All promptsOn page

Optimise an existing page without letting the model rewrite it

Returns a short list of anchored edits to a page you have already published, rather than a rewrite. It quotes the exact text to change and alters no fact.

Works in
ChatGPT, Claude, Gemini
You need
The full text of one page, including its title tag and meta description · The query the page should win · Any audit or Search Console findings you already have
Written for
chatgpt prompts for seo optimization
98A

Scored by our own engine

This page, run through the audit we sell. Measured 4 August 2026.

Score your own page →

Paste a page into any assistant, ask it to optimise the page for SEO, and you get a rewrite. It reads more smoothly than what you had, it has lost the price, the caveat and the sentence a customer once quoted back to you, and it is unreviewable. The prompt below cannot rewrite anything, because every change it proposes has to quote the exact text it changes.

Why the default answer is always a rewrite

A rewrite is the cheapest thing a model can produce and the most impressive thing it can hand back. Given a page and a vague instruction, reproducing it in a cleaner voice is a single pass, and it looks like effort in a way that eleven small edits do not.

What it costs you is everything the model had no source for. The specific numbers, the hedge you added after a support complaint, the phrasing that came out of a sales call: those are the parts most likely to be smoothed away, because they are the parts that read as odd to a system predicting the next likely sentence. Anchored edits are also reviewable in one sitting. A rewrite has to be diffed against the original before you can trust a word of it, and nobody does that.

The edits worth making are the ones a machine reads

The title tag, the meta description, the H1, the first sentence under each heading, and the anchor text of the links pointing at the page. Those five decide what gets indexed and what an assistant can lift out of the page whole.

Everything else is a matter of degree. That is why the prompt makes the model group its edits and why the instructions tell you to apply that group on its own: it is small, it is reversible, and it is the only group whose effect you can attribute afterwards.

When the honest answer is that the page is fine

Sometimes nothing on the page is wrong. A page can match its query well, answer it in the first line, and still not rank, because another page on your own site is competing for the same query.

A prompt that can only produce edits will always produce edits, which is what rule 7 exists to prevent. If the model tells you the page is well matched and you still see nothing in Search Console, stop editing and go up a level: run your keyword list through the keyword clustering prompt, which flags exactly this, the two pages of your own that should never have been aimed at one query.

The prompt 429 words
You are a technical SEO editor. I am going to give you the full text of one
page that is already published. Return edits to it. Do not return a new
version of the page.

Rules you must follow:

1. Every edit must quote the exact text on the page that it changes, copied
   character for character from what I pasted. If you cannot quote the
   existing text exactly, you may not propose that edit.
2. Every edit must be one of four types: replace this text with this text,
   insert this text after this text, delete this text, move this section to
   here. Nothing vaguer. "Tighten the introduction" is not an edit.
3. Preserve every fact, number, name, price, date and claim exactly as
   written. You cannot verify any of them, so you may not change one, round
   one, or add one. If a claim looks wrong or unsupported, put it in the
   check this list instead of editing it.
4. Group every edit as one of: what a machine reads, what a person reads,
   style. The first group is the title tag, the meta description, the H1, the
   first sentence under each heading and the anchor text of internal links.
   Do not attach a percentage, a traffic estimate or a ranking position to any
   edit. You cannot know those.
5. Give me at most twelve edits. If you have more, keep the twelve that matter
   and tell me what you dropped and why.
6. Do not increase how often the target query appears. If it is missing from
   the title, the H1 or the first hundred words, propose one edit for each of
   those and no others.
7. If the page is already well matched to its target query, say so and stop.
   Do not manufacture edits to fill the table. A page can underperform for
   reasons that are not on the page.

Output a markdown table with these columns:
# | Type | Exact existing text | Proposed text | Group | Why this helps

After the table, give me three things:
- The check this list: claims, numbers or dates I should verify myself.
- What you deliberately did not touch, and why it is working.
- The single edit to make first, with one sentence of reasoning.

Target query: [THE QUERY THIS PAGE SHOULD WIN]
What the page is for: [THE ONE ACTION I WANT A READER TO TAKE]
Known problems: [PASTE AUDIT OR SEARCH CONSOLE FINDINGS, OR WRITE "none"]
The page: [PASTE THE FULL TEXT, INCLUDING TITLE TAG, META DESCRIPTION AND EVERY HEADING]

What to change

Everything in square brackets is yours to replace. Nothing else needs editing.

