I have written about my fact-checking routine before. This post shows it applied to one actual draft, from start to finish. What arrived, what I flagged, what I rejected, and in what order.
What the draft arrived with
The draft was about publishing habits. It came in at about 900 words with three statistics, one named study, and a claim about what “most blogs” do. It read smoothly. That smoothness is exactly the problem. A language model predicts likely text, and likely text sounds confident whether or not it is true.
The first pass
What I would do first is read the whole thing once without editing anything. I am not looking for style. I am looking for every claim a reader might act on. I mark them as I go: the three numbers, the study, and the “most blogs” claim. Five claims total. That is a normal count for a draft this length.
Sorting the claims
Not every claim costs the same to check. The “most blogs” claim is vague enough that I cut it rather than verify it. It adds nothing a reader can use. The three numbers are worth chasing, because a wrong number in a post on my site is a real problem, not a cosmetic one. The named study I treat as unconfirmed until I can find it myself.
Chasing the sources
Two of the three numbers traced back to something real, though one had been rounded in a way that changed its meaning. I replaced it with the original figure and a link to the source. The third number I could not trace at all. It had the shape of a real statistic, which is what makes these hard to spot. I re-derived that point from the source I did have, or I would have cut it.
The named study did not exist as described. The draft had the right idea attached to the wrong citation. I rejected that section outright. Rejecting costs me a paragraph, and rewriting it from a real source costs maybe twenty minutes. Publishing it wrong would cost more.
What I kept
After the pass, the draft kept its structure and most of its sentences. What changed was the load-bearing content: the numbers now point somewhere, the fake study is gone, and the vague claim is cut. The value of the routine is not that it catches everything. It is that I run it every time, in the same order, so nothing skips past me.
The lesson from this one
Most blogs stop at reading the draft and fixing the tone. That is the easy part. The part that matters is treating every claim as unconfirmed until you have seen the original source yourself. The AI draft is a starting point. The accountability for what gets published stays with me.