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
Anti-Patterns
1---2name: ai-content-audit3description: Audit a content library, docs site, or blog for AI-generated filler that's eroding trust and search performance — and triage what to fix, rewrite, or delete. Use when asked to find slop in a content library, audit AI-written content quality, explain why content engagement or rankings dropped after scaling with AI, or set a quality bar for AI-assisted publishing. Produces an audited inventory with per-piece verdicts, the detection signals used, a triage plan, and a publishing quality gate that prevents recurrence. For a single article's AI-citability use aeo-optimizer; for the strategy itself use content-calendar or seo-content-brief.4---5
6# AI Content Audit Skill
7
8Teams 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.
9
10## What This Skill Produces
11
12- An **audited inventory** with per-piece verdicts: keep / enrich / rewrite / delete-and-redirect
13- The **detection signals** found, quoted — so verdicts are checkable, not vibes
14- A **triage plan** sequenced by traffic and trust impact
15- A **publishing quality gate** for AI-assisted content going forward
16
17## Required Inputs
18
19Ask for (if not already provided):
20- **The corpus** — pieces or URLs to audit (or a sample; state the sampling), with publish dates
21- **Performance data if available** — traffic, engagement, rankings over time (the audit works without it, but verdicts get sharper)
22- **What the content is *for*** — SEO, docs, thought leadership, support deflection (the quality bar differs)
23- **Production context** — when AI-assisted publishing started, at what volume (the before/after seam is diagnostic gold)
24
25## Detection Method
26
27Slop isn't "AI wrote it" — it's *content with nothing inside*. Audit each piece for the signals, quoting instances:
28
291. **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.
302. **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.
313. **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.
324. **Fluency without grounding** — claims with no source, stats with no year, "studies show" with no study; internally contradictory sections (the tell of stitched generations).
335. **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.
34
35**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).
36
37## The Quality Gate (prevention)
38
39For AI-assisted publishing going forward, every piece passes before shipping:
40- **The density test** — a named reviewer deletes the anywhere-sentences; ≥50% must survive
41- **One of three** must be present: proprietary data/experience · a named example with specifics · a defensible stance someone could disagree with
42- **Claims carry sources**; stats carry years
43- **The read-aloud test** — one paragraph aloud; if it sounds like nobody, it ships under nobody's name and that's the problem
44The gate is a checklist with an owner, not a sentiment.
45
46## Output Format
47
48### AI Content Audit: [property] — [n] pieces ([sampling noted])
49
50**Headline:** [keep/enrich/rewrite/delete counts + the one-line diagnosis]
51
52**The seam:** [what changed at the AI-volume transition, if data allows — cohort chart described]
53
54| Piece | Traffic | Signals found (quoted) | Verdict |
55|---|---|---|---|
56
57**Triage plan:** [sequence: high-traffic enrichables first → deletions batched with redirects → rewrites scheduled; owner + dates]
58
59**The quality gate:** [the checklist above, adapted to this org, with its named owner]
60
61## Quality Checks
62
63- [ ] Every non-keep verdict quotes at least one concrete signal from the piece
64- [ ] The audit is provenance-blind on keepers — good AI-assisted content is not penalised for its origin
65- [ ] Deletions come with redirect targets, not just removal
66- [ ] The triage is sequenced by traffic × trust impact, not by ease
67- [ ] The gate has an owner and a pass bar, not aspirations
68
69## Anti-Patterns
70
71- [ ] Do not use "AI-detector" scores as evidence — they misfire both ways; the signals are about emptiness, not origin
72- [ ] Do not delete by publish-date cohort — some AI-era pieces are good and some human classics are slop
73- [ ] Do not enrich everything — a piece with no reason to exist gets deleted, not decorated
74- [ ] Do not install the gate without an owner — a checklist nobody signs is the slop pipeline with extra steps
75- [ ] Do not frame the report as anti-AI — the finding is a *quality* failure that AI made cheap to commit at scale
Mohit Aggarwal