The 15-Minute AI Content Audit Before You Publish
An AI-assisted draft can be grammatically clean and still be unpublishable. The danger is not a stray em dash or a suspiciously tidy list. It is a page that makes no original contribution, cannot support its claims, and sounds as if nobody would notice if the byline changed.
This audit is a practical last pass for writers, editors, marketers, and small teams. Start with a 15-minute timebox, then extend the review whenever the subject, claims, or source risk demands it. It does not try to prove whether a machine touched the draft. It asks the more useful question: has a responsible human made this worth publishing?
First, Set the Standard
Google’s guidance does not ban AI-assisted writing. It asks publishers to focus on accuracy, quality, relevance, and context about how content was created. Its spam policy draws the line at scaled, unoriginal pages made mainly to manipulate rankings, regardless of whether people or software produced them.
Those are useful minimums, but a reader deserves more than “not spam.” A publishable article should do at least one thing a generic generated answer cannot: report an observation, make a defensible judgment, show a tested process, organize primary evidence, or explain a hard-earned lesson.
If you need examples of the patterns that make a draft feel weightless, open why AI slop feels so obvious before you begin. If you need line-level repair tactics, keep how to humanize AI writing beside this checklist.
The 100-Point Anti-Slop Scorecard
Score each category from 0 to 5, then multiply by its weight. Do not award points for intent. Award them only for evidence visible in the draft or its working files.
| Category | Weight | 0 points | 3 points | 5 points |
|---|---|---|---|---|
| Original value | ×5 | Repeats common summaries | Adds a useful framework or example | Contains firsthand work, original analysis, or a genuinely useful synthesis |
| Claim support | ×4 | Claims have no traceable sources | Most consequential claims are linked | Every material claim traces to an appropriate primary or authoritative source |
| Specificity | ×3 | Could fit any topic or brand | Includes some concrete names, dates, or steps | Uses precise details throughout and explains why they matter |
| Reader outcome | ×3 | No clear job for the reader | Gives advice but leaves gaps | A defined reader can complete a useful task after reading |
| Human judgment | ×2 | Avoids decisions and trade-offs | Makes a few recommendations | Clearly states choices, limits, trade-offs, and reasons |
| Voice and rhythm | ×2 | Uniform, generic, padded | Mostly natural with a few dead zones | Sounds consistent, intentional, and appropriate to the publication |
| Transparency | ×1 | Ownership and process are unclear | Byline and sources are present | Byline, relevant expertise, methods, and material AI use are clear where useful |
The maximum is 100. Use these gates:
| Score | Decision | Required action |
|---|---|---|
| 85–100 | Publish after proofread | Fix only factual or mechanical errors; do not sand away the voice |
| 70–84 | Hold for one focused edit | Repair the two lowest-scoring categories, then rescore |
| 50–69 | Rewrite around original evidence | Keep useful research, but rebuild the structure and argument |
| Below 50 | Do not publish | Return to reporting, testing, or source collection before editing prose |
A high total cannot cancel a zero in claim support. If the piece makes unsupported medical, legal, financial, safety, or reputation-sensitive claims, it is a hold no matter what the arithmetic says.
The 15-Minute Audit, Step by Step
Minute 0–2: Write the reader contract
At the top of your working copy, complete this sentence:
This article helps [specific reader] do [specific job] by giving them [specific evidence, process, or decision tool].
If you cannot finish that sentence without using “learn about,” the draft probably has no outcome. Rewrite the promise until it describes an observable result: choosing an option, checking a claim, repairing a draft, or completing a task.
Then compare the title, introduction, headings, and conclusion with that contract. Delete sections that answer a nearby question merely because the model produced them. A shorter page that completes one job beats a comprehensive-looking page that completes none.
Minute 2–5: Mark every checkable claim
Highlight names, dates, numbers, quotations, causal claims, policy descriptions, and statements about what an organization “recommends.” For each one, open the linked source and verify:
- The source says what the draft says.
- The source is current enough for the claim.
- The link goes to the original document when one is available.
- The draft preserves important limits and context.
Do not cite a search snippet, an AI answer, or a page that merely cites another page. If the trail ends in “studies show,” either find the study or remove the claim. Google’s people-first guidance explicitly asks whether content has clear sourcing, demonstrable expertise, and easily verified factual accuracy.
