Where a human has to intervene when a model writes your content
The workflow around the model rather than the prompt inside it: the three points a person has to intervene, and a review prompt that holds a draft.
- Works in
- ChatGPT, Claude, Gemini
- You need
- A finished draft, pasted as text · One sentence on what you sell and to whom · Who the reader is
- Written for
- how to use chatgpt for seo content writing
Scored by our own engine
This page, run through the audit we sell. Measured 4 August 2026.
The prompt is the smallest part of this. A model can produce sentences and it cannot decide what is worth saying, so the question worth asking is not which prompt to use but where a person has to stop and do something the model cannot. There are three of those points, and the prompt below is the last of them: a review that reads a finished draft and holds it.
The three points where a person has to intervene
Before the draft, at every number, and on the last read. Everything between those three can be handed to a model, and nearly all of the quality of the published page is decided at them.
Before the draft is where the specifics come from: something that actually happened, an objection you get in real conversations, a constraint you hit. The drafting prompt covers that half and refuses to start without them. Every number is a checkpoint because a model asked to write persuasively reaches for a figure, and the figure it reaches for looks exactly like a measured one. The last read is structural, and it is the one people skip.
Why a review beats another rewrite
Because a rewrite changes the sentences and leaves the gaps where they were. Ask a model to improve its own draft and you get a different draft missing the same evidence, which is how an afternoon disappears.
A review does the opposite. It quotes the failing sentence, names the gate, and hands the work back to you. The gate that catches most drafts is the first one, fewer than three concrete details in the entire piece, and no rewrite fixes that. Somebody has to go and find something out.
People add “reddit” to this query for a reason. Most vendor advice here stops at a list of banned words, and practitioners worked out some time ago that swapping vocabulary changes the surface and nothing underneath it.
What the gate cannot see
Whether the article is worth publishing at all. Every gate measures the draft against itself: evidence present, answers first, tells absent. A draft can pass all five and still be a page nobody needed.
That judgement stays with you, and it is the reason this is a workflow rather than a prompt. Once the verdict is READY and the page is live, run it through the AI content readiness check to see the structural half measured from the outside, on the HTML a crawler actually receives rather than on the markdown you pasted.
You are reviewing a draft before it is published. Do not rewrite it. Report
what a human has to fix, and do not pass it while anything is unfixed.
Run these gates in order and report each one separately, by number.
Gate 1, specifics. Quote every sentence in the draft that contains a concrete
detail: a named example, a figure with a source next to it, a date, a decision
somebody actually made. If there are fewer than three, the gate fails. A
hypothetical illustration is not a specific, however detailed it is.
Gate 2, unsupported claims. Quote every sentence that states a number, a
percentage, a study, a survey, a trend or a market size, and write next to
each one either the source the draft itself gives, or UNSOURCED. Never supply
a source yourself, never estimate a figure, and never confirm one from memory.
If the draft cites something you cannot see, write CANNOT VERIFY.
Gate 3, answer first. For every H2, quote the first sentence under it and say
whether it answers the heading in 40 words or fewer. A sentence that restates
the heading, or that sets up the answer, fails.
Gate 4, tells. Quote the sentences that read as machine written and say why
for each one. What counts: even coverage that concedes nothing, a run of three
adjectives, an opening that repeats the heading, a closing section that
summarises the piece, a sentence of the shape "it is not just X, it is Y". Do
not give me a list of banned words. Quote the actual sentences.
Gate 5, the missing question. Name the one question this reader would ask that
the draft does not answer, and say who would have to be asked to answer it.
Rules:
1. Quote exactly, never paraphrase a sentence you are flagging. I have to find
it in the document.
2. Do not rewrite, improve, or offer replacement copy anywhere in the output.
This is a review. Suggesting the fix is how the specifics get replaced by
things you made up.
3. Do not pass a draft to be encouraging. If a gate fails, it fails, and you
say which sentences caused it.
