The most expensive mistake I have seen with AI-assisted content was a single statistic.
It was specific, plausible, attributed to a research firm that exists, and completely invented. It went out in a client newsletter to several thousand people, and one of them was a former analyst at that firm.
Everything after that was recoverable but unpleasant.
So this is the process I have used since, which takes about ten minutes per article and catches four distinct categories of problem.
Do it in this order, because it is ordered by how badly each failure hurts.
Quick Info
One: Every Number, Every Name, Every Citation
Do this first and do it without exception.
Fabricated citations remain a normal failure mode, and the fabrications are convincing. Real journal names, real-sounding studies, plausible percentages, sometimes real authors attached to papers they never wrote.
The rule I work to:
No number goes in without a link to its primary source. No exceptions, no matter how good the number sounds.
In practice: search the claim, find the original source rather than an article citing it, confirm the figure and the date, and link to it. If you cannot find the primary source in two minutes, delete the claim.
I lose two or three good statistics per article this way. I have never lost a client over an accurate one.
Two: Check the Dates on Everything
The second most common problem and the least discussed.
Models draw on material of varying age, so content confidently describes a tool's pricing, a platform's features, a regulation or a best practice as current when it has changed.
What to check specifically:
This category is why I distrust AI-assisted content about software more than about principles. Principles age slowly. Pricing pages do not.
Three: Find the Confident Nonsense
The third category is claims that are not verifiably false but are stated with more certainty than anyone could have.
Phrases like studies show, experts agree, research consistently demonstrates, with no study, expert or research named. These read as authoritative and they are placeholders where evidence should be.
Search the draft for those constructions specifically. Then either name the source or soften the claim to what you can actually support, which is often your own observation.
Saying in my experience is honest and it is also more interesting than pretending to a consensus.
Four: The Voice and Substance Pass
Now the editorial work, which is what separates useful content from filler.
Five questions per draft.
Then read it aloud. It is the fastest way to find the sentences that are technically fine and unmistakably not yours.
The Words That Give It Away
A shortlist worth searching for, because they cluster in generated prose and they signal it clearly to a regular reader.
Delve, tapestry, landscape used metaphorically, navigate used metaphorically, unlock, elevate, seamless, robust, leverage as a verb, in today's fast-paced world, it is important to note, when it comes to, at the end of the day.
Also structural tells: every paragraph the same length, every list exactly three items, an opening that restates the title, and a closing paragraph that summarises what you just read.
None of these are forbidden words. They are just the most predictable available choice each time, and predictability is the actual problem.
Who Signs It
The last step is not a check, it is a decision.
Somebody's name goes on the piece, and that person is accountable for every claim in it. If the named author has not read every word, do not publish it under their name.
This is not only an integrity point. Accountable authorship with real relevant experience is part of what makes content eligible to be trusted and cited at all, which I go into in my E-E-A-T article.
So the process ends with a person reading it properly. There is no version of this that works without that step, and it is the step under most pressure when volume goes up.
A Realistic Time Budget
For a fifteen hundred word article: two minutes on numbers and citations, two on dates, one on unsupported certainty, five on voice and substance.
Ten minutes. Against an article that might take an hour to produce, that is a reasonable insurance premium, and it is the difference between publishing something and publishing something you can stand behind.
If you cannot spare ten minutes per article, publish fewer articles. That is not a slogan, it is the actual recommendation.
Frequently Asked Questions
Yes. Fabricated but plausible sources remain a normal failure mode, including real publication names attached to studies that do not exist. Verify every one.
Search for the claim, find the primary source rather than an article citing it, confirm the figure and date, and link to it. If you cannot find it in two minutes, remove the claim.
Because pricing, features, policies and regulations change faster than the material models were trained on. Anything time-sensitive needs checking against the current source.
Add something that could only come from your own experience, state an opinion, cut a quarter of the words, and read it aloud.
Yes. Accountability is the point, and accountable authorship with relevant experience is also part of what makes content trustworthy enough to be cited.
About ten minutes for a fifteen-hundred-word piece. If you cannot spare that per article, publish fewer articles.
Before You Go
Numbers, dates, unsupported certainty, then voice. Ten minutes, in that order, every time.
My editorial checklist, including the word list to search for, is in the resources section.
If you want an editorial process built for a team publishing at volume, get in touch.
No number without a link, Tariq
I'm a marketing consultant, entrepreneur and content creator. I help businesses grow through practical marketing, websites, SEO, content and AI.
More About Tariq →