AI Content Audit Skill
Teams that scaled content with AI are discovering the bill: libraries full of fluent, structurally identical, information-free pieces that readers bounce off, search engines quietly demote, and — worst — that erode the trust the good content earned. This skill audits the library for slop with named signals, triages it, and installs the gate that stops the refill.
What This Skill Produces
- An audited inventory with per-piece verdicts: keep / enrich / rewrite / delete-and-redirect
- The detection signals found, quoted — so verdicts are checkable, not vibes
- A triage plan sequenced by traffic and trust impact
- A publishing quality gate for AI-assisted content going forward
Required Inputs
Ask for (if not already provided):
- The corpus — pieces or URLs to audit (or a sample; state the sampling), with publish dates
- Performance data if available — traffic, engagement, rankings over time (the audit works without it, but verdicts get sharper)
- What the content is for — SEO, docs, thought leadership, support deflection (the quality bar differs)
- Production context — when AI-assisted publishing started, at what volume (the before/after seam is diagnostic gold)
Detection Method
Slop isn't "AI wrote it" — it's content with nothing inside. Audit each piece for the signals, quoting instances:
- Information density — the core test: delete every sentence that any competitor could have written, and measure what's left. Slop survives at <20%. Look for: zero proprietary data, zero named examples, zero opinions with an owner, zero specifics a reader could act on.
- Structural monoculture — the same skeleton repeating across pieces (intro-restating-the-title → 5 H2s → "in conclusion"); listicles whose items are definitions, not judgments; FAQ sections answering questions nobody asked.
- Hedged voicelessness — "it's important to note", "in today's fast-paced world", both-sides-ism on questions the brand should have a stance on; the absence of anything a lawyer would ever have flagged.
- Fluency without grounding — claims with no source, stats with no year, "studies show" with no study; internally contradictory sections (the tell of stitched generations).
- Reader evidence, where data exists — engagement collapse relative to the library's pre-AI baseline, rising pogo-sticking, ranking decay cohort-matched to the AI-volume era. Correlate verdicts with the seam from the production context.
Verdicts: Keep (dense, differentiated — AI-assisted or not; the audit is provenance-blind on keepers) · Enrich (sound skeleton, hollow middle — inject data, examples, stance) · Rewrite (topic worth owning, execution beyond saving) · Delete & redirect (nothing inside, no traffic worth saving — thin pages drag the domain).
The Quality Gate (prevention)
For AI-assisted publishing going forward, every piece passes before shipping:
- The density test — a named reviewer deletes the anywhere-sentences; ≥50% must survive
- One of three must be present: proprietary data/experience · a named example with specifics · a defensible stance someone could disagree with
- Claims carry sources; stats carry years
- The read-aloud test — one paragraph aloud; if it sounds like nobody, it ships under nobody's name and that's the problem The gate is a checklist with an owner, not a sentiment.
Output Format
AI Content Audit: [property] — [n] pieces ([sampling noted])
Headline: [keep/enrich/rewrite/delete counts + the one-line diagnosis]
The seam: [what changed at the AI-volume transition, if data allows — cohort chart described]
| Piece | Traffic | Signals found (quoted) | Verdict |
|---|
Triage plan: [sequence: high-traffic enrichables first → deletions batched with redirects → rewrites scheduled; owner + dates]
The quality gate: [the checklist above, adapted to this org, with its named owner]
Quality Checks
- Every non-keep verdict quotes at least one concrete signal from the piece
- The audit is provenance-blind on keepers — good AI-assisted content is not penalised for its origin
- Deletions come with redirect targets, not just removal
- The triage is sequenced by traffic × trust impact, not by ease
- The gate has an owner and a pass bar, not aspirations
Anti-Patterns
- Do not use "AI-detector" scores as evidence — they misfire both ways; the signals are about emptiness, not origin
- Do not delete by publish-date cohort — some AI-era pieces are good and some human classics are slop
- Do not enrich everything — a piece with no reason to exist gets deleted, not decorated
- Do not install the gate without an owner — a checklist nobody signs is the slop pipeline with extra steps
- Do not frame the report as anti-AI — the finding is a quality failure that AI made cheap to commit at scale