[THE QUERY THIS PAGE SHOULD WIN]
One query, and ideally one the page already gets impressions for. Optimising a page toward a query it has never appeared for is a rewrite wearing an edit costume, and the honest version of that job is a new page.
[THE ONE ACTION I WANT A READER TO TAKE]
Without this the model optimises for reading rather than for doing, and quietly proposes edits that soften your call to action. "Book a survey" and "download the spec sheet" pull the same page in different directions, so say which one you are actually paid for.
[PASTE AUDIT OR SEARCH CONSOLE FINDINGS, OR WRITE "none"]
What you already know is wrong, in evidence form. A line like "no H1, and the answer to the query is in the fourth paragraph" turns the model from a guesser into an editor working from a brief. The free content readiness check gives you exactly this in a form you can paste.
[PASTE THE FULL TEXT, INCLUDING TITLE TAG, META DESCRIPTION AND EVERY HEADING]
The whole page as text, not a URL and not a summary. The title tag and the meta description are the two highest leverage lines on most pages and neither one appears in the visible text, so a copy and paste from the rendered page silently removes them from consideration.

How to run it

  1. 01
    Copy the page as text, including the parts you cannot see

    Take the visible text, then open the page source and add the title tag and meta description at the top of what you paste. Keep the headings marked as headings. Do not paste a URL and ask the model to fetch it, because what it retrieves is often a different version of the page or nothing at all.

  2. 02
    Run the free check and paste the findings in

    Run the URL through the AI content readiness check and copy what it reports into the known problems field. That converts the run from an opinion into an edit brief, because the model is then working from measured faults rather than guessing which of a thousand possible improvements you need.

  3. 03
    Pick the query the page already competes for

    Open Search Console, filter to the page, and look at the queries it gets impressions for. Choose one of those rather than the query you wish it ranked for. A page with impressions is a page that is close, and close is what an edit can fix.

  4. 04
    Verify every anchor before you apply anything

    Search your page for each piece of exact existing text in the table. Anything you cannot find character for character was reconstructed from memory rather than read, and the whole edit goes, not just the quote. This one check catches most of what goes wrong with this kind of prompt.

  5. 05
    Apply the machine reads group first, on its own

    Make the title, meta description, H1, opening sentence and link text edits together and leave the rest for now. That group is small, it is reversible, and it is the group whose effect you can actually attribute later. Changing everything at once means learning nothing from the result.

  6. 06
    Re run the check and wait for a recrawl

    Run the content readiness check again to confirm the structural faults are gone. Then wait for the page to be recrawled before judging anything, because the version being ranked is the version the crawler last saw, not the one on your screen.

Questions people ask

Why does ChatGPT rewrite my page when I ask it to optimise it?

Because a rewrite is easier to produce than an edit and it looks like more work. Asked to improve a page, a model reproduces it in its own voice, which feels like progress and quietly deletes the specifics only you knew. Getting edits instead requires forbidding the rewrite explicitly and forcing every change to quote the text it changes.

Should I paste the page text or give the model the URL?

Paste the text. Assistants that fetch a URL often get a cached copy, a cookie wall or a version rendered without JavaScript, and none of those is the page you are editing. Pasting also lets you include the title tag and meta description, which are not part of the visible page but are usually where the useful edits are.

How many edits should I apply at once?

Apply one group at a time, starting with the edits a machine reads. If you change twelve things in one sitting and the page moves, you have learned that something worked. If you change five related things, you have learned which kind of thing worked, which is worth more on the next page.

Can I use this on a page that already ranks well?

Yes, and it is where the prompt earns its constraints. The rules that preserve every fact, cap the edit count and allow the model to conclude that nothing needs changing all exist so a working page survives the exercise. A page that ranks has earned something, and a rewrite forfeits it.

What do I do with an edit that quotes text my page does not contain?

Delete the whole edit rather than fixing the quote. A wrong anchor means the model was working from a paraphrase of your page rather than the page, so the reasoning attached to it was formed about text that does not exist. The rest of the table is still usable, and that row is not.

A new prompt, most days One working prompt for a real SEO or AI visibility job, what to change in it, and a worked example. No sequences, no offers dressed as newsletters.

Unsubscribe in one click. We never pass your address on.

Run your first audit
in about a minute

Free account, no card. Paste your URL and get a real, scored report of your AI and search visibility.

Measuring rankings in