Every AI-assisted draft fails in the same handful of ways. The prose is fluent, the structure is reasonable, and then somewhere in the third section there is a statistic attributed to nobody, a claim that sounds authoritative and cannot be checked, and a paragraph that repeats the introduction.
Because the failures are consistent, they can be caught with a checklist rather than a full re-read. The five checks below take about ten minutes per piece, and each one targets a specific way the text breaks.
What actually goes wrong
It helps to know why these failures happen, because the cause determines the fix. A model asked to write about a topic without supplied specifics will produce the average of what it has encountered. That average includes real facts, plausible-sounding near-facts, and rhetorical structures that imitate the shape of evidence without containing any.
The result is a piece that reads as informative while carrying much less information than its length implies. Reviewing it is not about finding errors of grammar; it is about finding sentences that look like content and are not.
Check one: invented specifics
Scan for every number, name, date, and attribution. Then sort them into two piles: things you can point at a source for, and everything else.
Pay particular attention to percentages, survey findings, and named research. These are the easiest things for a model to produce convincingly and the most damaging when they are wrong, because readers who notice a fabricated statistic stop trusting the rest of the page.
The fix is rarely to hunt for a real number that supports the claim. More often the sentence is stronger without it: replace "studies show a majority of teams report X" with a description of what actually happens, or with a plain statement of your own observation. A claim with no number is unremarkable. A claim with a fake number is a liability.
Check two: claims nobody can check
Different from invented specifics, these are statements that are technically unfalsifiable — "this approach delivers better results", "teams see significant improvements", "the most effective way to do this is…".
The test is to ask what evidence would prove the sentence wrong. If nothing could, the sentence carries no information, regardless of how confident it sounds.
Two ways to repair one. Add the condition: instead of "this works well", write "this works when the source material is already structured; it fails when you are starting from a rough transcript". Or replace the evaluation with the mechanism: rather than asserting that something is better, explain what it does and let the reader judge.
Search for the words "significant", "effective", "best", "powerful", and "robust". They mark the sentences most likely to need this treatment.
Check three: structure
Read only the headings, in order, and ask whether each one advances the piece. Generated structures often contain a section that exists because the format expects it — a "benefits" section, a "challenges" section, a "conclusion" that restates the opening.
Remove any heading whose content a reader could predict before reading it. If the reader already knows what "Challenges" will say, the section adds length and no value, and it costs you the reader's patience for the sections that do matter.
Then check the ending. A conclusion should either answer the question the piece opened with or tell the reader what to do with what they have learned. A restatement of the introduction is neither, and it is the most common structural defect in generated drafts.
Check four: originality
Not plagiarism — that is a separate concern — but whether the piece contains anything a competitor's page on the same subject would not.
Run through the draft and mark every passage that could only have come from your specific experience: a detail about how the work is actually done, a constraint you know about, a failure you have seen. If the count is zero, the piece is a competent summary of the public conversation and nothing more, and it will not earn links, returns, or rankings against pages that already hold that ground.
When the count is low, the fix is not to rewrite the prose. It is to add the missing material and then cut an equivalent amount of generic text, so the piece does not simply grow longer.
Check five: reader trust
The final pass is about whether the piece is honest with the person reading it. Three things to look for.
- Does the headline promise what the article delivers? A title that overpromises makes an otherwise solid piece feel like a disappointment, and the reader's sense of having been misled attaches to the whole site.
- Is anything overstated? Confidence that outruns the evidence reads as sales copy. Readers forgive uncertainty far more readily than they forgive exaggeration.
- Is the AI use disclosed if the reader would reasonably expect it? Where a piece presents personal experience it does not have, or speaks in a voice that is not real, stating that plainly is part of the content being trustworthy.
Running the checklist in ten minutes
Do not attempt a full read for each check. Each one is a scan for a specific pattern, and they are much faster when done separately.
- Numbers and attributions: three minutes, using search.
- Unverifiable claims: two minutes, searching five words.
- Headings and the ending: one minute.
- Originality count: two minutes.
- Trust pass: two minutes, reading the headline and the first and last sections only.
Ten minutes against a piece that took ninety to produce is a cheap trade, and it is the difference between a site that accumulates an editorial standard and one that only claims to have one.
What to do when a piece fails
Failures cluster. A piece that fails the originality check usually fails the unverifiable-claims check too, because both are symptoms of the same cause: the draft was written without supplied substance.
When two or more checks fail, do not patch. Return to the input and rebuild the middle with real material. Patching a hollow draft produces a longer hollow draft, and the additional work is spent making the emptiness less visible rather than filling it.
One further rule is worth adopting as a habit: if a piece cannot pass the originality check after a reasonable attempt, publish nothing and repurpose the research into a piece where you do have something specific to say. A smaller site with thirty pieces that each contain something unrepeatable performs better than one with three hundred that contain nothing.