THE SHORT ANSWER
Review whether the work is safe and useful to publish—not whether it merely reads well.
Use three distinct passes. First, verify sources, facts, claims, privacy and risk. Second, test whether the content answers the brief, adds genuine value and sounds like the brand. Third, confirm the page, metadata, links and approval record are ready for the intended destination.
Do not begin by polishing sentences. Fluent wording can make unsupported claims harder to notice, and early line editing wastes time if the argument or evidence later needs to change. The release decision belongs to a named person who is qualified for the content's level of consequence.
The purpose of QA is not to prove the model performed well. It is to decide whether the organisation should stand behind the finished work.
BEFORE REVIEW
Match the review path to the consequence of error.
Not every content job needs the same reviewers. A metadata draft based on an approved page carries different consequences from a financial claim, clinical explanation, customer story or public response to a sensitive issue.
Assign the risk tier before generation. Define the required sources, checks and approvers for that tier. If the draft introduces a higher-risk claim or use than the brief allowed, move it to the stronger pathway rather than asking the original reviewer to make an unfamiliar judgement.
- Intended audience, channel and action
- Consequence if a material statement is wrong
- Personal, confidential or restricted information involved
- Legal, compliance, technical or subject expertise required
- Named content owner and final publication authority
- Conditions that require escalation or stopping the job
The Australian Government's current Guidance for AI Adoption groups responsible implementation around accountability, impact, risk management, transparency, testing and human control. The practical implication for marketing teams is simple: quality assurance needs an owner, acceptance criteria and a route to intervene.
PASS ONE / SOURCE TRACEABILITY
Check every material statement against an identifiable source.
Start with facts that could change the reader's decision: product specifications, prices, dates, eligibility, performance statements, market data, regulatory information and quotations. Record where each came from and whether that source is current and approved for public use.
Do not accept a reference merely because it looks plausible or includes a link. Open the source. Confirm that it supports the exact statement, not just the general topic. Check the publication date, scope, geography and whether later guidance has replaced it.
Internal sources need the same discipline. A published page may be outdated. A sales deck may contain indicative wording rather than an approved claim. Where sources conflict, stop and ask the named owner to resolve the difference.
- Every material fact has an identifiable source
- The source supports the exact wording and scope
- Dates, currency, geography and definitions match
- Internal information is current and approved
- Quotations are exact and permission is confirmed
- Contradictory sources are resolved by an accountable owner
PASS ONE / CLAIMS AND LIMITATIONS
Separate evidence, interpretation and recommendation.
A draft can become misleading without containing a completely fabricated fact. It may turn a correlation into causation, present an estimate as an observed result, generalise beyond the sample, omit an important condition or convert a vendor capability into a guaranteed outcome.
Label the type of statement. Facts need evidence. Interpretations need transparent reasoning. Estimates need assumptions and ranges where appropriate. Recommendations need a clear relationship to the audience's objective and constraints.
Review superlatives, comparisons and promises closely. Words such as “best”, “leading”, “always”, “secure”, “compliant” and “guaranteed” carry a larger burden than they appear to. Remove them when the organisation cannot support the exact claim.
- Evidence and interpretation are clearly distinguished
- Causal language is supported
- Estimates disclose their basis
- Limitations and material conditions are visible
- Comparisons use a fair, current basis
- The content does not promise rankings, citations, compliance or results outside the organisation's control
PASS ONE / PRIVACY AND INFORMATION BOUNDARIES
Check what entered the workflow and what appears in the output.
Look for personal information, customer identifiers, confidential commercial material, unpublished product details and content subject to contractual restrictions. Review both the source pack and the generated draft; a summary can still expose information that should not be public.
The New Zealand Office of the Privacy Commissioner advises agencies not to rely on generative AI output without fact checking and to use human review before acting on it. Its guidance also warns against entering personal or confidential information unless retention and disclosure have been explicitly addressed.
Privacy, legal and security owners should determine the organisation's rules. The content reviewer's role is to recognise the boundary, follow it and escalate uncertainty—not to improvise a policy during final approval.
- No unapproved personal or confidential information
- Permissions and intended use are confirmed
- Customer examples cannot enable re-identification
- Images and quotations have appropriate approval
- The output remains within the agreed use case
- Any incident or near-miss is recorded and escalated
PASS TWO / BRIEF FIT
Return to the job the content was meant to do.
Re-read the original brief without looking at the draft. Confirm the audience, situation, decision, channel and intended next action. Then assess whether the content does that job. A polished explanation for the wrong audience is not a successful output.
Check that the main answer appears early enough, the level of detail matches the reader's knowledge and the conclusion resolves the original question. Remove sections that are interesting but do not help the reader progress.
If the brief itself was weak, do not force the draft through review. Send the job back to briefing with the missing decision identified. Quality assurance should reveal upstream problems, not hide them with better editing.
