When to run this checklist
Run it on every piece where an AI tool wrote a meaningful share of the text or produced an image, at the point where a person reviews the draft and before it goes to publishing or to the send queue. That includes ad variations generated in bulk inside an ad platform: the review can be lighter for a short headline, but a headline can still carry a false claim or a price that changed last week. An AI answer used for research or planning rather than published needs its own check before it travels, set out in the step-by-step routine for fact-checking an AI answer.
The reviewer should be someone who knows the subject well enough to spot a plausible error, not only someone who can spot a typo. For regulated topics, such as health, finance or legal matters, the review also needs whoever normally signs off claims in that area. The accuracy and review workflows guide explains how to match the depth of review to the risk of the piece.
Before you start
- The brief, with the target query or audience, the purpose and any required claims or offers.
- The promptThe instruction and material given to an AI language model, which works like a brief: the model knows language, not the audience or product, unless told. and the tool used, so you know what the model was given and what it could not know.
- The primary sources for facts the piece relies on: product pages, price lists, research, the people being quoted.
- The brand voice guide and the business's policy on disclosing AI-assisted content.
- Comparable human-written pieces, so later performance can be judged against something.
The checklist
Work through the groups in order. A piece that fails accuracy goes back before anyone polishes its voice, and a piece with a privacy problem stops even if everything else is perfect. Ticks are saved in your browser only, so record the sign-off itself in the content tracker.
Accuracy
Quality and originality
Privacy and disclosure
Rights and sources
Sign-off
Where each check is explained
The widget keeps each line short enough to tick; the reasoning sits in the guides. Each check links to the page that explains it.
| Check | Explained in |
|---|---|
| Facts, statistics and quotes against primary sources | Accuracy and review workflows |
| Unsubstantiated claims removed | Approvals and compliance reviews |
| Product names, prices and terms current | Accuracy and review workflows |
| First-hand experience, examples or data added | AI-assisted content with human review |
| Generic filler and repetition cut | Writing for the web |
| Search intent answered | Search intent |
| Voice checked against brand guidelines | Brand identity and voice |
| No customer or confidential data in the tool | Disclosure and data privacy |
| Disclosure policy applied | Disclosure and data privacy |
| Generated images checked for rights and likeness | AI for creative |
| Sources credited and linked | E-E-A-T and content quality |
| Named reviewer and recorded approval | Accuracy and review workflows |
| Performance tracked against human-written content | AI-written content not performing |
If an item fails
An accuracy or privacy failure stops the piece. A fact that cannot be traced to a primary source is removed or replaced with one that can; it is not softened with "around" or "reportedly". A claim the business cannot substantiate comes out, whatever it does for the copy. If customer records, unreleased plans or anything covered by a confidentiality agreement went into a tool whose terms allow the vendor to keep or train on it, the piece is not the main problem any more: report it under the business's data policy, then decide whether the draft can be used at all. Customer records and other First-Party DataData a business collects directly from its own customers, visitors and users. are what this rule exists to protect.
Originality failures send the draft back to the writer, not to the model. Asking the tool to "make it more original" produces different generic text. The fix is something only the business has: a customer story, a result from its own data, a photograph from its own work, the opinion of someone who does the job. Google's guidance in the sources below is clear that search engines judge the result, not the tool, and the E-E-A-T guide shows how E-E-A-TExperience, Expertise, Authoritativeness, Trust. signals come from exactly this kind of experience.
Voice, source credits and image checks are usually quick fixes. The sign-off group should never be skipped: a piece with no named reviewer is a piece nobody owns when a reader finds the error.
A kombucha brewery's blog post on gut health
A kombucha brewery uses an AI tool to draft a blog post on fermentation and gut health. The reviewer, the head brewer, checks it against the list. The draft cites a study with a precise figure, 37% of adults, that no source supports; it is cut. A sentence saying the drink "supports immunity" is a health claim the brewery cannot substantiate, so it is removed. The draft lists last year's price for the mixed case. The brewer adds a paragraph on how the brewery's own second fermentation works and two photographs from the brewhouse, applies the brewery's disclosure line for AI-assisted posts, and records the approval in the tracker. Three months later the post is compared with the brewery's two human-written posts on fermentation.
Common mistakes
- Reviewing for style, not substance. The reviewer fixes the tone and the commas, and the invented statistic sails through because it was well written.
- Checking facts against another AI answer. A second model agreeing with the first is not a source.
- Leaving the model's structure in place. Every section has the same shape and the same summary sentence, and readers notice the pattern even when they cannot name it.
- Pasting customer data to "personalise" a draft. The convenience is real and the data exposure is permanent.
- Disclosing inconsistently. One post carries an AI disclosure and the next, made the same way, does not, which looks worse than either policy applied evenly.
- Never comparing results. Without the comparison against human-written content, nobody can tell whether AI-assisted production is saving time or costing traffic.
The checklists library holds the other pre-flight lists; for web pages, run the SEO page publishing checklist after this review.
