What should a human approve before AI-drafted content is published?

Before AI-drafted content is published, a named person should approve its claims, page elements, purpose, authorship and disclosure, and approval record. Each claim should trace to a source the approver opened, and the sign-off should be dated and stored, so a later measurement knows what changed and when.

• Google calls it critical to manually fact-check and review all AI-generated content before publishing, and does not say who should do it (Google, Search Central, read 30 October 2026).
• Penlid's method holds a draft for human review when any claim in it is contradicted, whatever its overall score (Penlid, Methodology, read 30 October 2026).
• In Penlid's simulated-site tests, pages picked because they had been declining, then left unchanged, were falsely called "improved" 5.50% to 9.20% of the time (Penlid, Methodology, read 30 October 2026).

Why does AI-drafted content need an approval step?

AI-drafted content needs an approval step because a generative model predicts likely wording instead of looking facts up, so a fluent draft can still be wrong. Google's guidance describes such errors as hallucinations. It also names title elements, meta description elements, structured data and image alt text, and tells publishers to validate structured data markup, so the checklist below covers them.

The step is about what the page says, not about hiding AI use. Google's February 2023 Search Central Blog post states that Appropriate use of AI or automation is not against our guidelines. The same answer adds that generating content primarily to manipulate search rankings is against its spam policies; what Google's policies say about AI-generated content has the detail. NIST's voluntary generative AI profile warns that people may defer too much to generative AI, so we define the checks below instead of trusting a skim.

What should a human check before approving an AI-drafted page?

A human should check five things before approving an AI-drafted page: its claims, its page elements, its purpose, its authorship and disclosure, and its approval record. None of the sources listed below prescribes an approval workflow, so this checklist is our own design; the table gives the reason for each check.

In our reading, claims come first, because a corrected claim can change the title, the purpose test and the record. A note about AI use has its own guide: how to disclose AI assistance on a page.

What should stop an AI-drafted page from being approved?

One claim that its source contradicts should stop an AI-drafted page from being approved, however good the rest of the draft is. Penlid's methodology describes checking each claim in a draft against live sources, and it counts a source about a related but different statistic as no support: A source discussing an adjacent statistic is recorded as unverifiable, not supported. It also lets one failure hold a draft: A draft with any contradicted claim is held for human review regardless of its overall score.

Our approval rule borrows that logic: a high overall score cannot outvote one contradicted claim, and a claim that no source states must be sourced or cut before sign-off. The checking steps are in how to fact-check AI-written content against primary sources. Two more findings should hold a page; these are our rules, not a source's:

• A purpose that fails. Google's helpful-content page asks whether you are mainly summarizing what others have to say without adding much value, and its spam policies list using generative AI tools to generate many pages without adding value for users as an example of scaled content abuse.
• A fabricated byline. Google describes fabricating creator profiles, such as with AI-generated headshots, made-up names or false credentials, as a form of deception.

Anthropic's documentation says its techniques, such as asking a model to find a supporting quote for each claim, reduce hallucinations but do not eliminate them, and says to always validate critical information, especially for high-stakes decisions. That last check is the approver's job.

Who should approve, and what do they need in front of them?

The approver should be a named person who knows the topic well enough to spot a wrong claim, who sees each source beside the draft and who can hold the page. That a person approves is our rule: Google's guidance says to manually fact-check and review but does not say who does it, although its helpful-content questions ask whether someone with demonstrable knowledge of the topic wrote or reviewed the content.

OpenAI's safety guide, written for applications built on its API, advises that wherever possible a human review outputs before use, calls this especially critical in high-stakes domains and says reviewers should have the information needed to verify outputs. We apply that to publishing by analogy. Three rules follow; they are our reasoning, not a source's:

• Put each source beside its claim, so the check is a comparison and not a search.
• Add a second reviewer for high-stakes topics. Google's helpful-content page says its systems give even more weight to strong E-E-A-T (experience, expertise, authoritativeness and trustworthiness) for topics that could significantly impact health, financial stability, safety or well-being, so an error there costs more.
• Cap the queue at what the approvers can read in full. Google's helpful-content page says its quality raters are trained to treat generative AI used to produce large amounts of text without manual oversight or curation as little to no effort; the example below shows the arithmetic, and what counts as scaled content abuse covers the policy.

What should the approval record hold so a later measurement can use it?

The approval record should hold the version approved, the approver, the date and how the page was chosen, because a later measurement can only use changes someone wrote down. Penlid's methodology states the same limit for its own measurements: We only know about changes we were told about.

• The page address, the version approved, the approval date written YYYY-MM-DD, the approver's name, the checks passed and every claim changed during review.
• The tool and model version that drafted the page, and the date each source was read.
• How the page was chosen for AI drafting, and which comparable pages you left alone.

The last item matters most for measurement. Penlid's methodology says that when a result matters, choose which pages to change without looking at how they were trending. In its simulated-site tests, pages whose last 28 days sat at least one standard deviation below their own average, then changed in no way at all, still came back "improved" 5.50% to 9.20% of the time, against at most 3.10% when pages were not picked that way; it names regression to the mean as the reason. What a content operations ledger should record lists the fields, and how to choose pages to change without bias covers the selection.