Claude can write a list of 100 or more long-tail keyword ideas from one prompt. What it cannot do is tell you how many people search for them. A language model has no live search data, so a list is a set of guesses until you check it.
This guide gives you a prompt you can copy, shows what that prompt returned in our own run, and then shows how to check the list with three free or low-cost sources: Google autocomplete, Search Console and Keyword Planner.
From our own run, 9 November 2026. We gave Claude the prompt below with the seed “technical seo audit” and the audience “small business owners”. It returned 111 phrases, all different, each 5 to 10 words long. We then asked Google’s public autocomplete about every one. Only 4 of the 111 were offered back exactly as written; 77 of our test queries returned no suggestions at all.
So treat the list as ideas to test, not as proof of demand. The rest of this guide shows how to test them.
Table of Contents
ToggleWhich prompt should you use?
Paste this into a new Claude conversation, so older messages do not colour the answer. Replace the two bracketed parts. The first is your seed topic. The second is the person who is searching, such as “small business owners” or “WordPress beginners”. Giving Claude an audience stops it from drifting into general phrases.
You are an SEO researcher. Seed topic: [TOPIC] Audience: [WHO SEARCHES, for example small business owners] Return 100 or more long-tail search phrases that this audience could type into Google about the topic. Rules: 1. Each phrase has 5 to 10 words. 2. No duplicates, and no near-duplicates that only swap one word. 3. Cover these types: how to, why, which or what, for a specific audience or platform, checklists and templates, cost and time. 4. One phrase per line, lower case, no numbering, no explanations. 5. Do not invent search volumes and do not say any phrase is popular.
Two rules in the prompt matter more than the rest. Rule 5 stops Claude from making up search volumes, which it would otherwise do with confidence. Rule 3 forces a spread of question types, so you do not get 100 versions of the same “how to” phrase.
What did the prompt return?
For the seed “technical seo audit” and the audience “small business owners”, Claude returned 111 phrases. We counted them ourselves.
- All 111 were different from each other.
- Length ran from 5 to 10 words: 11 phrases had 5 words, 32 had 6, 24 had 7, 20 had 8, 18 had 9 and 6 had 10.
- Most were questions: 68 began with “how to”, “how long” or “how often”.
| Type | Phrases | Example |
|---|---|---|
| How to, how long, how often | 68 | how to find orphan pages without paid tools |
| Audience or platform | 20 | technical seo audit for shopify stores |
| Checklist, template, cost, tools | 14 | technical seo audit template google sheets |
| Which, what or versus | 7 | screaming frog vs sitebulb for technical audits |
| Why | 2 | why is google not indexing my new pages |
The spread is lopsided. We asked for “why” phrases and got two. If a type matters to you, add a count to rule 3, for example “at least 15 why phrases”, and run it again. The prompt follows what you ask for, not what you hope for.
Were the phrases real searches?
Google says autocomplete predictions reflect real searches done on Google, and that its systems look for common queries that match what someone starts to type [1]. That makes autocomplete a quick, free signal. We used it as a test.
Method. For each of the 111 phrases we sent Google’s public autocomplete endpoint the phrase without its last word, with the language set to English and the country to the UK. We then looked for the full phrase in the list that came back.
Result. Four phrases came back exactly as Claude wrote them:
- how to fix duplicate title tags in wordpress
- how to fix discovered currently not indexed
- how to improve interaction to next paint
- how to check server response time
For 34 of the 111 test queries autocomplete returned some suggestions, and for 77 it returned nothing. That is normal for long phrases: autocomplete shows common queries, and a specific eight-word question is rarely common enough.
What it does and does not tell you. A phrase that appears is a sign that people search for it. A phrase that does not appear is not proof that nobody does. It only means it is not common enough to be suggested. Autocomplete also varies with location, trending interest and, for signed-in users, past searches [1], so your results may differ from ours.
We did not measure search volume for any phrase. We have no volume data for this list, and we do not claim any.
How do you clean the list?
Before you check anything, remove the noise. Long lists from any AI tool tend to carry numbering, repeated phrases and phrases that only reorder the same words. Paste your list here to clean it. It runs in your browser and sends nothing anywhere.
Paste a keyword list from Claude. The tool removes numbering, exact duplicates and phrases that only reorder the same words, sets aside phrases under 4 words, and counts what is left by type.
What is the whole process?
The steps below are the order we would follow. You can zoom and pan the picture, or expand it. The same steps are written out under the graphic.
Read the workflow as text
- 1. Pick a seed and an audience One topic and one reader. Example: technical seo audit, for small business owners.
- 2. Run the prompt Ask for 100+ phrases of 5 to 10 words, grouped, with no duplicates and no invented volumes.
