Work out which half of your keyword question a model can actually answer
What a language model is genuinely good at in keyword research, what it cannot know, and a prompt that scores your request before it tries to answer it.
- Works in
- ChatGPT, Claude, Gemini
- You need
- The keyword request you were about to make · Whatever real search data you can reach
- Written for
- how to use chatgpt for seo keyword research
Scored by our own engine
This page, run through the audit we sell. Measured 4 August 2026.
Every other page in this category hands you an artefact to copy. This one is about the method around them: which parts of a keyword question a language model can answer, which parts it will answer anyway, and how to tell the two apart before you build anything on the output. The prompt above does that audit on your own request rather than on a general example.
The line that runs through all of it
A keyword tool can only report demand somebody already measured. A language model knows how people write and nothing about how many of them write it.
That is not a shortcoming to prompt around, it is the shape of the instrument. Every prompt in this library that works respects it, and every AI keyword workflow that produces nonsense crosses it, usually at the moment somebody asks for the list to be sorted by volume. The model has no volume, so it produces some, and the numbers are indistinguishable from real ones.
Four labels, and why the middle two matter most
LANGUAGE and DATA are the easy ends. The interesting scores are JUDGEMENT, where the model is reliable only because you supplied the evidence, and UNKNOWABLE, where nobody can answer and a tool would be no better.
JUDGEMENT is where most of the real value sits. Grouping a list by intent, spotting the two pages of yours that would compete, ranking your terms by how closely they match what you sell: all of these are strong, and all of them collapse if the inputs are thin. The confidence figure exists to make that visible, because a model reasoning about a business it knows one line about will still answer in full sentences.
Running the method
Audit the request, collect the data the audit names, then run the narrowed prompt. It is slower than asking for a keyword list and it is the difference between research and generated text.
Once you know which stage you are at, the artefacts do the work: expansion for phrasings, clustering for a page plan, one page one query for a single decision, and triage once pages are ranking. The full sequence is here, and if the underlying question is what keyword research is at all, start with the definition.
You are auditing a request before answering it. I am going to paste a
keyword research task I was about to ask an assistant to do. Do not do the
task. Break it into subtasks and score each one on whether you could do it
honestly.
Score every subtask against this rubric and use these four labels exactly:
LANGUAGE. Answerable from how people write and speak. Phrasings, synonyms,
the words a person uses before they know your product exists, what a query
implies about the person typing it. You are strong here and this is the
reason to involve a model at all.
JUDGEMENT. Answerable only from data I supply, and then reliable. Grouping
my list, spotting two of my pages that would compete, ranking my terms by
fit to what I sell. Say which of my inputs it needs.
DATA. Not answerable by you at all. Search volume, difficulty, competition,
cost per click, trend direction, what currently ranks, whether a SERP has
changed. Name the specific report or tool that answers it instead. Do not
soften this with "approximately" or "typically".
UNKNOWABLE. Not answerable by anyone in advance, including a tool. Whether
a page will rank, how long it will take, what a competitor will do next.
Then give each subtask a confidence figure from 0 to 3, where 3 means you
would stand behind the answer without qualification and 0 means you would
be guessing, plus the one input that would raise it.
Rules you must follow:
1. Do not answer any part of my request, including the parts scored
LANGUAGE. This pass is classification only.
2. Every subtask scored DATA must name a real source: Google Search Console,
Google Keyword Planner, Google Trends, a named paid tool, or a manual
search. "A keyword tool" is not a source.
3. If my request contains no LANGUAGE subtask at all, say so and tell me an
assistant is the wrong instrument for this entire request.
4. Do not invent a subtask I did not ask for in order to fill the rubric.
5. Finish by rewriting my request as a prompt containing only the LANGUAGE
and JUDGEMENT parts, with the DATA parts moved into a short "get these
first" list above it. The rewrite must not mention any figure I have not
given you.
Output as a scored rubric and not as a table:
For each subtask, four lines: the subtask in my own words, the label, the
confidence figure out of 3, and the one input that would raise it.
Then two headed blocks: GET THESE FIRST, and THE PROMPT THAT REMAINS.
