A keyword clustering ai prompt should do one job: sort a seed list into groups that share a single search intent, then tell you which groups deserve their own page and which should be folded into a page you already have. Give the model your seeds, the site it serves, and an explicit duplicate-intent rule, and it will return a draft you can act on.
What it will not return is judgement about your business. A model cannot see your search volume data, your current rankings, or the pages already live on your site. It groups language well. It does not know what you can credibly write about. This guide gives you the prompt, the dedupe pass that stops two pages chasing the same query, and the checks that catch a bad cluster before it becomes a brief.
The keyword clustering ai prompt that separates seeds into intent groups
You are clustering keywords for [site, product or brand], which serves [audience in one line] and currently has these live pages: [list existing URLs or page types, or write "none yet"]. Below are [number of seeds, e.g. 100] seed keywords taken from [source, e.g. a search console export, a brainstorm, a third-party list]. Group them by shared search intent. Intent means the job the searcher is trying to finish, not shared vocabulary. Use only these intent labels: [labels, e.g. what is it, how do I do it, which should I buy, comparison, pricing, troubleshooting]. For each group return: the group name, the seed keywords in it, the question one reader is asking, the page type that would answer it, and a one-line note on whether the group is large enough to justify its own page. Rules. Do not invent search volume numbers, keyword difficulty scores, or ranking estimates. If two seeds could be answered by a single page, flag them as a duplicate-intent conflict instead of quietly placing them in separate groups. Do not create a group with fewer than [minimum seeds] members without saying why it matters. List the three groups you are least confident about and explain the doubt. Seeds: [paste list].
Fill in these details
- Site or brand: The thing the clusters will serve. This keeps groups tied to a real offer rather than a generic topic list.
- Audience: Who is searching. Two readers with the same words can need different pages.
- Existing pages: Without this, the model will happily propose a second page on a topic you already cover.
- Seed source: Say where the list came from. A list from your own search console is worth more than a generated list.
- Intent labels: Give the model a fixed set. Free-form labels produce near-duplicate categories.
- Minimum seeds: A floor below which the model must justify a group, usually three.
If you have no seed list, ask the model for candidates first, then treat them as unverified hypotheses and confirm them in a keyword tool before spending a page on them.
A second pass that turns each group into a page decision
Using the clusters you produced, build a decision table with these columns: cluster, primary intent, proposed slug, page type, the single question the page answers, existing page it overlaps with, and a verdict. The verdict must be one of: build now, build later, fold into an existing page, or drop. Give one sentence of reasoning per verdict. Never recommend two new pages for the same intent. Never estimate traffic or difficulty. If a cluster is genuinely undecidable, write "needs research" and name the specific check that would decide it.
Run this after you have corrected the first pass. The value of the second prompt is the verdict column, which forces a decision rather than another taxonomy.
The dedupe pass that keeps two pages out of the same query
Most prompt libraries mention duplicate intent and stop there. It is the rule that decides whether your next six months of writing helps or cancels itself out, so it deserves its own pass.
Two clusters are duplicates when a single page could satisfy both honestly. If the same reader asks both questions in one sitting, they belong on one page. If answering both well makes the page incoherent, they are separate.
Review these clusters for duplicate intent. For each pair that a single page could serve, output: the two cluster names, the shared reader question, the merged cluster name, and the reason they are not worth separate pages. Then list any cluster that is too small to stand alone and name the page it should fold into. Do not merge clusters that need different formats, such as a tutorial and a comparison table. Do not merge across different stages of a decision, such as learning something and buying it.
Worked example: a hundred seeds for a small service business
Suppose a bookkeeping firm exports a hundred seeds from a brainstorm and its search console, then runs the first prompt. A workable output looks like this.
| Cluster | Intent | Seeds | Proposed page type | Verdict |
|---|---|---|---|---|
| Choosing a bookkeeper | Comparison | best bookkeeper for small business, when to hire a bookkeeper | Comparison guide | Build now |
| Bookkeeping software vs a person | Comparison | quickbooks vs bookkeeper, do i need a bookkeeper | Comparison guide | Merge into the row above |
| GST filing deadlines | Informational | gst due date, gst penalty for late filing | Reference page | Build now, one per deadline type |
| What bookkeeping costs | Cost | how much does a bookkeeper charge | Pricing explainer | Fold into services page |
The merge is the important row. "Quickbooks vs bookkeeper" and "best bookkeeper for small business" reach the same reader on the same afternoon. Two pages would split the authority and leave both thin.
Check the deadline cluster before building it. Deadlines change, and a page that is confidently wrong about a penalty date does real harm to the reader who trusts it.
How to check a cluster before it turns into a brief
1. Read the group's keywords as one sentence
Say them out loud in order. If they form a coherent question a person would type, the cluster probably holds. If they read as a list of unrelated nouns sharing a suffix, split it.
2. Confirm one page can answer the whole group
Take the widest seed in the group and ask whether your page will genuinely satisfy it. The widest seed sets the page's real topic. If the page cannot answer it well, the cluster is a cluster of two clusters.
3. Cross-check against pages you already have
Search your own site for each group's primary question. An existing page that already answers it needs an edit, not a new URL. Publishing a second page on the same question is the most common and most avoidable mistake in this workflow.
4. Drop what you cannot serve
A cluster for a subject you have no experience writing about is a liability, not an opportunity. Rank it as "build later" and move on.
Common keyword clustering prompt mistakes
- Grouping by shared words: The model will happily create a "best" cluster. Override it and group by intent.
- No existing-page list: Expect near-duplicate proposals. Paste your URLs.
- Asking for volume or difficulty: The model will produce confident numbers that mean nothing. Forbid them in the prompt.
- One page per cluster by default: A verdict column forces the merge or fold decision.
- Trusting the least confident groups: The model flags them for a reason. Check those first.
- Clustering before you have real seeds: Generated keywords are guesses. Confirm volume and difficulty with a real tool first.
Frequently asked questions
Can ChatGPT do keyword clustering on its own?
It can group a seed list by intent reasonably well once you give it intent labels and a duplicate-intent rule. It cannot supply reliable search volume, difficulty, or SERP data, so treat its output as a draft structure rather than a finished plan.
How many keywords should I cluster at once?
Work in batches that fit in one conversation, usually 50 to 100 seeds per pass. Larger lists get summarised rather than clustered. Keep the same intent labels across batches or the groups will not line up.
Do I need a keyword tool if I use prompts?
For grouping, no. For deciding whether a cluster is worth a page, yes. Tools tell you whether anyone is searching; the prompt tells you what those searches want. Use each for the job it is actually good at.
Should every cluster become a page?
No. Most good clusters should be folded into an existing page or built much later. A calendar of twelve new pages a quarter is a plan most small teams cannot execute, and unexecuted plans do not compound.
Cluster one seed list, then keep the file
Run the first prompt on your real seeds, run the dedupe pass, and write the verdicts into a spreadsheet you keep. The value is in the decisions you record, not the clusters themselves. For more prompts covering the rest of the search workflow, browse the SEO optimization prompt collection.