Rewrite a finished draft for search without flattening the writing
An editing prompt that rewrites only the passages failing one of four named tests, quotes every original beside its replacement, and states the cost.
- Works in
- ChatGPT, Claude, Gemini
- You need
- A finished draft you or a writer already wrote · The query it should win · 150 words of your own writing you would not change
- Written for
- chatgpt prompts for seo writing
Scored by our own engine
This page, run through the audit we sell. Measured 4 August 2026.
An SEO rewrite is an editing job, and most prompts for it are drafting prompts wearing a different hat. Hand a model a finished draft, ask it to optimise the writing for search, and it returns the piece in its own voice: smoother, emptier, and unreviewable without diffing every line. The prompt below rewrites at most eight passages, puts each original next to its replacement, and has to name what the rewrite costs you.
Four faults are worth rewriting for
Four things in a draft genuinely cost search performance: the answer arriving late under a heading, a passage that cannot be read on its own, a heading nobody would say out loud, and a subject never named in plain words.
Those four have something in common. Each one is judgeable from the text in front of the model, with no data it does not have. Everything else an assistant flags when you ask it to improve a draft is preference, and preference is unlimited, which is why an unconstrained prompt rewrites the whole piece.
The second test is the one people find odd and it is the one that matters most for being quoted. Cover the paragraph above a passage and read it cold. If it now refers to “this approach” or “the problem above”, an assistant cannot lift it into an answer, because the passage does not contain its own subject.
Make the model say what it costs
Every rewrite trades something away. Forcing the model to name the loss on its own line turns an invisible trade into a line you can reject in two seconds.
The trades are consistent once you start reading them. Moving the answer to the first sentence usually costs an opening image. Making a passage self-contained costs a repeated noun the reader already met. Both are often worth it. Losing the one concrete detail in the paragraph never is, and without the cost line that loss arrives disguised as better prose.
Rule 3 does the other half of the same job. No new facts, no statistics, no sources, and a visible [NEEDS FACT] marker wherever the rewrite wanted one. A model asked to strengthen a paragraph reaches for a figure, and the figure it reaches for looks exactly like a real one.
Editing cannot give a draft something to say
It cannot. Structure is the only thing this prompt improves, and a well structured piece with no specific in it is an organised nothing that competes against every other page saying the same.
If the left alone list comes back longer than the rewrites, that is the diagnosis. Go back to the article drafting prompt, which refuses to start until you supply a real example and a real objection. For a page that is already published and already ranking, use the anchored edit prompt instead, and when this draft ships, run it through the AI content readiness check. If what you are holding is a draft a model wrote rather than one you did, start with how to humanize AI content, because search structure applied to empty sentences leaves you with well organised filler.
You are editing a draft I have already written. You are not writing it again.
Work passage by passage and change only the passages that fail one of the four
tests below.
The four tests. A passage may be rewritten only if it fails at least one:
A. The first sentence under a heading does not answer that heading. It
introduces, sets up or warms up instead.
B. The passage cannot be understood on its own. Read it with the paragraph
above it covered: if it now refers to something unnamed, it fails.
C. The heading above it is not a phrase anybody would say out loud when
asking about this. Headings that are labels rather than questions or
claims fail.
D. The passage is about the subject of my target query and never names that
subject in plain words.
Rules you must follow:
1. Rewrite at most eight passages. If more than eight fail, keep the eight
that cost me the most and tell me which ones you left and why.
2. Quote the original passage in full above every rewrite, copied character
for character from what I pasted. If you cannot reproduce it exactly, you
may not rewrite it.
3. Preserve every fact, number, name, date, price and claim exactly as I wrote
it. If a rewrite would need a fact that is not in my draft, write
[NEEDS FACT: what is needed] and leave the gap visible. Never supply the
fact yourself. No statistics, no studies, no named sources, no examples you
invented.
4. Use my vocabulary. Do not introduce a word that appears nowhere in my
draft, with one exception: the plain wording of my target query. Match the
sentence lengths in the voice sample I give you.
5. For every rewrite, write one line beginning "Costs:" naming what the new
version loses against the original. Every rewrite loses something. If you
cannot name the cost, you have not read the original closely enough.
6. Do not add the target query to a passage that already contains it, and add
it at most once to a passage that does not. Frequency is not the job.
7. Never add a summary paragraph, a closing section that restates the piece, a
transitional opener such as "Moreover", "Furthermore" or "Additionally", or
a bulleted list where I wrote prose.
8. A passage that fails none of the four tests may not be touched, however
much you would like to improve it. List it as left alone instead.
9. British spelling: optimisation, behaviour, analyse.
Output in this order. Do not use a table.
For each passage you rewrote:
PASSAGE [n] | Under the heading: [heading] | Test failed: [A, B, C or D]
Original: [the original, verbatim]
Rewrite: [the new version]
Costs: [what the rewrite loses against the original]
Then, in this order:
LEFT ALONE: every passage you considered and did not touch, with one line
each on which tests it passed and what is working in it.