What I sell and to whom: [ONE SENTENCE ON WHAT YOU SELL AND TO WHOM]
Real data I can reach today: [SEARCH CONSOLE, KEYWORD PLANNER, A PAID TOOL, OR "none"]
The request I was about to make: [PASTE YOUR KEYWORD QUESTION, WORD FOR WORD]What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[ONE SENTENCE ON WHAT YOU SELL AND TO WHOM]- This decides which subtasks are judgement rather than language. Grouping terms by fit to what you sell is only reliable when the model knows what you sell, so a request scored against "we sell software" comes back with lower confidence figures than the same request scored against a named buyer.
[SEARCH CONSOLE, KEYWORD PLANNER, A PAID TOOL, OR "none"]- What you can reach today, which changes the GET THESE FIRST list from advice into a route. Answering "none" produces a genuinely different list, built around Keyword Planner and manual searching rather than around a subscription, and that list is the honest one for most small sites.
[PASTE YOUR KEYWORD QUESTION, WORD FOR WORD]- Paste what you actually typed, including the vague parts. Tidying the request before pasting removes the exact ambiguity the rubric is there to catch, and "find me the best keywords for my business" scoring badly across four subtasks is the most useful result this prompt produces.
How to run it
- 01Write the request you were about to send
Open a new note and type the keyword question in your own words, as you would have asked it. Do not clean it up. The value of this pass is in seeing which parts of your own phrasing were quietly asking for data, and a polished request hides them.
- 02Run the audit and read the DATA rows first
Paste the prompt with your three inputs. The DATA rows are the parts you would otherwise have received as invented numbers, delivered in the same confident tone as the parts the model can do. Every one of them names where the real answer lives instead.
- 03Collect the GET THESE FIRST list
Go and get those figures before running anything else. This is the slow step and it is the only one that puts evidence into the process. Search Console for terms you already appear for, Keyword Planner for terms you do not, and a manual search for what the results actually look like.
- 04Run the rewritten prompt with your data pasted in
The rewrite contains only the parts a model can answer, so run it with your collected figures included. The output will be narrower than what you originally asked for, and it will be made of things that are true, which is the trade this whole method is.
- 05Sanity check one conclusion against a live search
Take one conclusion and search the query in an incognito window. A model cannot see a search result page, so any statement about what ranks or what a searcher finds is inference. Two minutes of looking settles what an hour of prompting cannot.
- 06Check the page you write can actually be quoted
Once the research turns into a page, run it through the AI content readiness check. Keyword research decides what a page is about, and structure decides whether an assistant can lift an answer out of it. The two failures are independent and a page can pass one while failing the other.
Questions people ask
How do you use ChatGPT for SEO keyword research properly?
Use it for language and judgement, and use a real data source for demand. It is strong at generating the phrasings people use before they know your product exists, at grouping a list by intent, and at spotting two of your pages that would compete. It has no access to search volume, difficulty or what currently ranks.
Does ChatGPT know search volume?
No, and it will produce numbers anyway if you ask. Language models have no connection to search data and generate plausible figures rather than admitting the gap, which is the single most common failure in AI keyword research. Get volume from Google Keyword Planner, from Search Console for terms you already appear for, or from a paid tool.
Can ChatGPT browse Google to check what ranks?
Some assistants can fetch pages when browsing is enabled, and what comes back is not what a searcher sees. Results are personalised, localised and increasingly assembled per session, so a fetched result page is one sample from one context. For anything that depends on the live SERP, look yourself in an incognito window.
What is a language model actually better at than a keyword tool?
Phrasings nobody has searched often enough to be recorded. A tool can only show you demand it has already measured, so the way a customer describes a problem before they know the product category exists is invisible to it. That gap is the one place a model contributes something a tool cannot.
Is AI keyword research reliable enough to build a content plan on?
The language and grouping parts are, once you have supplied real figures. The demand parts are not, at any level of prompting, because the limitation is architectural rather than a matter of the model trying harder. A plan built on generated volume looks identical to one built on measured volume and behaves nothing like it.
Unsubscribe in one click. We never pass your address on.