Take one keyword to an annotated outline, then to a draft, with a stop in between
A two stage prompt that builds an annotated outline from the results you paste, stops and waits for approval, then drafts only from the outline you approved.
- Works in
- ChatGPT, Claude, Gemini
- You need
- One keyword · The titles and H2s of the results currently ranking for it · One thing you know that those results do not say
- Written for
- chatgpt prompts for seo blog writing
Scored by our own engine
This page, run through the audit we sell. Measured 4 August 2026.
Most blog writing prompts go from keyword to finished post in one pass, which means the structure is decided somewhere inside a thousand words nobody has read yet. The prompt below builds an annotated outline, stops, and refuses to write until you reply GO. The stop is the point. Everything expensive happens after it.
The outline is where the post is decided
An outline is only useful if it commits to something. This one gives every H2 the question it answers, the answer in one sentence, the evidence that section needs, why it exists, and what it must not repeat.
The reason line does the most work. “Because the ranking pages have it” is explicitly rejected as a reason, which strips out the sections that exist only because everybody else has them. Those sections are why so many posts on the same keyword read identically: each was assembled from the headings of the others.
Rule 2 makes the judgement visible. The model has to name what the ranking results all cover that it is deliberately leaving out, and say where each one belongs instead. That list is the only part of the output that explains itself, and it is the part worth reading first.
Paste what ranks, do not ask
A model cannot see a search engine. Ask it what ranks for your keyword and it will produce a fluent description of pages that may not exist, and the outline built on that description will feel researched.
So the prompt takes the results as an input and tells the model to say so when it has none. Copying the titles and H2s from a handful of real results takes a few minutes in an incognito window, and it is the difference between planning against your actual competition and planning against a memory of the internet.
The other input doing real work is the last one: what you know that those results do not say. Rule 5 checks it, and if nothing you offered is genuinely missing from the pages you pasted, the prompt tells you the post has no reason to exist yet. That is an unwelcome answer and usually a correct one.
The answer sentence carries into the draft
Stage 2 opens every section with the exact sentence you approved in the outline, word for word. Nothing else guarantees the answer first structure survives drafting.
Left to itself, a model rewrites its own answer line into something warmer that introduces the section rather than answering it, and the passage stops being liftable. Copying it verbatim removes the choice. Once the post ships, confirm the outline actually made it into the markup with the heading structure checker, then take the draft rewriting prompt to whatever your own editing pass loosened.
You are planning and then writing one blog post. This runs in two stages and
you must stop between them.
STAGE 1. Build the outline. Do not write the post.
For the keyword I give you, produce an annotated outline. For every H2:
- The question that section answers, phrased the way somebody would ask it
out loud.
- The answer to that question in one sentence of forty words or fewer. This
sentence is not a placeholder. Write it properly, because it becomes the
first line of the section in stage 2.
- What evidence the section needs, and who has to supply it: me, a public
source I would have to find, or nothing beyond reasoning.
- Why this section exists, in one line. "Because the ranking pages have it"
is not a reason. If the only reason is that competitors cover it, drop it.
- What this section must not repeat from any other section.
Rules for stage 1:
1. Build the outline from the results I pasted, not from what you assume
ranks. You cannot see a search engine. If I wrote "none", say so at the top
of your output and build the outline from the reader and the keyword alone,
and do not describe what is currently ranking.
2. Name the sections the ranking results all cover that you are deliberately
leaving out, and say why each one belongs elsewhere or nowhere.
3. Give me at most six H2s. A seventh section is nearly always two sections
that should have been one, or a second post.
4. Every H2 must answer a question no other H2 answers. If two overlap, merge
them in front of me and say what you merged.
5. Mark the one section that carries the thing I said the ranking results do
not say. If nothing I gave you is genuinely absent from those results, tell
me the post has no reason to exist yet and stop.
6. Do not invent statistics, studies, survey results, dates, company names or
customer quotes anywhere in the outline.
Then stop. End your output with: "Reply GO to draft, or tell me what to
change." Write nothing further until I reply.
STAGE 2. Draft from the approved outline and nothing else.
1. Every section opens with the exact answer sentence from the approved
outline, copied word for word as its first sentence. Then the detail.
2. Write only the sections in the approved outline, in that order. No new
sections. No introduction that summarises the post. No conclusion that
restates it.
3. Where the outline said a section needs evidence I have not supplied, write
[NEEDS EVIDENCE: what is needed] in place of the claim and continue. Never
supply the evidence yourself.
4. Every section must be readable with the section above it covered. Do not
open a paragraph with "this", "that approach" or "as mentioned above".
5. No paragraph past four sentences. At least one paragraph that is a single
short sentence. No list where prose would do.
6. Banned: "in today's fast paced world", "in the digital landscape", "it is
worth noting", "delve", "leverage", "seamless", "unlock", "elevate". Do not
begin a sentence with "Moreover", "Furthermore" or "Additionally".
7. British spelling: optimisation, behaviour, analyse.
End the draft with the list of every [NEEDS EVIDENCE] marker you left.
