Publication readiness needs more than polished prose
Imagine an article that reads smoothly on the first pass. Its opening sounds confident, its headings are tidy and its recommendation appears commercially sensible. A closer review finds that one citation doesn't exist, a broad compliance claim has no qualified source, a client detail remains in an example and nobody has been named to approve publication.
This constructed example reflects recognised risks. NIST's Generative Artificial Intelligence Profile identifies confidently stated false content and misleading or non-existent citations. The Office of the Australian Information Commissioner asks organisations using commercially available AI products to consider personal information, accuracy, tool suitability, transparency and meaningful human oversight in its privacy guidance for AI products.
Fluency doesn't resolve any of those problems. Before publication, the finished output needs a repeatable decision. Pass it, return it for specific repairs, or reject it because the purpose or an unresolved risk makes publication indefensible.
This audit starts after drafting. If recurring failures show that the production method itself is weak, use the AI-assisted content workflow to redesign how the work moves from brief to publication.
Audit the finished output one draft at a time
The unit of review is one publishable draft. That includes the body copy, claims, sources, byline, dates, metadata, links, media and next reader action. A team can later compare records across a batch, but each publication decision belongs to a specific output and an accountable person.
Google's people-first content self-assessment asks practical questions about original information, expert review, factual errors, sourcing, authorship, audience purpose and useful context about automation. These are editorial prompts, not a mechanical ranking score. They help a reviewer inspect what is actually on the page.
A pass means every gate in this audit is met. Revise means a failure can be repaired safely, with an owner and a clear recheck condition. Reject means the page lacks a worthwhile purpose or an important failure cannot be resolved. Don't average the results into a percentage. A serious failure isn't cancelled out by polished prose elsewhere.
Record the draft URL or filename, reviewer, review date, decision and required repair. That small audit trail makes the judgement inspectable and stops an unresolved version from drifting into publication.
Start with purpose and reader value
Compare the draft with its approved brief and the content already on the site. It should serve a named reader at a recognisable decision moment. It should also make a useful contribution rather than repeat a neighbouring article in smoother language.
Google's guidance for using generative AI in content tells publishers to focus on accuracy, quality and relevance, including original and added value. It doesn't say that AI assistance creates a ranking advantage or an automatic penalty. The finished page still has to deserve its place.
Ask what the reader can understand or decide after reading this page that they couldn't get from the existing content estate. If the answer is vague, revision may require a sharper argument or a better example. If the page has no distinct job, reject it rather than polishing a duplicate.
Find the expert contribution
Source summary is useful, but it isn't the business's contribution. The draft should contain judgement someone inside the business can stand behind. That might be a relevant caveat, an interpretation of the evidence, a first-hand delivery pattern or a worked example with enough detail to be useful.
Google's people-first questions cover demonstrable expertise, first-hand knowledge and expert review. For an Off Piste article, the practical standard goes one step further. Evidence should appear before interpretation, and the interpretation should explain what changes for a serious operator.
Mark paragraphs that merely restate a source. Then ask the subject-matter owner to add the decision, boundary or consequence the source alone can't provide. Citation-worthy website content offers a deeper standard for combining evidence, specificity and original interpretation.
If no qualified person will own the recommendation, don't hide that gap behind a confident edit. Return the draft to the relevant expert or narrow its claim.
Verify facts, sources, and claim fit
Open the original source for every material factual claim. Confirm that it exists, then check its date, scope, population, method, qualifiers and exact relationship to the wording in the draft. A real source can still be a poor citation when it supports a narrower or different claim.
This direct check matters because NIST's profile covers both confabulated content and misleading citations. Google's people-first assessment also asks whether content contains easily verified factual errors and whether clear sourcing supports trust.
Fabricated citations, invented quotations, unsupported high-risk claims and material source mismatches are hard stops. Pause publication until the statement is corrected, qualified, supported by an appropriate source or removed. For a detailed record of claims, evidence and ownership, use the claim-level evidence audit. When the source itself needs scrutiny, follow the source evaluation method to test provenance, method, scope and claim fit.
A source check asks whether the link is real. An evidence check asks whether that source earns the sentence the business wants to publish.
