AI-assisted content QA: what to approve before publishing

A draft can read well and still be unfit to publish. An unsourced figure, an outdated comparison or an explanation that repeats another page remains a problem even when the prose is clean. Editorial quality assurance should end with a decision and a named person accountable for it.

This process is for content leads who use AI for research, drafting or translation. It produces an approval record for each article, not an authorship verdict or a ranking score. The guide to writing content AI can cite covers how to present an answer; this guide checks whether that answer has enough support to publish.

Separate Google's policy from your editorial standard

Google's guidance on generative AI content explains that automation can help with research and structure. It also calls for accuracy and relevance, including in metadata and structured data.

Its scaled content abuse policy addresses large numbers of pages created primarily to manipulate results rather than help users, regardless of the production method. It provides no safe publishing quota and does not classify every use of AI assistance as spam.

The checks below are Mentio's proposed editorial process. They are not an official Google test, and passing them does not guarantee rankings, citations or error-free content. Their purpose is to record what was checked before publication.

Open a review record before editing sentences

Record the proposed URL, the reader, their problem and the decision they should be able to make after reading. Add the closest existing article and describe the difference in a sentence that does not rely on swapping a keyword.

Name the draft's author, the fact reviewer and the person who authorizes publication. If one person holds several roles, record that rather than implying an independent review. Keep the exact version examined and note where AI was used. Retain enough information to reconstruct the review without adding private conversations or customer data.

Before continuing, check whether an existing page already solves the same problem. If an update would suffice, return the proposal to the editorial calendar owner. A publishing schedule does not justify two URLs offering the same answer.

Check claims that affect a decision

Start with prices, figures, product capabilities, comparisons, dates and cause-and-effect claims. For each material claim, retain the exact passage and the evidence you actually read. A URL in a bibliography does not establish that the source supports the article's wording.

If a source reports correlation, do not approve a causal promise. If a feature is available only on one plan, keep that qualification in the article. When sources disagree, explain the disagreement or narrow the claim rather than silently choosing the convenient source.

For proprietary data, link the method and keep the sample, period and exclusions visible. The proprietary research guide covers that work. Do not claim the team ran tests it never performed: an explanation based on documentation is not first-hand product experience.

Compare the contribution with the closest existing page

Compare the promise in the title, the H2 headings and the task the reader completes. Then read both bodies. Different headings can conceal the same answer.

Ask the reviewer to identify what the draft helps readers do that the existing page does not. That might be a verification procedure, a relevant limitation or a decision supported by checked sources. Not every article needs an original study, but a longer paraphrase does not create value by itself.

Also examine variations across a series. If only the industry, city or tool name changes while the advice stays the same, check whether each variation has specific evidence behind it. Where that evidence is missing, stop extending the series and decide which URL should hold the answer. An automatic similarity percentage cannot make that decision for you.

Test usefulness before polishing the prose

Walk through the article's procedure using the materials it requires. Can you identify the inputs, perform each step and recognize the result? Flag any gap that depends on an unmentioned tool, permission or data source.

Then review the language. Remove interchangeable introductions, conclusions that merely repeat the previous section and unsupported urgency. Preserve qualifications, exceptions and longer explanations where they are needed. The aim is clearer understanding, not identical sentence lengths throughout the blog.

Review the translation separately. Check negations, units, feature names, markets and levels of certainty. A Spanish "podría" must not become an English "will". If the approved original changes, review the equivalent passage and metadata in the other language again.

Decide what blocks publication and what needs revision

Do not average the problems into a passing score. A single unsupported central claim can invalidate the conclusion even when the rest of the text is sound.

These are proposed approval rules, not thresholds published by a search engine. Give each issue a location, an owner and evidence of resolution. The final status should be ready to publish, needs revisions or requires more research. A label saying only "reviewed" leaves the decision unclear.

Keep an approval record you can reconstruct

Use the following as a working record:

• URL, language and version examined.
• Reader, problem and intended outcome.
• Neighboring URL and the draft's specific contribution.
• Material claims, sources checked and unresolved limitations.
• Procedure check and translation review results.
• Issues, corrections and evidence of resolution.
• Approver, decision and date.
• Trigger for the next review: a change in product, source, data or procedure.

Before closing the record, open the published page. Check that its title, description, FAQ and structured data match the approved content. Check the images, links and language too. Retain the final version so a later change can be distinguished from an error in the original review.

Across a series, record recurring problems and where they originate. If the same brief produces unsourced claims in several articles, correct the brief and review the affected batch. The measurement governance guide helps document definitions when an article uses visibility metrics; editorial approval of those claims remains a separate task.

Frequently asked questions

Can an AI detector approve an article?

Do not use it as an approval criterion. Its output does not verify sources, usefulness or accuracy. Base the decision on the content examined and the issues resolved, without inferring authorship from stylistic signals.

How many AI-assisted articles can I safely publish?

The cited Google policy establishes no safe number. Review the purpose, value to readers and your actual capacity to verify each publication. A quota cannot replace those checks.

Does every post need proprietary research?

No. A well-sourced synthesis can help readers complete a useful task. It should distinguish sourced information, inferences and recommendations without claiming experience or data that do not exist.

What if I cannot confirm a central claim?

Keep the draft on hold until you verify the claim, remove it or revise the conclusion to a defensible scope. Repeating a question to a model does not turn missing evidence into certainty.

Does passing QA guarantee inclusion in AI answers?

No. This QA process documents editorial review; it does not control which sources a system selects. Measure visibility separately, and do not treat a mention as validation of every claim in the article.

Want to know if AI mentions your brand?

Discover your visibility in ChatGPT, Claude and Gemini in minutes.

Analyze your brand for free

Related articles