Decide what one page should target, and make the model quote the page as evidence
A prompt that takes one URL and a shortlist of queries and returns a single verdict on what that page should target, with a sentence from the page as proof.
- Works in
- ChatGPT, Claude, Gemini
- You need
- One URL and the text of that page · Three to six candidate queries · The other pages on your site that touch the same topic
- Written for
- chatgpt prompt for keyword research
Scored by our own engine
This page, run through the audit we sell. Measured 4 August 2026.
Most keyword research prompts return a spreadsheet. Sometimes the actual question is much smaller: there is one page, there are four or five queries it could plausibly be about, and somebody has to choose. The prompt above makes that one choice, refuses to hedge it into a primary plus secondaries, and has to quote a sentence from your page before it is allowed to answer.
One page, one query, one decision
A page written for one query and a page written for four look identical in a plan and completely different once published. The four query version opens with a general paragraph, because it cannot commit to which question it is answering first, and a general opening is the thing both a searcher and an assistant skip.
Committing to one is not a claim that the page will only rank for one. It ranks for whatever it deserves. It is a decision about what the title says, what the H1 says, and what the first sentence answers, and those three are worth deciding on purpose.
The refusal is the feature
Rule 2 says the model must quote a sentence from your page as evidence, and stop if it cannot find one. That rule turns an agreeable assistant into something that can tell you no.
Ask a model whether your page suits a query and it will find a way to say yes, because saying yes is cheap and it has no stake in being right. Ask it to produce the sentence and the page either contains one or it does not. When it does not, you have learned that the page is about a topic rather than a question, and no amount of retargeting fixes that.
Where this sits in the workflow
Late. This is the last decision before writing, and it assumes you already have a shortlist worth choosing between.
If your shortlist came out of your own head, expand a seed keyword first and validate the results in a real tool. If you have a whole export and no plan, cluster it rather than running this page by page. If the page is already live and collecting impressions, triage what you already rank for instead, because Search Console has already answered part of this question for you.
You are a technical SEO strategist. I have one page and a shortlist of
queries. Choose the single query this page should target. One decision, with
reasoning. Do not produce a keyword list, a plan or a table.
Rules you must follow:
1. Pick exactly one primary query from my shortlist. Not a primary and two
secondaries, not a cluster, not a range. If the right answer is not on my
shortlist, say so plainly and name what is missing rather than choosing
the least wrong option.
2. Quote one sentence from the page text I pasted as evidence that this page
already answers that query. If you cannot find a sentence, say the page
does not currently target anything and stop there. A page that supports
no quotable sentence has a content problem, not a keyword problem.
3. Take every rejected candidate in turn and say in one sentence what a
searcher typing it wants, and why this page is not it. The rejections are
the part I am paying attention to.
4. If any other page I listed already answers the chosen query, do not pick
it. Say which page owns it and tell me either to merge the two or to
retarget this one, and which.
5. Do not estimate search volume, difficulty or competition. If I supplied
figures, use them and say you are using mine. If I supplied none, decide
on intent and fit alone, and say in one line that the decision contains
no demand information.
6. Judge the page as it is written, not as I described it. If my one line
summary of the page and the pasted text disagree, follow the text and
tell me they disagree.
Output in this shape, as short headed blocks and not as a table:
VERDICT: the one query.
EVIDENCE: the sentence from my page, quoted exactly.
CONFIDENCE: high, medium or low, plus the single thing that would raise it.
WHY NOT THE OTHERS: one line per rejected candidate.
WHAT THE PAGE MUST CHANGE: at most three edits that would make the page
unambiguously about the chosen query. If the answer is nothing, say nothing.
WHAT I CANNOT SEE: one line naming what you would need to be sure.
The page URL: [THE URL, OR "not published yet"]
Candidate queries: [THREE TO SIX QUERIES, WITH VOLUME IF YOU HAVE IT]
My other pages on this topic: [URLS AND WHAT EACH ONE IS ABOUT, OR WRITE "none"]
The page text: [PASTE THE H1, THE HEADINGS AND THE FULL BODY]What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[THE URL, OR "not published yet"]- The URL matters because the path is a signal you have already committed to. A page at /guides/rota-templates/ that you now want to target "shift planning software" is a retarget with a redirect attached, and the model should be told that before it recommends one.