MARKERS: every [NEEDS FACT] you wrote, gathered into one list.
FIRST: the single rewrite to apply first, and one sentence on why.
My target query: [THE QUERY THIS DRAFT SHOULD WIN]
Who is reading: [WHO THE READER IS AND WHAT THEY HAVE ALREADY TRIED]
My voice, do not deviate from it: [PASTE 150 WORDS OF THE DRAFT YOU WOULD NOT CHANGE]
The draft: [PASTE THE FULL DRAFT WITH EVERY HEADING MARKED]What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[THE QUERY THIS DRAFT SHOULD WIN]- One query, in the words a reader would type. It decides test D and nothing else, which is deliberate. "How to stop GPTBot being blocked" makes the model check that the draft names the crawler in plain words somewhere; "AI visibility" is broad enough that every passage passes and the test does nothing.
[WHO THE READER IS AND WHAT THEY HAVE ALREADY TRIED]- What they have already tried is the half that matters. "They have already added FAQ schema and nothing changed" tells the model that an explanatory passage is not a passage failing test B, it is a passage that should be cut, and it will say so in the left alone list rather than politely rewriting it.
[PASTE 150 WORDS OF THE DRAFT YOU WOULD NOT CHANGE]- The paragraph you are proudest of, pasted twice: once here and once inside the draft. This is the constraint that stops the rewrites converging on house style. Pick a passage with your real sentence rhythm in it, including the short sentences, because the model matches what it is shown rather than what you describe.
[PASTE THE FULL DRAFT WITH EVERY HEADING MARKED]- The whole draft as text with the headings still marked as headings, since three of the four tests are about the relationship between a heading and the passage under it. A draft pasted as flat prose loses that relationship and the model falls back to rewriting on taste.
How to run it
- 01Finish the draft first, badly if necessary
This prompt edits, so it needs something to edit. A half draft produces rewrites of sentences you were going to cut anyway, and the model has no way to tell a placeholder from a position you hold. Get to a complete piece with every section written, however rough, before you open an assistant.
- 02Choose your voice sample honestly
Read the draft and find the 150 words you would defend against an editor. It is usually not the opening. Whatever you paste becomes the target the rewrites are matched against, so choosing a bland paragraph gets you blander rewrites and choosing your best one raises the floor for the whole run.
- 03Run it and read the left alone list first
The passages the model refused to touch tell you more than the rewrites do. A run that left almost nothing alone means the tests are matching everything, which usually means your headings are labels rather than questions and the structural problem is above the sentence level.
- 04Reject on the cost line, not on the rewrite
Read the Costs line before the rewrite itself. A rewrite that trades a good opening image for an answer in the first sentence is often worth taking, and a rewrite that trades your one concrete detail for a cleaner sentence never is. The cost line makes that trade visible in a way a clean replacement does not.
- 05Resolve every marker by hand
Supply the fact or cut the sentence. Do not let a [NEEDS FACT] marker survive into the published piece and do not ask the model to fill it, because the figure it produces will look exactly like one you sourced. This is the step that decides whether the edited draft is more accurate than the original or less.
- 06Check the structure survived your own editing
Publish, then run the URL through the AI content readiness check to confirm the answer first structure is still there. Writers restore the windup they were trained to write, usually while doing one last pass for flow, and a first sentence that warms up is the thing that stops a passage being liftable.
Questions people ask
What is the best ChatGPT prompt for SEO writing?
For an existing draft, one that can only edit and has to justify every edit against a named test. Prompts that say "optimise this for SEO" return the piece rewritten in the model default voice, which reads more smoothly and has quietly lost the specifics you put in. Naming the four faults worth fixing is what turns a rewrite into an edit.
Will rewriting a draft with AI make it worse?
It can, and the mechanism is predictable. A model smooths whatever reads as unusual, and the unusual sentences are often the ones carrying your actual experience: the hedge added after a support complaint, the number from your own logs, the odd phrasing a customer used. The protections that matter are keeping the original next to every rewrite and forbidding any change to a fact.
How is this different from asking for SEO optimisation of a page?
That job runs against a live published page and should produce anchored edits rather than new prose, which is a separate prompt. This one runs against a draft nobody has seen, where rewriting whole passages is allowed because nothing has earned a ranking yet. Confusing the two is how a page that already ranks gets rewritten and loses ground.
Does answer first writing make an article read badly?
It reads blunt to a writer and fine to a reader. The cost is real: the sentence you would have opened with, the one that set a scene, moves down or goes. That is the trade the prompt makes you look at rather than hiding, because the same property that makes a passage abrupt is what lets a search engine or an assistant lift it whole.
Can I use this on a draft a model wrote?
Yes, and it will expose the real problem rather than fixing it. A generated draft usually passes test A and fails at having anything to say, so the rewrites come back small and the left alone list comes back long. If the output looks like there was nothing to do, the draft needs a specific only you have, not an edit.
Unsubscribe in one click. We never pass your address on.