Last time I went digging into AI search advice and found that most of the confident statistics everyone repeats trace back to a single website that no longer exists.
The worst offender? A study that scored content using an AI, then announced that its AI scores predicted what Google would do.
So you’d think I’d be the last person on earth to tell you to run an AI content audit on your own writing.
But here’s the thing… those are two completely different jobs, and the difference is the whole point of this article.
An AI can tell you whether your paragraph makes sense on its own. It cannot tell you whether Google will cite it.
One of those is editing. The other is fortune-telling with extra steps.
Below are the six prompts I actually use before I hit publish, exactly as I paste them. Plus the three things AI genuinely cannot do here — which matters more than the prompts do.
Ask this, not that
Before the prompts, the mindset. Most people use AI wrong at this stage, and it’s always the same mistake.
| ❌ Don’t ask | ✅ Ask instead |
| “Is this article good?” | “What’s missing from this section?” |
| “Rate my post out of 10” | “List every claim here with no source” |
| “Will this rank on Google?” | “Does this paragraph work if read alone?” |
| “Make this better” | “What would a sceptic attack first?” |
| “Is this statistic true?” | “What source did I give for this?” |
See the pattern? Every question on the left asks for a verdict. Every question on the right asks for a list.
Verdicts are opinions, and an AI’s opinion of your work is worth roughly nothing. Lists are observations, and those are genuinely useful.
Why you can’t just ask “is this any good?”
Try it. Paste in anything you’ve written — a masterpiece, a first draft, a grocery list — and ask an AI to rate it.
It’ll tell you it’s great.
This isn’t a bug you can prompt your way around with a firmly worded instruction. These tools are built to be agreeable. Ask for praise and you’ll get praise, and you’ll feel wonderful, and your article will still have the same holes in it.
So every single prompt below is designed to find problems, not hand out marks. Notice how many of them are phrased as list or identify rather than evaluate.
That one reframe is worth more than all six prompts put together. Honestly, if you close this tab now and remember nothing else, remember that.
The six prompts
If you’ve read how to learn AI for beginners, you already know the critique prompt — what’s weak about this? These six are that same instinct, aimed at an article you’re about to publish.
Copy these as they are. They work in any chatbot — Claude, ChatGPT, Gemini, take your pick.
They’re numbered for reference, not for sequence. I don’t run them in this order — there’s a section at the end with the order I actually use, and why.
1. Does this paragraph survive on its own?
AI Overviews don’t quote whole articles. They lift out one passage and show it to somebody who never sees the rest of your page.
So the question is: would that passage still make sense?
Copy/Paste versions in boxes.
Below is one section from an article. Read ONLY this section — assume you have not seen the rest of the page and never will.
1. What questions does this section leave unanswered that a reader would need answered before they could act on it?
2. List every pronoun or back-reference ("this", "that approach", "these", "the tool", "as mentioned above") that points at something not defined inside this section.
3. Does the first sentence answer the question implied by the heading? Yes or no.
Do not rewrite anything. Just report.
HEADING: [paste your heading]
SECTION: [paste one section]
Why it works: this test needs a reader with zero context — and being instructed into total amnesia is one thing AI is genuinely good at.
What to do with the output: every dangling “this” and “that approach” is a place where someone reading the extract would be lost. Fix those first. They take seconds.
2. Is there any reason for this to exist?
This is the uncomfortable one. Run it before you get attached to your draft.
Here are the top five articles currently ranking for "[your target search]":
[paste their titles and main points, or just the URLs]
Here is my draft:
[paste your draft]
LIST 1 — everything in my draft that NONE of the five mention. (This is my original material.)
LIST 2 — everything in my draft that three or more of them also cover. (This is the overlap.)
If LIST 1 is empty, say so plainly in one sentence.
Be blunt. I would rather not publish than publish a duplicate.
What to do with the output: if LIST 1 comes back empty, you’ve written a summary of what already exists. Don’t publish it.