PASS TWO / USEFULNESS AND ORIGINALITY
Ask what the page contributes beyond a competent summary.
Generic AI can create fluent category explanations quickly. The publication standard should be higher: original expertise, a useful framework, first-party evidence, local operating context, a practical decision tool or a clear point of view grounded in experience.
Google's current people-first guidance asks whether content provides original information or analysis, offers substantial value and leaves the reader feeling they have learned enough to achieve their goal. Its guidance for generative search similarly emphasises valuable, non-commodity content rather than pages created for every possible query variation.
Delete filler that repeats the heading, restates widely available advice or stretches the page to a target length. Add the organisation's real contribution where it is missing. If there is no contribution to add, reconsider publication.
- The content answers a real audience need
- The organisation adds expertise, evidence or a practical method
- The argument goes beyond summarising other sources
- New Zealand or Australian context is substantive where relevant
- The reader can make a better decision or take a useful action
- No section exists only to increase length or repeat a keyword
PASS TWO / BRAND QUALITY
Review meaning and behaviour before surface-level tone.
Brand consistency includes more than voice. Check the position the content takes, the customer problem it prioritises, the product names and descriptions it uses, the proof it relies on and the behaviour it asks from the reader.
Then review language: vocabulary, formality, sentence rhythm, point of view, reading level and calls to action. Remove generic AI habits such as inflated openings, repeated summaries, false certainty and phrases the brand would not normally use.
Do not edit everything towards one uniform voice. A product page, technical guide and social post can express the same brand differently. Judge whether the variation fits the channel while preserving the organisation's non-negotiable decisions.
- Positioning and product language are current
- Tone fits both brand and channel
- Prohibited or inflated language is removed
- The reader is addressed respectfully and clearly
- Calls to action match the commercial journey
- The content sounds recognisably from this organisation
PASS TWO / SENSITIVITY AND ACCESSIBILITY
Check who could be excluded, misrepresented or harmed.
Review examples, assumptions, imagery and language for the people affected by the content. Look for stereotypes, unnecessary demographic assumptions, inaccessible explanations and advice that may not hold across the intended audience.
For Aotearoa New Zealand audiences, check spelling, terminology, pronunciation marks and cultural context with people who hold the relevant knowledge. Do not use translation or cultural references as decorative signals. For Australian audiences, confirm that local terminology, currency and regulatory context have not been copied across without review.
Use readable headings, short paragraphs, descriptive links and meaningful alternative text where images convey information. Accessibility is part of publication quality, not a technical afterthought.
PASS THREE / SEARCH AND AI DISCOVERY
Make the useful answer easy to find, understand and connect.
Confirm that the title and main heading describe the page accurately, the opening answers the primary question, headings reflect the real structure and internal links help readers move to related explanations or commercial next steps.
Check the meta description, canonical URL, indexation intent, sitemap inclusion and relevant structured data. Structured data should describe visible content accurately; it does not compensate for weak information or guarantee a search feature, ranking or AI citation.
Make sure important text is available in rendered HTML and not hidden only inside an image or inaccessible interaction. The page should work for people first and remain clear to systems that crawl and retrieve it.
- Descriptive page title and one clear H1
- Useful answer and subject are clear near the start
- Logical headings and crawlable internal links
- Accurate meta description and canonical URL
- Correct indexation, sitemap and structured data
- Visible dates and ownership where currency matters
- No unsupported SEO or AI-discovery claims
PASS THREE / DESTINATION READINESS
Review the content in the place where people will use it.
Copy can change when it enters a page template, email platform, social format or campaign build. Headings wrap. Disclosures become separated from claims. Links break. Tables become unusable on mobile. Complete the final review in context rather than approving a document and assuming the implementation will preserve it.
Test every link and action. Confirm the owner, publication date, images, captions, tracking requirements and legal text. Preview common screen sizes and check that the most important answer and action remain clear.
If content is repurposed across channels, each version needs its own destination check. Approval of a long-form source does not automatically approve every shortened variation.
THE RELEASE DECISION
Record who approved what—and what happens next.
The final approver should make one of three decisions: publish, return with specific changes, or stop and escalate. Avoid vague approval such as “looks fine” when evidence, risk or ownership remains unresolved.
Keep a light release record: content identifier, version, source set, required reviewers, approval date, destination and next review date. Record material corrections after publication and use recurring patterns to improve shared sources, rules, briefs or training.
Scribe is designed to apply approved brand knowledge, source material and repeatable content rules before the draft reaches a reviewer. That can reduce predictable correction, but it does not remove accountable human approval. A working demonstration is most useful when you bring one real draft, its source pack and the release standard it must meet.
- Evidence and claims verified
- Privacy and information boundaries satisfied
- Brief, audience and commercial role fulfilled
- Original contribution and brand quality present
- Specialist reviews completed
- Destination and technical checks passed
- Named person authorises publication
- Next review date and learning captured