Minute 5–7: Run the substitution test
Replace the topic or brand name in three random paragraphs. If the prose still works, it is probably generic.
Typical offenders include:
- “In today’s fast-paced digital landscape…”
- “This powerful tool can revolutionize the way you…”
- “By leveraging these strategies, businesses can unlock…”
- Conclusions that repeat the introduction without making a decision.
Delete these sentences before trying to improve them. Then insert the missing substance: the exact constraint, a concrete example, a threshold, a failure mode, or the reason one option beats another. Our AI slop vs. human content comparison shows why specificity is a stronger signal than decorative imperfection.
Minute 7–10: Perform the evidence-to-paragraph check
For every major section, label the kind of value it contains:
- Observed: something the author saw, tested, measured, or experienced.
- Sourced: a claim grounded in a linked document.
- Reasoned: an explicit conclusion drawn from stated premises.
- Procedural: a step the reader can perform.
If a section contains none of the four, it is atmosphere. Cut it. If the entire piece is “sourced,” it may be a summary that adds no original value. Add analysis, a decision tool, or a procedure that helps the reader use the evidence.
Minute 10–12: Restore decisions and voice
AI drafts often hide behind symmetrical lists: three benefits, three risks, no conclusion. Force the author to decide.
Ask:
- What would we actually do?
- What is the strongest objection?
- When does this advice stop applying?
- Which trade-off is acceptable, and why?
- What detail would only someone close to this work include?
Then read the draft aloud. Break any rhythm that sounds mechanically uniform. Combine choppy repetitions. Split overloaded sentences. Replace abstract nouns with actors and actions. Keep a sharp sentence if it is accurate; “humanizing” does not mean making good prose worse on purpose.
Minute 12–14: Check the who, how, and why
Google recommends making it clear who created content, how it was created when that context matters, and why it exists. Confirm that the byline is real, the publication has a relevant purpose, and any material automation is disclosed in a way that helps the reader rather than performing virtue.
Disclosure is especially useful when automation produced data, translations, summaries, or images that a reader might otherwise misinterpret. The Associated Press’s current generative-AI standards keep human journalists accountable for accuracy and require people to edit and vet experimental AI-assisted outputs. Your publication may use different rules, but ownership cannot be delegated to a tool.
Minute 14–15: Make the publish-or-hold call
Add the weighted score. Record the two lowest categories and one sentence explaining the decision. This gives the next editor something better than “it feels AI-ish.”
Use this handoff:
| Field | Example |
|---|---|
| Score | 76/100 |
| Lowest categories | Original value; claim support |
| Decision | Hold |
| Required repair | Replace the generic trend section with interview notes and link each policy claim to the issuing organization |
| Owner | Named editor |
Before and After: Edit the Thinking, Not Just the Tone
Before:
AI is transforming content creation in many exciting ways. By using the right tools and following best practices, brands can streamline workflows while maintaining quality. However, human oversight remains essential to ensure accuracy and authenticity.
The paragraph is smooth, harmless, and empty. It makes three claims without defining a tool, workflow, quality test, or oversight action.
After:
Use the model for the first structural pass, then make a named editor verify every external claim against its source. If that editor cannot identify one observation, test, or judgment the article contributes, return the draft before line editing. Faster production is not a win when it creates a larger fact-checking queue.
The revision does not merely swap words. It names a use case, an owner, a gate, and a trade-off. That is what the audit is designed to produce.
Primary Guidance Used in This Audit
- Google Search: Creating helpful, reliable, people-first content — self-assessment questions, sourcing, expertise, and the “who, how, and why” framework.
- Google Search: Guidance on using generative AI content — accuracy, quality, relevance, metadata, and giving users context.
- Google Search spam policies — the definition and examples of scaled content abuse.
- Associated Press: Standards around generative AI — human accountability, verification, and limits on publishable material.
The One Rule Worth Remembering
Do not ask whether the draft looks human. Ask whether a human has made—and can defend—every consequential publishing decision.
If the score says hold, hold it. Spend the next hour gathering something real instead of spending it disguising something generic. Then take the repaired piece through the human editorial-pass workflow before it goes live.