[THREE TO SIX QUERIES, WITH VOLUME IF YOU HAVE IT]- A real shortlist, not everything you found. Three to six forces you to have already thrown out the obvious losers, which is where the model is least useful. Include your figures if you have them, and if you do not, rule 5 keeps the decision honest about it.
[URLS AND WHAT EACH ONE IS ABOUT, OR WRITE "none"]- This is what turns the verdict from a guess into a decision. Without it the model will happily assign a query your best performing page already owns, which is how one page targeting one query becomes two pages splitting it. Three or four neighbouring URLs is enough.
[PASTE THE H1, THE HEADINGS AND THE FULL BODY]- The evidence rule 2 is checked against. Paste the real text rather than describing it, because the gap between what a page says and what its author believes it says is the thing this prompt is most useful at exposing. For an unpublished page, paste the outline and the model will judge the outline.
How to run it
- 01Cut your candidates down to six before you start
Throw out anything you would not defend in a sentence. The model is weakest at sifting a long list and strongest at judging fit between a specific page and a specific query, so give it the second job. If you cannot get below six, you have a clustering problem rather than a targeting one.
- 02Copy the page text, not your description of it
Take the H1, every heading and the full body from the live page or the draft. Do not summarise. Rule 6 exists because authors describe the page they meant to write, and the model can only be useful here if it is reading the page that actually exists.
- 03List the neighbouring pages honestly
Include any URL of yours that a search engine might reasonably serve for one of your candidates, even the old post you forgot about. This is the input that activates rule 4, and rule 4 is the one that stops the verdict creating a competing pair.
- 04Read the rejections before the verdict
The verdict is one line and the rejections are where the reasoning is. If the model rejects a candidate for a reason you disagree with, that disagreement is the actual decision and it is now explicit, which is more than a keyword tool gives you.
- 05Act on the evidence line, or accept the refusal
If the model could not quote a sentence, do not rerun it with a different shortlist. It has told you the page does not answer anything specific yet, and the fix is writing rather than targeting. That refusal is the most valuable output this prompt produces.
- 06Make the page unambiguous, then check the structure
Apply the edits under WHAT THE PAGE MUST CHANGE, then run the URL through the heading structure checker. A page targeting one query needs one H1 and sections a machine can read as answers, and a clear target with a broken outline still fails to get quoted.
Questions people ask
What is the best ChatGPT prompt for keyword research on a single page?
One that forces a single choice and demands evidence for it. Asking a model which keyword a page should target usually returns a primary keyword and several secondaries, which is not a decision. Requiring one query, a quoted sentence from the page proving it fits, and a stated reason for every rejection is what makes the output actionable.
Should a page target more than one keyword?
A page ranks for many queries and should be written for one. Those are not in conflict. The primary query decides the title, the H1 and the opening answer, and the variations follow from writing that subject properly. Writing for three at once produces a page that is the best answer to none of them.
Why does the prompt make the model quote my page back to me?
Because it is the one check that cannot be passed by agreement. A model asked whether a page suits a query will usually say yes, since agreeing costs nothing. Requiring a verbatim sentence as evidence turns that into a test the page can fail, and a page that fails it needs writing rather than a different target.
What if the right query is not in my shortlist?
The prompt is told to say so rather than pick the least wrong option, which is what most models do by default. When that happens, the shortlist was built from your product words and the query you need is phrased in the words a customer would use. Run an expansion pass and rebuild the shortlist.
How is this different from clustering a keyword export?
Clustering runs across a whole list and decides how many pages you need. This runs on one page you already have or are about to write, and decides what it should own. Clustering is planning and this is a single commitment, which is why the output is a verdict rather than a table.
Unsubscribe in one click. We never pass your address on.