The keyword research prompts worth running, and the order they go in
The five stages of keyword research, which prompt belongs to each, and a router prompt that tells you which stage you are actually at.
- Works in
- ChatGPT, Claude, Gemini
- You need
- An honest description of what you already have · Whether you have any real search data
- Written for
- best chatgpt prompts for keyword research
Scored by our own engine
This page, run through the audit we sell. Measured 4 August 2026.
Eight prompts on this site cover keyword research and they are not alternatives to each other. Each one consumes what the previous one produces, so the useful question is not which prompt is best but which stage you are actually at. The prompt above answers that and refuses to do anything else, including the research itself.
The sequence
Five stages, and the middle one has no prompt because no language model can do it.
- Expansion. One idea, no list. Expand a seed keyword into a set worth validating, which generates the phrasings a keyword tool never recorded because nobody searched them often enough.
- Validation. Candidates, no evidence. This happens in Keyword Planner, Search Console or a paid tool. There is nothing to prompt.
- Clustering. A validated list, no page plan. Turn a keyword export into topical clusters, which decides which terms share a page and names the two pages that would compete.
- Targeting. A plan, and one URL still undecided. Decide what one page should target, which is one decision rather than a spreadsheet.
- Triage. Pages that already rank. Triage the keywords you already rank for, starting from a Search Console export rather than from candidates.
Where to start if none of that made sense
At the definition, then the method. If keyword research is a new phrase, what keyword research is explains it using something you sell rather than an example about running shoes.
If you have a business and no data at all, how to do keyword research works from first principles and assumes no tool. If your question is really about the model rather than the method, how to use ChatGPT for keyword research sets out what it is and is not good at here.
The stage everybody skips
Validation, every time. It is the slow one, it costs money or patience, and skipping it feels harmless because the output of stage one already looks like a keyword list.
It is not one. It is a list of phrasings a model considers plausible, and clustering it produces a page plan built on demand nobody has measured. The stop condition for stage two is boring and worth holding to: every term you plan a page around has a figure next to it that came from somewhere other than a language model.
You are a technical SEO strategist. I am going to describe what I currently
have. Diagnose which stage of keyword research I am actually at, and tell me
the one thing to do next. Do not run the keyword research itself.
These are the five stages, in order. A stage cannot be skipped, because each
one consumes what the previous one produces.
1. EXPANSION. I have an idea of what I sell and no list. The job is
generating candidate phrasings, including the ones that contain none of
my product words.
2. VALIDATION. I have candidates and no evidence anybody searches them. The
job is getting real figures from a real data source. No language model
can do this stage and no language model should be asked to.
3. CLUSTERING. I have a validated list and no page plan. The job is deciding
which terms share a page and which two planned pages would compete.
4. TARGETING. I have a plan or an existing page and one undecided URL. The
job is choosing the single query one page should own.
5. TRIAGE. I have pages that already rank and appear in Search Console. The
job is deciding what to do about positions I already hold.
Rules you must follow:
1. Name exactly one stage. If my description spans two, name the earlier one
and say which sentence of mine made you pick it. People consistently
believe they are one stage further on than they are.
2. Give me one next action. Not a plan, not a list of options, not a
sequence. One. If I do that one thing I should reach the end of the
current stage.
3. Tell me the stop condition: the observable thing that will be true when
this stage is finished. Not a feeling of completeness. Something I could
show somebody.
4. Name the stage I am most likely to jump to instead, and say in one
sentence what breaks if I do. Almost everybody tries to skip validation.
5. Do not produce keywords, volumes, difficulty scores or a content plan. If
I asked for those inside my description, refuse and say which stage they
belong to. You have no search data and this prompt is not the place to
pretend otherwise.
6. If my description is too vague to place, ask me exactly one question and
stop. Do not guess a stage and do not ask three questions.
Output in this shape, as short headed paragraphs and no table:
STAGE: the number and name, plus the sentence of mine that decided it.
WHY NOT THE STAGE AFTER: two sentences.
DO THIS NEXT: one action, stated as an instruction.