- 3. Clean the list Remove duplicates and near-duplicates. Drop phrases under 4 words. The cleaner below does this.
- 4. Check demand Autocomplete first, then Search Console for queries you already show for, then Keyword Planner estimates.
- 5. Give each phrase one URL Merge phrases with the same intent into one page. Do not write one thin page per phrase.
- 6. Publish, then re-check Look at the Performance report after a few weeks and keep the phrases that earn impressions.
How do you check demand properly?
Use the three sources in this order, from quickest to slowest. Each answers a different question.
1. Google autocomplete: do people type this?
Type the phrase, minus its last word or two, into Google and see whether the full phrase is offered. It costs nothing and takes seconds. As we saw above, it will miss most long phrases, so use it to confirm winners, not to reject the rest.
2. Search Console: do you already show for it?
The Performance report shows which search queries bring impressions and clicks to your site. By default it shows the past three months, and you can change the metrics, dimensions and dates [2]. It only covers queries your site already appears for, so it is best for checking a new list against what you rank for today. If one of Claude’s phrases already earns impressions, write or improve a page for it first.
3. Keyword Planner: roughly how many searches?
Google describes Keyword Planner as a way to discover keywords and view estimates of the searches they receive [3]. Its help page says you need to enter billing information and complete your account setup to use the keyword suggestion features. Treat the numbers as estimates, not counts.
| Source | Answers | Cost | Weak spot |
|---|---|---|---|
| Autocomplete | Do people type this? | Free | Misses uncommon phrases |
| Search Console | Do we already show for it? | Free | Only your own queries |
| Keyword Planner | About how many searches? | Account setup needed | Estimates, not exact |
What do you do with the phrases that pass?
Do not write one page per phrase. Many of the phrases in our list mean the same thing. “how to find orphan pages without paid tools” and “how to find pages with no internal links” can live on one page. Group phrases by what the searcher wants to do, then give each group one URL and one main phrase (our keyword mapping guide shows how).
Ask Claude to do the first pass with a follow-up prompt, then read its grouping yourself:
Group the phrases above by what the searcher wants to do. Give each group a short name and list its phrases. Keep the groups to ten or fewer. Do not add new phrases.
Which prompt variations are worth trying?
Change one thing at a time so you can see what changed. Each of these edits goes into the same prompt.
- Local intent. Add “Every phrase includes a place name or the words near me” to rule 3, and name the city in the seed.
- Questions only. Replace rule 3 with “Every phrase is a full question.”
- Competitor gap. Paste the headings of a competitor page and ask for phrases it does not cover. Check the answer against the page yourself.
- Another language. Ask for the phrases in the language of your market. Check a few with a native speaker, because direct translations are often not how people search.
What mistakes should you avoid?
- Treating the list as data. It is a list of ideas. Check it with the sources above before you plan content around it.
- Asking for volumes or difficulty. A model that has no search data will still give you a number. Rule 5 in the prompt blocks that.
- Writing a thin page per phrase. Many phrases share one intent. One strong page beats ten near-copies.
- Skipping the audience. Without it, the phrases drift towards general topics that big sites already cover.
- Leaving the placeholders in. If you paste the prompt with [TOPIC] still in it, you will get a list about the word TOPIC.
A note on the earlier version of this guide. It quoted several statistics and time-saving figures. When we rebuilt it in November 2026 we could not trace those figures to a source page we could open, so we removed them. Everything numeric above comes from our own run or from the Google help pages listed below.
Which guides should you read next?
These guides on this site pick up where the list ends.
- Keyword mapping: how to give every keyword one URL explains how to assign the phrases that pass your checks to pages without overlap.
- Cluster internal linking shows how to link the pages that come out of a keyword group.
- Internal link audit with Search Console, Screaming Frog and Ahrefs covers the Search Console side in more depth.
- Featured snippet testing and tracking helps with question-style phrases.
- Voice search analytics covers spoken, conversational queries.
Common questions
Can Claude tell me how many people search for a phrase?
No. A language model has no live search data. Use Search Console for queries your site already shows for, and Keyword Planner for estimates. The prompt in this guide tells Claude not to invent numbers.
Is a long-tail keyword always better than a short one?
Not always. Long phrases are more specific, which usually means fewer searches. Whether they bring better visitors depends on your site and your audience, so test with your own Search Console data.
How many phrases should one page target?
One main phrase and the closely related ones that share its intent. If two phrases need different answers, they need different pages.
Will other AI tools give the same result?
The prompt is plain text, so you can paste it anywhere. We only ran it in Claude, so we cannot say how other tools compare.
Sources
All three pages were opened on 9 November 2026. Our own run used Claude and Google’s public autocomplete endpoint on the same date.