My keyword: [THE KEYWORD, IN THE WORDS SOMEBODY TYPES]
Who is searching it: [WHO IS SEARCHING AND WHAT THEY HAVE ALREADY TRIED]
My business: [ONE SENTENCE ON WHAT YOU SELL AND TO WHOM]
What currently ranks: [PASTE THE TITLES AND H2s OF THE TOP RESULTS, OR WRITE "none"]
What I know that those results do not say: [THE THING ONLY YOU CAN ADD]What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[THE KEYWORD, IN THE WORDS SOMEBODY TYPES]- One keyword, not a title and not a topic. "How to check if GPTBot is blocked" produces an outline with a sequence in it. "AI crawlers" produces an outline that surveys a subject, which is the shape of post that covers everything a reader might want and answers nothing they came for.
[WHO IS SEARCHING AND WHAT THEY HAVE ALREADY TRIED]- The second half decides how many sections you get. "They have already read the vendor blog posts and want to know what to actually change" removes two explanatory H2s before they are written, and those removed sections are most of what makes a post feel like every other post on the query.
[ONE SENTENCE ON WHAT YOU SELL AND TO WHOM]- This tells the model what the post is allowed to assume and where it is allowed to be technical. Without it the outline is written for a general reader, and a general reader is nobody. It also keeps the model from planning a section that quietly pitches something you do not sell.
[PASTE THE TITLES AND H2s OF THE TOP RESULTS, OR WRITE "none"]- Search the keyword in an incognito window, open the first handful of results, and copy their titles and headings. This is the input that makes rule 2 work, because leaving something out on purpose requires knowing what everybody else put in. A model asked what ranks will describe pages it has never seen.
[THE THING ONLY YOU CAN ADD]- One thing you have seen that is absent from every result you pasted: a failure mode you keep hitting, a step the guides skip, a number from your own data. Rule 5 checks it against those results and will tell you the post has no reason to exist if it is already covered, which is worth knowing before a writer is paid.
How to run it
- 01Collect what ranks before you open an assistant
Search your keyword in an incognito window and copy the titles and the H2s from the first handful of results into a note. It takes a few minutes and it is the difference between an outline grounded in the page a reader will compare you against and an outline built from a model guessing what a search engine shows.
- 02Write down the thing you can add
Read the headings you just collected and find the gap you can genuinely fill. If everything you know is already in there, stop: the honest conclusion is that this keyword does not need another post from you. Rule 5 will reach the same conclusion, and it is cheaper to reach it before you have run anything.
- 03Run stage 1 and edit the answer sentences yourself
The one sentence answers under each H2 are the most important lines in the whole run, because stage 2 copies them into the draft verbatim. Rewrite the ones that hedge or that restate the question. Correcting them here is a single pass; correcting them after the draft means editing the same idea in two places.
- 04Read the deliberately excluded list before you approve
Rule 2 gives you the sections everybody else covers that this outline leaves out. That list is the only place the outline explains its own judgement. If it excluded something your reader genuinely needs, add it back before replying GO rather than patching the draft afterwards.
- 05Reply GO only when the outline is one you would defend
The stop exists so the expensive part runs on a plan you agreed to. Approving an outline you have skimmed produces a draft you rewrite, which costs more than the ten minutes the outline needed. If more than one or two H2s are wrong, change the inputs and rerun stage 1 rather than editing the outline in place.
- 06Check the headings that shipped are the headings you planned
Publish, then run the URL through the heading structure checker. Outlines drift in a CMS, where an H2 becomes a styled paragraph or a level gets skipped, and headings are where both search engines and assistants decide where one answer stops and the next begins.
Questions people ask
What is the best ChatGPT prompt for SEO blog writing?
One that separates the outline from the draft and refuses to cross the line without your approval. A single prompt that goes from keyword to finished post gives you a thousand words built on a structure nobody checked, and the structure is the part that decides whether the post is worth reading. Approving an outline takes minutes and saves a rewrite.
Can ChatGPT see what is currently ranking for my keyword?
Not reliably, and it will answer as though it can. Even with browsing enabled, what a model retrieves is often a cached or partial version of a page, and without browsing it produces a confident description of results that do not exist. Paste the real titles and headings yourself. It is the single input that most changes the quality of an outline.
How many H2s should a blog post have?
As many as there are distinct questions, which is usually fewer than a model offers. The prompt caps it at six because a seventh section is nearly always two sections that should have been one. A post organised into a few sections that each answer something is easier to read and easier for an assistant to lift an answer out of than one broken into twelve.
Why does the draft have to reuse the outline sentences word for word?
Because that is the only way answer first structure survives the drafting stage. Left free, a model rewrites its own answer sentence into a warmer opening that sets the section up rather than answering it, and the passage stops being quotable. Copying the approved line verbatim makes the structure a fact rather than an instruction the model has to keep in mind.
Will this write a publishable post on its own?
No. It will write a well structured draft with visible gaps where evidence should be, and closing those gaps is your job. A model cannot be prompted into having something to say, which is why the prompt makes you name what you know that the ranking results do not, and stops if the answer is nothing.
Unsubscribe in one click. We never pass your address on.