Go and get something first-hand — run the thing, screenshot the result, count something nobody’s counted. And then write.
Harsh? A bit. But an empty LIST 1 means there is no reason for your page to be chosen over the five that got there first.
This isn’t just my opinion, either. Google’s own guidance asks the same thing in almost the same words:
“If the content draws on other sources, does it avoid simply copying or rewriting those sources, and instead provide substantial additional value and originality?”
That’s the whole test, straight from Google. This prompt is just a way of answering it honestly.
3. Where did I get that number?
This is the Two-Click Test from part one, pointed at yourself.
Previously, I used it on other people’s articles: take any statistic, click the source it credits, then click the source that one credits.
If you don’t reach an actual methodology within two clicks, it’s folklore.
Here, you run it on your own draft before anyone else gets the chance.
Find every factual claim in this draft that contains a number — percentages, multipliers, correlations, counts, dates.
Put them in a table: the claim | the source I named | did I link it?
Flag every number with no source. Do NOT guess what the source might be and do NOT fill any in.That last line is not optional. Leave it off and the model will “helpfully” invent plausible-looking sources for your unsourced numbers — which is precisely the disease we’re trying to cure.
What to do with the output: for every flagged number, go and find the actual study. Not a blog quoting the study. The study. If you can’t find one in two clicks… cut the number. Your article survives without it.
4. Am I being vague when I could be specific?
1. List every named thing in this draft — people, organizations, products, tools, places.
2. List every place where I referred to something generically ("the tool", "the company", "one study", "a recent report") where a specific name could go instead.Why it matters: search engines understand named things far better than they understand “a popular tool.” Swapping “a scheduling app” for “Buffer” costs you nothing and makes the sentence better anyway.
What to do with the output: work down LIST 2 and name things. If you can’t name it, that’s usually a sign you didn’t check it closely enough.
5. Do my sections start with the answer?
People skim. AI extracts. Both reward putting the answer first and the throat-clearing… nowhere.
For each H2 heading in this draft, quote the first 30 words underneath it, then tell me whether those 30 words answer the question the heading implies.
One line per heading: heading | yes/no | why.What to do with the output: every “no” is a section that opens with a warm-up paragraph. Delete the warm-up. Your actual first sentence is usually hiding two sentences down.
I do this constantly and it is always an improvement.
6. Tear it apart
Save this for last, once the draft is basically finished.
You are a skeptical expert in this field who believes this article is wrong.
Identify the three weakest claims and explain exactly how you would attack each one. Do not be polite and do not balance it with praise.What to do with the output: ignore anything that’s just style grumbling. Look for the attacks you can’t answer. Those are real holes, and better you find them than a commenter does.
The other five prompts check your structure. This one checks whether you’re right.
The three things AI genuinely cannot do here
This is the important section, and it’s the one that’s missing from every other “AI prompts for bloggers” post you’ll read.
It cannot tell you whether Google will cite you. Nobody can. If a tool hands you a “GEO score” out of 100, ask what that score is measured against… and watch what happens.
We actually have a decent test case for this. Ahrefs tracked 1,885 pages that added schema markup against 4,000 pages that didn’t, and measured the citations.
The result was no uplift at all — a small 4.6% decline in AI Overviews, and changes elsewhere too small to mean anything.
Their conclusion: “Adding schema produced no major uplift in citations on any platform.”
If a real study of nearly 6,000 pages can’t find the effect, a chatbot guessing at your draft certainly can’t.
It cannot supply your original material. This is the big one.
The whole reason original work earns citations is that AI can’t generate it. If it could, it wouldn’t be original.
Your screenshots, your results, the thing that broke on step four: that’s yours or it doesn’t exist.
AI can polish it. It cannot produce it.
Worth knowing there’s real research behind this rather than just vibes. The peer-reviewed GEO study I went through in part one of this article series — around 10,000 searches — found three edits that reliably improved visibility: adding statistics, quoting credible named people, and citing your sources. All three require something you went and got. None of them can be generated.