4. If I have given you only a URL and not the draft text, say so and stop.
End with one line, in exactly this shape:
VERDICT: HOLD, fix gates <numbers>
or
VERDICT: READY
My business: [ONE SENTENCE ON WHAT YOU SELL AND TO WHOM]
The reader: [WHO IS READING AND WHAT THEY HAVE ALREADY TRIED]
The draft: [PASTE THE FULL DRAFT INCLUDING ITS HEADINGS]What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[ONE SENTENCE ON WHAT YOU SELL AND TO WHOM]- Sets what the draft is allowed to assume. "We audit sites for AI visibility, for marketing teams at software companies" makes gate 5 ask a technical question. Leave it out and gate 5 comes back with "what is AI visibility", which helps nobody.
[WHO IS READING AND WHAT THEY HAVE ALREADY TRIED]- The second half is what makes gate 5 useful. "They have already added FAQ schema and it changed nothing" tells the model which questions are already closed, so the one it names is the next real question rather than the first obvious one.
[PASTE THE FULL DRAFT INCLUDING ITS HEADINGS]- The whole draft as text, with the markdown headings intact, because gate 3 walks the H2s and cannot do that on prose with the structure stripped out. Paste rather than attach: some models summarise an uploaded file before reading it, and a summary has no sentences left to quote.
How to run it
- 01Collect the specifics before you open a model at all
Spend twenty minutes in your support inbox, your call notes or your own logs and find one thing that actually happened. This step is the whole workflow. A model cannot be prompted into having something to say, and every later step in this process is checking whether you did this one.
- 02Draft however you like, then leave it alone
Use a drafting prompt, dictate it, or write it yourself. The review does not care where the sentences came from, which is the point of separating the two jobs. Come back to it after a break, because you cannot review a draft you are still writing.
- 03Run the review and read gate 2 first
Paste the prompt with the full draft into ChatGPT, Claude or Gemini. Go straight to the unsourced claims, because those are the sentences that would do real damage in public, and they are usually the ones you did not notice yourself adding.
- 04Fix by cutting, not by asking for a rewrite
Deal with each flagged sentence by supplying the source, cutting the sentence, or writing the specific yourself. Do not hand the list back to the model and ask it to fix them. A rewrite pass restores fluency over the gaps and you lose the only signal you had.
- 05Answer the gate 5 question or narrow the article
Gate 5 names one question the draft leaves open. Either go and get the answer from whoever holds it, or narrow the piece so that question honestly sits outside its scope. Publishing with it open is what makes a reader feel the article stopped short.
- 06Run it again, then check the published page
Rerun the review on the edited draft until the verdict is READY, then publish and run the live URL through the AI content readiness check. Editing quietly reintroduces windups at the top of sections, and a first sentence that warms up is the one thing that stops a passage being quotable.
Questions people ask
How do you use ChatGPT for SEO content writing without it being obvious?
Treat the model as the drafting step in a workflow rather than the workflow. A person supplies the specifics before drafting, checks every number afterwards, and reads the final version for structure. The output reads as written when it contains things only you could contribute, and no prompt can supply those.
Should I review the draft with a model or edit it myself?
Both, in that order, because they catch different things. A model reads a long draft evenly and will quote every unsourced number and every windup without getting bored, which is exactly what a human editor stops doing by page three. Judging whether the piece is worth publishing is still yours.
Why does the prompt forbid the model from suggesting fixes?
Because a suggested fix is a sentence the model invented, and the reason a draft is thin is that it has too many of those already. Reviewing and rewriting in one pass produces a smoother draft with the same missing evidence. Keeping them separate forces the gap to stay visible until a person closes it.
Does Google penalise content written with AI?
Google guidance addresses content produced at scale to manipulate rankings rather than the tool used to write it. The practical risk is more ordinary: a draft with nothing specific in it competes with thousands of pages saying the same thing, so nothing has a reason to rank it or quote it.
How many times should I run the review?
Until the verdict is READY, which for a real draft is usually two passes. If it takes more than three, the problem is not the draft. Gate 1 failing repeatedly means the piece has fewer than three concrete details in it, and that is fixed by research rather than by another round of editing.
Unsubscribe in one click. We never pass your address on.