How to fact-check an AI-written article before you publish: a twenty-minute routine
By Thomas Lovas Lokoy · Published 16 October 2026 · 5 min read
A draft from a language model reads with the same confidence whether it is right or wrong. That is the whole problem, and it is why a review routine has to be mechanical rather than impressionistic: you do not read the article and ask whether it feels right, you work through a fixed list and ask of each item whether you can show where it came from. This is the routine, in the order that catches the most expensive mistakes first.
Step one: open every source (five minutes)
Start at the bottom, with the Sources section. In AI Orbit the writer is instructed to end every article with a Sources section listing only pages it actually read, and to omit the section if it read none; the article page also shows "Pages read while researching", the list of pages the research step opened. Open each link.
For each one you are checking three things: that the page exists, that it says what the article attributes to it, and that it is the kind of page you would be comfortable being seen to rely on. A source that exists but does not support the sentence citing it is the most common failure, and the sentence has to change, not the citation. If an article has no Sources section at all and makes factual claims, treat every claim as unverified and go to step two with that in mind.
Step two: every number (five minutes)
Read the body once looking only for digits and the words that stand in for them: per cent, half, most, double, a third. Each one gets one of three outcomes:
- Found in a source you opened in step one. Keep it, and make sure the sentence says where it came from.
- Found by you now in a primary source, the vendor's own page, a regulator, the original study. Keep it and add the source.
- Not found. Cut the number. Rewrite the sentence around what is known, or delete the sentence.
Do not soften an unverified figure into "around" or "roughly". A hedged invention is still an invention. The writer's brief in AI Orbit tells the model that an honest "it depends, and here is what it depends on" is worth more than a confident invention; hold the draft to the same standard.
Prices deserve a special rule. They change, and a draft that was right on the day it was written is wrong a month later. Unless you have just read the figure on the provider's own page, replace it with an instruction to check the current price, and date any figure you keep.
Step three: names, dates and quotations (three minutes)
Named people, named companies, named products, named laws and named studies are the next most expensive class of error, because a wrong one is defamatory or embarrassing rather than merely inaccurate. Check that each person exists and holds the role the article gives them, that each product still exists under that name, and that each law or rule is current. Any direct quotation you cannot find verbatim in a source goes; paraphrase it as a claim with attribution or remove it. A date that cannot be verified becomes a season or a year, or disappears.
Step four: the metadata (two minutes)
The title tag, meta description, excerpt and tags are content too, and Google's own guidance on generative AI content says the manual review should extend to titles, meta descriptions, structured data and image alt text. In AI Orbit these are generated at the end of the run and stored with the article; read them. The meta description in particular is a promise about what the page contains, written to be clicked. Check that the page keeps the promise. A description that claims a comparison the body does not make is a small lie at the top of the search results.
The cover image needs one glance: does the attribution line make sense, and does the licence linked from it suit where you are publishing?
Step five: read it once as the reader (five minutes)
Only now read the article start to finish, as the person it was written for. You are no longer checking facts; you are checking whether it answers the question that person had, whether the opening gets to the point, whether any section exists only to reach a word count, and whether the close tells them what to do next. The writer in AI Orbit is told to cut every sentence that only introduces the next one and never to pad by restating a point. Where it has done so anyway, cut.
This is also where you catch the things no checklist can: the paragraph that is technically true and completely beside the point, the example that would land in one country and baffle another, the tone that is wrong for the destination.
What to do with what you cannot verify
The rule is simple and uncomfortable: if you cannot verify it and you cannot rewrite around it, it does not go out. The cost of cutting a paragraph is a shorter article. The cost of publishing a wrong one is a correction, a reader who trusts you less, and in some subjects, finance, health, law, real harm to someone who acted on it. In those subjects, add the risk warning or the professional-advice line the subject requires, and consider whether a draft is the right starting point at all.
Making the routine faster over time
Three habits shorten it. Brief tighter: a topic sentence that names the country and the reader produces fewer vague claims to check, as the briefing guide explains. Keep research on where your plan has it, because a draft written from a brief of real pages has sources to check rather than claims to hunt. And keep a short list of the primary sources for your subject, so step two becomes a lookup rather than a search.
What the routine does not do is make the draft yours by virtue of having been checked. You still have to decide whether you agree with it. Read how the draft is produced so you know what the model was told, then run the routine on the Free tier before you pay for anything; the pricing page has the plans.