It cannot tell you whether a fact is true. It will invent a source with total confidence and zero hesitation. Never ask “is this true?” Ask “what did I cite?” — then go and look yourself.
That third one is not theoretical. While researching the last article I found a guide about proving your experts are real that used “Dr. Sarah Chen” as its example of a credentialed expert.
She doesn’t exist.
She’s one of the most common fake names AI models produce, and she’d been published without anyone checking.
AI never sounds unsure. That’s the whole problem in one sentence.
The order I run them in
Don’t run all six at once. You’ll drown in output and fix nothing.
- Prompt 2 first — always. It’s a gate, not a step. If there’s nothing original in your draft, everything after it is polishing something that shouldn’t be published.
- Prompt 1 on your two or three most important sections. Not all of them. Life is short.
- Prompt 5 to fix your section openings.
- Prompt 4 to name the vague things.
- Prompt 3 to hunt down your unsourced numbers.
- Prompt 6 last, when the shape is settled.
Whole thing takes about forty minutes on a 2,000-word post. Faster once you’ve done it twice.
What this caught on my last article
I ran this on part one of this series before publishing it. The source check especially — by hand, one link at a time.
It found four problems.
Two statistics were credited to an article that didn’t contain them. One link pointed at a study whose website had gone offline entirely and now redirects to something unrelated. And one sample size was overstated by a factor of seventy — a whole dataset quoted when the finding came from a small slice of it.
Every single number was correct. Four of the sources were wrong.
That’s what this catches. Not invented statistics — real statistics wearing the wrong label.
Once a writer sees data that looks and smells correct AND they don’t check it but use it anyway, then someone else sees their data and uses it, it becomes this big thing everyone is quoting and THEY ALL HAVE IT WRONG.
And that’s the sneaky one, because a wrong number often looks wrong. A right number with a bad link looks perfect right up until somebody clicks it.
Don’t be that writer!!!!
Frequently asked questions
Which AI should I use for this?
Any of the main ones. These prompts don’t rely on a special feature — they work in Claude, ChatGPT or Gemini. Use whichever you already pay for, if you pay for one at all. The free tiers handle this fine.
Won’t editing with AI flatten my voice?
Not if you use it this way. Every prompt here asks the AI to report, not rewrite — that’s deliberate. You’re getting a list of problems and fixing them yourself, in your own words. The moment you let it rewrite whole paragraphs, yes, you’ll sound like everybody else.
Can AI tell me whether my post will rank?
No. And neither can anything else, whatever it charges you. What you can control is whether your writing is original, sourced, and clear enough to be quoted. That’s it. That’s the lever.
Is it cheating to run an AI content audit on your own work?
No more than running spellcheck. You wrote it, you’re checking it, you’re doing the fixing. The line worth caring about is the one between “AI helped me find the holes” and “AI wrote it and I skimmed it.”
How long does this take?
About forty minutes for a 2,000-word article the first time, less after that. Prompt 2 alone takes five minutes and will save you from publishing things that had no reason to exist.
Do this before you publish anything else
Here’s the thing about the last two years of AI writing advice: everyone’s been asking how to produce more. Almost nobody’s been asking how to check what comes out.
An AI content audit isn’t glamorous. It’s forty minutes of being told your third section doesn’t make sense and your favourite statistic has no source. Nobody’s selling a course in it.
But it’s the difference between publishing something that gets quoted and publishing something that quietly disappears into the pile with everything else.
Ask for lists, not verdicts. Ask what’s missing, not whether it’s good. And never, ever ask an AI whether your facts are true.
Which of these did you try — and what did it catch that you weren’t expecting? Tell me in the comments… I collect these now, apparently. 🤭
Part one of this series: I traced the AI SEO numbers back to their source, and the source no longer exists. Worth reading first if you want to know why I’m so twitchy about unsourced statistics.