COLLECT FIRST: the inputs that action needs, as a short list.
DONE WHEN: the stop condition.
DO NOT SKIP TO: the stage I am tempted by, and what breaks.
What I sell and to whom: [ONE SENTENCE ON WHAT YOU SELL AND TO WHOM]
What I have right now: [DESCRIBE YOUR LIST, YOUR PAGES AND YOUR DATA, HONESTLY]
Real search data I can get: [SEARCH CONSOLE, KEYWORD PLANNER, A PAID TOOL, OR "none"]What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[ONE SENTENCE ON WHAT YOU SELL AND TO WHOM]- The buyer, not just the product. "We sell scheduling software to construction subcontractors" places you differently from "we sell software", because a business with a named buyer can start at expansion while one without a named buyer has an earlier problem than keyword research.
[DESCRIBE YOUR LIST, YOUR PAGES AND YOUR DATA, HONESTLY]- The input the whole diagnosis rests on, and the one people flatter. "A spreadsheet of about 80 keywords I typed from memory" and "an 80 row export with volume from Keyword Planner" are stages one and three, and describing the first as the second sends you to clustering with nothing to cluster.
[SEARCH CONSOLE, KEYWORD PLANNER, A PAID TOOL, OR "none"]- What you can actually reach today, not what you could buy. Writing "none" gets you a different and more honest answer, because the route through validation without a paid tool runs through Keyword Planner and your own Search Console rather than through a subscription you have not bought.
How to run it
- 01Write down what you have before you read any prompt
One paragraph, plainly. How many keywords, where they came from, whether any figure in the sheet came from a tool, and how many relevant pages are live. Written honestly this takes two minutes and it is the difference between running the right prompt and running a good one at the wrong time.
- 02Run the router and accept the earlier stage
Paste the prompt with your three inputs. When it places you a stage earlier than you expected, that is the prompt working rather than failing. Rule 1 exists because the common error is believing you have a keyword list when you have a list of guesses.
- 03Open only the prompt for that stage
Go to the linked prompt for the stage you were given and run it end to end before opening another. Running expansion and clustering in the same session produces clusters built from unvalidated candidates, which looks like a plan and is a list of assumptions with headings.
- 04Do the validation stage in a real tool
Stage two has no prompt on this site and never will, because it is the one stage a language model cannot do. Take your candidates into Keyword Planner, Search Console or a paid tool and get figures somebody measured. Everything downstream inherits whatever you accept here.
- 05Rerun the router when the stop condition is met
The DONE WHEN line is the thing to check against, not your sense that the stage went well. When it is true, run the router again with an updated description of what you have. It will move you on, and occasionally it will move you back, which is also information.
Questions people ask
What are the best ChatGPT prompts for keyword research?
The ones matched to the stage you are at, because a keyword research prompt run at the wrong point produces confident output nobody can use. Expansion generates candidates, clustering decides which share a page, targeting picks one query for one URL, and triage handles positions you already hold. Running clustering on unvalidated candidates is the most common way this goes wrong.
Can ChatGPT do keyword research on its own?
It can do four of the five stages and not the one that matters most. Language models have no access to search volume, competition or trend data and will produce plausible figures if asked, so validation has to happen in a tool with real data. The model is genuinely good at language, intent and grouping, which is the rest of the work.
Do I have to run all five stages?
Only from wherever you are. A site with three years of Search Console data starts at triage and may never need expansion. A brand new business with no pages starts at expansion and cannot start anywhere else. The order matters within a run, not as a checklist to complete once.
Which single prompt should I run if I only run one?
Clustering, if you already have a keyword export sitting unopened, because it turns a list into a build order. Triage, if you have pages ranking in Search Console, because it finds work on positions you have already paid for. If neither is true you do not have a keyword problem yet, you have a research problem.
Why is there no prompt for search volume?
Because there is no honest one. A model asked for volume produces numbers that look exactly like real ones and came from nowhere, and a plan built on invented demand is worse than a plan built on none. Every prompt in this category forbids the model from touching demand for that reason.
Unsubscribe in one click. We never pass your address on.