Mine long tail keywords a low authority site can actually win
A prompt that expands a seed into long tail queries, then throws out every one whose results page is already owned by a brand you cannot outrank yet.
- Works in
- ChatGPT, Claude
- You need
- One seed topic · Your domain and roughly how old it is · Any keyword export you already have, optional
- Written for
- long tail seo keywords
Scored by our own engine
This page, run through the audit we sell. Measured 6 August 2026.
Long tail keyword lists are easy to generate and mostly useless, because the part that decides whether a page is worth writing is missing from them. The prompt below expands a seed into specific queries, judges who currently answers each one, and then removes the ones you cannot win yet rather than handing you two hundred rows and leaving the triage to you.
The expansion is not the hard part
Any model will turn “cast iron pan” into fifty longer strings. Most of them are the same query with a stop word moved, and a good number of them are questions nobody asks. That output looks like research and functions as noise.
What turns it into a plan is the third column: who answers this now. A query held by a forum thread is an opening, because a forum thread is a partial answer with no structure and no maintenance behind it. A query held by three national retailers is not an opening for a site that launched in March. Both look identical in a keyword tool, and they lead to completely different quarters.
The rule people delete, and should not
Rule 6 flags queries whose answer is one sentence. It is the rule that gets cut, because a list is shorter without it and a shorter list feels like a worse result.
Keep it. A query with a single fact answer is answered on the results page, and increasingly inside an assistant, without anyone visiting a site. A page built for it earns impressions, a flat click through rate and a slow suspicion that SEO does not work. Knowing which queries those are before you write is worth more than ten extra rows.
One line in the constraints that had to be added
The first version of this prompt had no instruction against asking questions. Run against Claude with a four month old site as the input, it did not produce a list at all: it came back asking what the product does differently and whether the site has any impressions yet. Both are reasonable questions and neither is what you wanted from a prompt you are about to run twenty times.
The fix is the first constraint. Make an assumption, label it, produce the output, and put the questions at the end. It costs one sentence and it is the difference between a prompt that runs and a prompt that opens a conversation.
Where the model will be wrong
The competition column comes from what the model absorbed during training, not from the live results page, which it cannot see. On a topic that moved in the last year, it will be confidently out of date. That is why the prompt makes it choose “unclear” instead of guessing, and why step three asks you to check five verdicts yourself. Five is enough. If four of five are right, use the column. If two are right, use the expansion and do the competition read yourself.
You are a search strategist working for a small site with little
authority. I will give you a seed topic. Produce long tail queries that a
site like mine has a realistic chance of ranking for, and be honest about
the ones I do not.
Rules you must follow:
1. Expand the seed into specific queries of four words or more. Specificity
comes from a qualifier a real person would add: a situation, a
constraint, a tool, a job title, a comparison, a price, a mistake, a
first attempt, a failure. Not from padding a head term with filler words.
2. For every query, state the intent in one word: learn, compare, choose,
do, fix, or buy.
3. For every query, judge who currently answers it best. Choose one of:
"big brand", "niche site", "forum or community", "mixed", or "unclear".
If you are not confident, choose unclear rather than guessing. I would
rather see unclear than a wrong verdict I act on.
4. Then cut the list. Move every query you marked "big brand" into a
separate section called Not yet, with one line on what would have to be
true for me to attempt it.
5. Group whatever survives into page sized sets. Several long tail queries
usually belong on one page, and telling me I need forty pages when I
need six is the most common way this exercise wastes a quarter.
6. Flag any query whose answer is a single fact. A query answered in one
sentence is answered in the results page itself, and a page built for it
gets impressions and no clicks. Say so rather than including it silently.
Output a markdown table with these columns:
Query | Intent | Who answers it now | Page it belongs on | Worth building
Then give me:
- The Not yet list, with the condition attached to each.
- The single page you would build first, and why that one before the others.
- Anything in my seed that suggests a topic you would have expanded but
could not, because I gave you too little about the business.
Constraints:
- Do not ask me anything before you answer. If something you need is
missing, state the assumption you are making, label it as an assumption,
and produce the list anyway. Put your questions at the end.