Test originality and brand voice
An acceptable draft should sound like a business with a position, not a composite of category pages. Read it aloud. Look for specific nouns, active verbs, evidence before interpretation and a commercial consequence after the explanation. Natural contractions help when they make the sentence sound human. Formal source language should remain intact where precision matters.
Generic phrases often signal missing judgement. Claims such as “transform your business” or “unlock unprecedented growth” need proof and a defined meaning. Forced enthusiasm, inflated certainty and repeated summaries make the page feel less trustworthy, even when the grammar is clean.
Compare the draft with Lara's published work and the approved brief. If the idea is sound but the language is generic, revise it. If the page simply paraphrases existing sources or nearby posts, return to the purpose gate. Originality isn't a cosmetic voice pass.
Check privacy, disclosure, and accountable ownership
Review the prompts, uploaded source material, examples, working draft and final metadata. Look for personal information, confidential commercial material, client details and content placed into a tool that wasn't approved for that use.
The OAIC guidance recommends due diligence before adopting a commercially available AI product and emphasises privacy, accuracy, transparency and meaningful human oversight. The obligations that apply depend on the organisation, activity, information and relevant law. This audit is practical editorial guidance, not legal advice.
Unresolved exposure of personal, confidential or client information is a hard stop. So is the absence of a named person who owns the final publication decision. The practical AI governance policy provides the wider controls for approved tools, data handling, ownership and escalation.
Keep four questions separate. Quality asks whether the content is useful and supportable. Provenance records where media came from and how it changed. Disclosure explains relevant use of automation where context calls for it. Detection attempts to infer whether AI produced material.
The C2PA specifications describe tamper-evident Content Credentials for the provenance of synthetic and modified media. Those credentials don't prove that a claim is true, that a recommendation is useful or that written content meets an editorial standard. They are a bounded media mechanism, not a universal publishing requirement.
Disclosure is contextual. Don't assume every AI-assisted text needs the same label, and don't use a provenance record as a substitute for review. This audit also excludes AI detector scores as a quality proxy. Approve or reject the work using evidence that a reviewer can inspect directly.
Review structure, links, and the next reader action
Headings should advance the argument rather than divide a long list into arbitrary parts. Paragraphs should develop one idea without repeating it. Lists belong where the reader needs a genuine checklist, sequence, taxonomy or parallel decision set.
Check every internal link in context. Its anchor should tell the reader what problem the destination solves. Confirm that external citations lead to the source described, and that the title, author, publish date, update date and description tell the truth about the page.
Then inspect the next action. It should follow from the reader's current problem. A draft about quality control might lead to a deeper evidence check, a governance decision or a publishing-system improvement. It shouldn't jump to a generic sales prompt that ignores the diagnosis.
Once approved content is live, post-publication evidence maintenance keeps claims, links and sources accurate as conditions change.
Agree on the conditions that stop publication regardless of the draft's other strengths.
Make the pass, revise, or reject decision
Bring the findings into one editorial record. For every failed gate, name the criterion, the repair, its owner and what must be rechecked. NIST frames measurement as part of managing generative AI risk, while the OAIC guidance stresses due diligence and human oversight. The Off Piste interpretation is to make the publication judgement explicit without pretending a weighted score can resolve it.
Pass the draft only when it has a valid purpose, a real expert contribution, supported material claims, appropriate information handling, accurate metadata, useful structure and a named approver.
Choose revise when every failure has a safe and specific repair. Give each repair to a person who has the authority and knowledge to complete it. Recheck the affected gate after the change rather than relying on a general proofread.
Reject when the page lacks a worthwhile purpose or when a hard stop remains unresolved. Rejection doesn't always mean deleting the underlying work. Research notes may support a better brief later, but the current output isn't ready to represent the business.
Repeated failures point to a system problem
Start with one high-value draft and make a concrete decision. After several individual reviews, compare the records for recurring patterns.
Repeated invented references suggest weak source controls. Generic language can expose an inadequate brief or missing expert input. Privacy incidents may show that tool rules aren't understood. Drafts with no approver point to an ownership problem rather than a writing problem.
At that stage, another line edit won't fix the system. Rework the AI-assisted content workflow, strengthen governance or improve the content and website publishing setup that carries the work into public. The audit has done its job when the decision is clear, the repair has an owner and the business knows whether the problem sits in this draft or in the way drafts are produced.