/ai-visibility-writing

name: ai-visibility-writing
description: Audit, question, suggest, or fact-preservingly revise a press release, blog post, contributed article, or expert explainer so AI answer systems can more easily retrieve, understand, quote, and cite its useful information. Use when someone asks for AI visibility, AI search, answer-engine, AEO, GEO, AI Overview, or ChatGPT citation optimization of supplied writing; when they want an evidence-aware pre-publication audit; or when another Newsjack workflow needs a prose-level AI-discoverability pass. Do not use as a technical SEO audit, rank tracker, publishing system, or promise of rankings, mentions, citations, traffic, or coverage.
metadata:
  category: AI visibility

AI Visibility Writing

Make useful information easier to find and reuse without making the writing worse for people. Treat AI visibility as a probabilistic outcome with strong retrieval, authority, query, and platform confounders—not as a property a rewrite can guarantee.

This skill inherits the ethical floor from `skills/ETHICS.md`. If local instructions conflict with that doctrine, `skills/ETHICS.md` wins. Follow `skills/WHY-NOT-SPAM.md` when the document will support media outreach.

Choose the requested behavior

• **Audit:** diagnose the supplied draft and stop before rewriting.
• **Question:** return only the few answers needed before sound advice is

possible.

• **Suggest:** rank concrete edits without silently applying them.
• **Rewrite:** apply safe, grounded edits and report what changed.

If the request is ambiguous, audit and suggest. Do not rewrite by default.

1. Build the fact ledger first

Before judging style, record the material the output must preserve:

• every number, date, named entity, title, quotation, attribution, comparison,

qualification, and source relationship;

• the document type, intended publisher, audience, and stated purpose;
• supplied links or evidence, including which claim each source supports;
• claims that are promotional, unverifiable from the supplied material, or

likely to require `fact-check` before publication.

Treat the ledger as a ceiling. Never add a statistic, testimonial, citation, link, customer, credential, superlative, causal claim, or other fact merely to make a passage look authoritative. Preserve uncertainty words such as “may,” “estimated,” “in this sample,” and “as of.” If a proposed improvement needs new proof, ask for it instead of writing around the gap.

2. Read the audience and likely queries

State the primary human audience and two to four plausible information needs the piece can honestly answer. Prefer the user's target queries when supplied. Otherwise infer cautiously from the document and label the inference.

Distinguish:

• the reader's question;
• the answer this document can support;
• the answer the organization wishes it could support but cannot yet prove.

Ask a blocking question only when its answer would change the target query, the factual ceiling, the recommended intervention, or whether a rewrite is safe. Group questions by priority and ask no more than five at once. Do not ask for optional analytics, personas, or keywords merely to appear thorough.

3. Separate eligibility from writing

Give two clearly separated diagnoses:

• **Retrieval and authority limits:** indexing, crawl access, snippet

eligibility, page HTML, internal links, canonicalization, structured data, publisher reputation, backlinks, and off-site mentions. These can dominate visibility but are outside a prose-only edit.

• **Writing-level opportunities:** relevance, extractable answers, evidence,

attribution, entity clarity, structure, specificity, and human readability in the supplied text.

Do not imply that prose can repair an unindexed page or weak publisher authority. Google says its ordinary Search eligibility and people-first practices remain foundational for AI features and that no special AI markup is required. Treat platform behavior as changeable and cross-engine findings as non-universal.

4. Select only applicable levers

Use the smallest useful set. Label each recommendation `supported`, `promising`, or `speculative`; the label describes the evidence for the recommendation, not a prediction for this page.

1. **Add unique, verifiable information — promising.** Prefer firsthand data, a defined method, an expert observation, or a real example over a commodity summary. Ask for proof when it is missing. 2. **Put a scoped answer near its descriptive heading — promising.** Make the first useful sentence answer the likely question directly, then add nuance. Do not manufacture certainty or flatten a complex answer. 3. **Tie claims to evidence and attribution — promising.** Name who found what, when, in which population or context, and from which supplied source. More citations are not automatically better. 4. **Clarify entities and relationships — promising.** On first reference, disambiguate organizations, products, people, places, and acronyms when a reader could reasonably confuse them. 5. **Use descriptive sections and coherent chunks — promising.** Give each section one job. Use headings that name the subject and consequence; avoid vague labels such as “Overview” when a precise label is available. 6. **Match format to intent — promising and conditional.** Use steps for a real procedure and tables for genuine comparisons. Do not bolt FAQs, tables, or question headings onto announcements that do not need them. 7. **Expose legitimate freshness — promising and conditional.** State a real publication, update, measurement, or effective date for time-sensitive information. Never add, hide, or alter a date to simulate freshness. 8. **Replace promotion with specific, qualified prose — promising.** Trade empty superiority claims for the exact supported outcome, scope, and limitation. Keep necessary brand voice. 9. **Preserve useful precision — support