- Do not invent search volume, difficulty scores or traffic estimates. You
have no access to them. If I supply figures, use mine. If I supply none,
reason about intent and competition instead and say that is what you did.
- Do not produce variants that differ only by a stop word. "How to clean a
cast iron pan" and "how do you clean a cast iron pan" are one query.
- Stay inside what my business actually sells. A long tail list that
wanders into adjacent topics reads impressive and converts nobody.
My seed topic: [SEED TOPIC]
My site: [DOMAIN, ROUGHLY HOW OLD, AND WHAT IT SELLS]
Keywords I already have: [PASTE ANY EXPORT, OR WRITE "none"]What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[SEED TOPIC]- One topic, not a list. "Cast iron pan care" works. "Cookware" is too broad and produces a list of head terms with a word bolted on. If you have five seeds, run the prompt five times rather than pasting all five, because the model spreads its attention thin and the fifth seed always gets the worst output.
[DOMAIN, ROUGHLY HOW OLD, AND WHAT IT SELLS]- This is what makes rule 4 work. A site that launched two months ago and a site that has been publishing for six years should get different lists from the same seed, and the model cannot make that judgement if you do not tell it which one you are.
[PASTE ANY EXPORT, OR WRITE "none"]- Optional. Pasting what you already have stops the model handing back queries you have already covered, and it gives the grouping step something real to attach new pages to. Write "none" and you get a clean sheet, which is fine for a new site.
How to run it
- 01Pick one seed and write down what you sell
Take a single topic and one sentence on the business behind it. The sentence is not padding. Without it the model produces queries that are technically long tail and commercially irrelevant, which is the failure mode of every keyword tool that expands strings rather than intents.
- 02Run the prompt and read the Not yet list first
The Not yet list is where the useful information is. It tells you which parts of your topic are held by sites you cannot displace this year, which is the thing a volume column will never tell you and the thing that decides whether a page was worth writing.
- 03Check five verdicts against a live search
Search five of the queries yourself and look at who ranks. The model is judging competition from memory of the web, not from the results page in front of you, and it will be wrong on some of them. Five checks is enough to tell whether the whole column is trustworthy for your topic.
- 04Merge the groups down
Take the page sized sets and try to make them fewer. Two groups that a single page could satisfy should be one page, and the test is whether a reader arriving from either query would feel the page was written for them. Thin pages are what happens when this step is skipped.
- 05Write the first page and check it can be quoted
Build the page the model put first, then run it through the AI content readiness check. A long tail page that ranks on Google and cannot be lifted into an assistant answer is now half a page, and long tail queries are exactly the ones assistants answer most often.
Questions people ask
What are long tail SEO keywords?
Longer, more specific search queries that individually get very little search traffic and collectively get most of it. They are usually four words or more and carry a qualifier that narrows the intent, such as a situation, a constraint or a comparison. They matter to small sites because specificity is the one advantage a site with no authority still has.
How do I find long tail keywords without a paid tool?
Expand a seed with a model, then verify. A language model is good at producing the phrasings a person would type and bad at knowing how many people type them. Use it for the expansion, use your own eyes on the results page for the competition check, and treat any number it offers you as invented, because it is.
Can ChatGPT tell me the search volume for a long tail keyword?
No. It has no access to search volume data and will produce confident figures anyway if you ask for them. This prompt forbids it explicitly for that reason. If you need volume, pull it from a data source and paste it in, and the model will reason with your numbers instead of its own.
How many long tail keywords should one page target?
As many as one page can genuinely satisfy, which is usually between three and ten closely related phrasings. The test is whether a reader arriving from any of them finds the page was written for their question. Splitting near identical queries across separate pages is how sites end up competing with themselves.
Why does the prompt refuse to include some queries?
Two reasons, and both are worth keeping. Queries whose results are owned by large brands go to a Not yet list because building for them now spends a week for nothing, and queries answered in a single sentence get flagged because the results page answers them without a click. Both are pages that look like progress and produce none.
Unsubscribe in one click. We never pass your address on.