The review checklist we would put in front of every AI-assisted draft
Twelve items, in the order a careful editor meets them. Print it, adapt it, or paste it into your content tracker.
How to use the list
Do the items in order. Skipping ahead is how errors slip through, and most teams find that half of the checks take seconds while two of them eat the afternoon.
Fact verification and the rewrite for a real point of view are the slow ones. That is normal, and it is exactly where a human reviewer earns their place, so do not try to automate them away. Not worth it.
The source step uses a domain rating checker free of charge for single lookups, and the bulk call when a draft carries more than a handful of links. As of October 2026 the starter credit is $100, which equals 1,000,000 credits. Plenty.
Before the draft is allowed near the CMS
- A named human owns the piece, and their name is in the tracker
- The brief says who the page is for and what they should do afterwards
- Every factual claim has a source the editor has opened, not just seen listed
- Every cited domain was scored; zero-score domains were opened and read
- Quotes appear on the page they are credited to
- Dates, figures and product names were checked against the source
- Outbound links go to pages you would be happy to defend
- Nothing in the text claims experience the author does not have
- The piece contains at least one thing the model could not know: your data, your process, your photographs
- Repeated phrasing and filler openers were cut
- Any required disclosure or label follows your own policy
- The reviewer's note and date are stored with the draft
Who signs off
The editor who ran the checks, not the person who wrote the prompt
Separating the prompter from the approver is cheap and catches a lot. If your team is one person, wait a day and re-read cold before you publish.
- Small teams: one reviewer, a day between draft and approval
- Larger teams: reviewer plus a spot audit of 10 percent of the month's output
Shortcuts we see teams take
Every team cuts corners eventually. The question is which ones hurt. Skipping the source check hurts most, because a wrong citation is public and easy for a reader to spot, while a dull intro costs only a little.
Skipping the owner step hurts second. A piece with no named editor has nobody who feels responsible when a reader writes in with a correction, and the error then stays on the page for months. Nobody owns it.
Cutting the original-contribution step is the slow poison. A site made only of rearranged model output gives readers little reason to return, and in Google's terms it starts to look like content created mainly to fill a keyword list. Avoid that.
What to do when an item fails
Fixing beats arguing. If a claim has no source, cut it or find one. Then move on to the next item without ceremony, because a checklist that invites a debate on every line stops being used within a month.
If the same failure keeps appearing, change the prompt or the brief and not just the output, because repeated unsourced claims usually come from a vague instruction asking the model to sound authoritative. Fix the cause. Not the symptom.
For source problems specifically, our troubleshooting table lists the common causes and the fix for each one, in the order that editors tend to meet them. For cost planning of the automated parts, see the pipeline cost table. Prices are in credits.
Quick questions about the list
Do we need all twelve items for a short news brief?
Not always. For a 300-word update, the owner, the source check, the quote check and the dates matter most. The original-contribution step can be lighter, but never skip the person who signs it off, even for short pieces.
Can the checks run automatically?
Some can. The domain scoring runs well from a script, and a link-opening check is simple too. Claims, tone and originality still need a person, so automate the list-making and keep a human on the judgment calls.
How long should a review take?
It varies with length and subject. A reasonable aim is to spend more time reviewing a draft than the model spent writing it, which is not hard, and to avoid rubber-stamping.