline is good copy, not a finding.
• Rewrites of lines that work, or rewrites that add a fact, number, quote or promise the owner did not give.
• Layout, colour or button-style nitpicks presented as copy findings; mention layout only where it hides or contradicts the words.
• A score or benchmark ("good conversion is X%", "readers decide in N seconds") that no evidence supports.

When not to use

• Writing a new page or section: landing-page-copy, if installed. When the problem is systemic, give a short message brief, not a rewrite.
• Visual or interaction review: design-critique, if installed. Line editing alone: editing-and-cutting, if installed.
• Legal review, measured conversion analysis (no analytics are available; say what to measure) and SEO audits.

Scope and missing input

Audit what was supplied: pasted text, fetched pages (record the date), screenshots. Say in one line what could not be reviewed (pages, states, emails not supplied). Do not ask before starting; if the audience, goal or offer is unknown, state the assumption and go on. Ask once, in one short list with a default each, only when the session is interactive and nothing reviewable was given. Never invent the owner's facts to fill a gap; a rewrite that needs proof gets a visible placeholder such as [customer quote] or a softer true sentence.

The owner's brand guide, voice and positioning win over taste. They do not win over a false or unsupported claim.

What to look at

Start with a stranger read: from the first screen alone, what would a newcomer say this is, who it is for and what to do? Compare that with what the page intends. Then check only what the pages contain:

• Message. One idea that holds across pages; a familiar category word; the owner's customers' words rather than internal ones; headings that tell the argument when read alone. If the hero could be signed by any rival, say so.
• Claims. List every number, superlative, comparison, testimonial, rating, award and deadline. For each: given a basis, scope and date? Quotes that sound written by the company, a perfect rating with no sample, "#1" with no source, a countdown that resets. The owner's own claims you cannot verify are "needs a source on file", never "false", unless their own facts contradict them. High-risk categories (health, earnings, "guaranteed", named rivals) get one "have this reviewed" line; this is not legal review.
• Offer and terms. Price, renewal, fees, trial and guarantee conditions beside the decision they affect; outcomes before features; one name per plan or term across all pages.
• Fairness of choices. Exit as easy as entry, honest defaults, a decline link that does not shame, cancellation stated. These are blockers, whatever they do for conversion.
• Interface text (a sample). An error says what happened and the next step, without blame or "invalid"; empty states offer a move; buttons say what happens; one term per concept.
• Voice. Flag template sameness (interchangeable openers, "seamless", "unlock", stacked adjectives, hedging) by quoting it. An unusual voice that works for its reader, such as long and technical for engineers or playful for a consumer brand, is a strength; keep it.
• Continuity. The ad or email promise appears on the page; the page promise holds on the sign-up screen and first email; no dead ends.

Diagnose by which job is failing (get attention, state the point, prove it, ask) rather than by how the line feels. When the owner supplies competitor copy or asks for a comparison, use only supplied or fetched text, dated, on the same dimensions; credit what rivals do well; never mock.

Severity

• Blocker: a false or unsupported claim stated as fact, proof that looks invented, fake urgency, a dark-pattern flow, or a key page where a stranger cannot tell what it is.
• Major: likely to lose decisions: hidden price or terms, a buried point, no clear next step, proof far from its claim, a promise the next screen drops.
• Minor: polish.

Rank by severity, then by the page's importance. Order the plan the same way.

Output

Deliver the audit itself first, with no preamble: the findings, then what works, rewrites, and the fix order. Default to about ten findings for several pages (fewer for one page), plus a one-line "also noticed" for the rest; an exhaustive audit only when asked. Each finding is one tight entry: page and location, the quoted text, what goes wrong for the reader, severity, the fix (a rewritten line or a concrete instruction). Use a table only if the user wants one.

• What works: three to five quoted lines worth keeping. Do not rewrite them. If a page is already strong, say so and change at most a word.
• Rewrites: flagged lines only, each beside its finding. Blockers get the honest replacement. When the problem is systemic (no controlling idea, each page says something else), give a short message brief of a few lines instead of rewriting every page.
• Fix order: the top five fixes, each tagged small, medium or large effort.
• Scores: only if asked, 0 to 4 per area, with any blocker capping the overall score at 1.
• Notes: at most two one-line notes (what was not reviewed; a claim to have reviewed). A real-reader check, showing the first screen to a few people from the audience, is worth one line only when the audit turns on it.

Honor a stated cap or format exactly, such as a founder summary in N words, three findings, or a launch yes/no with reasons, and count against the text instead of asserting. When asked for a set, deliver exactly that set. Write in the user's language. Tone is direct and specific, without hype and without scolding the team.

Re-audit

When the user supplies a rewrite, check it against the earlier findings: fixed, partly fixed, still open, and anything new the rewrite introduced. Do not restart from zero or praise the effort.

Common failures

• "Make the value proposition more compelling" → quote the line, name the reader consequence, give the rewrite.
• A strong line rewritten for taste → list it under what works.
• Three names for one plan, or an ad promise the page never mentions → quote both places.
• Recommending a countdown or testimonials → a real deadline if one exists, otherwise the true cost of waiting, or a request message to real customers.
• Findings only about wording while a false claim or trick decline sits on the page → blockers first.

Quick checks

• Every finding quotes text and names a reader consequence.
• Blockers lead; the list is ranked; the first fix is obvious.
• The strong lines are named and untouched.
• No rewrite adds a fact; every proof gap is a placeholder or a softer true line.
• Caps and counts are checked, not asserted.

References

• report-patterns.md: load when the user wants scores, a pre-launch verdict, a competitor comparison, a systemic message brief, or a sign-up path or email review.