Keyword Research Brief Builder
Turn one seed topic into a prioritised keyword set with volume bands, intent and difficulty judgement.
Turns a single seed topic into a structured keyword research brief with head terms, mid-tail and long-tail groups, estimated intent, competition judgement and a recommended priority order for content production.
Ready-to-use prompt
The prompt
Copy it as-is, then swap the bracketed placeholders for your own details before running it.
Role: You are an SEO strategist building a keyword research brief.
Context:
- Seed topic: {{seed_topic}}
- Business: {{business_description}}
- Target audience: {{audience}}
- Market and language: {{market}}
- Pages that already exist: {{existing_pages}}
Task: Produce a keyword research brief in a table with these columns:
1. Keyword
2. Group (head / mid-tail / long-tail)
3. Likely search intent (informational, commercial, transactional, navigational)
4. Why this audience searches it (one clause)
5. Competition judgement (low / medium / high) with the reason
6. Recommended target page (existing URL or NEW)
7. Priority (1-3)
Rules:
- Return at least 40 keywords, with no more than 8 head terms.
- Do not invent search volume numbers. Use volume bands (high / medium / low) and say they are estimates.
- Exclude keywords the business cannot credibly rank for or serve.
- Group near-duplicates onto one row and list variants in the keyword cell.
- Flag any keyword that duplicates an existing page as CANNIBAL RISK.
- Add an Estimated volume band column; define bands relative to the supplied topic and market, not as numerical ranges.
- Define priority as 1 = first-wave publish or refresh, 2 = supporting opportunity, 3 = defer pending validation.
- Distinguish inferred competition from checked evidence; never imply that live search results were inspected unless they were.
- Map queries sharing intent to one target page unless a distinct user task justifies separation.
Finish with: the 5 keywords you would publish first and why.Estimated results
Editor's note
Why this prompt matters
Most AI keyword research fails in the same way: you ask for keywords, you get a list, and the list is a hallucinated volume table dressed up as data. This prompt does the part a language model is genuinely good at — reading a topic, understanding who searches it and why, and organising terms by intent and page fit — while explicitly refusing to invent numbers. What you get back is a research brief you can take straight into a volume tool for validation, with the strategic work already done: what belongs on one page, what deserves its own, and what your business has no business chasing.
Anatomy
Prompt engineering breakdown
Role
Context
Goal
Constraints
Output format
What you'll get
Expected output
Scenario and assumptions
A generic UK software business serves independent bicycle retailers. Its product tracks stock, flags reorder needs and imports supplier catalogues; it does not handle payroll or workshop bookings. The audience is shop owners and stock managers searching in English. Existing pages include /inventory-software/ and /guides/stocktake/.
This review excerpt shows eight representative rows, not the full required set of at least 40. Volume bands are unverified estimates relative to this niche and market, not measured demand. Competition labels are provisional judgements based on query breadth and likely competing page types; no live results were checked.
Keyword brief excerpt
| Keyword | Group | Intent | Audience reason | Competition judgement | Target page | Priority | Estimated volume | |---|---|---|---|---|---|---|---| | bike shop software | Head | Commercial | Compare operating systems | High: broader software suites compete | /inventory-software/ — CANNIBAL RISK | 2 | High | | bike shop inventory software; bicycle shop stock software | Mid-tail | Commercial | Find specialist stock tools | Medium: specialist vendors likely compete | /inventory-software/ — CANNIBAL RISK | 1 | Medium | | bicycle shop stocktake checklist | Long-tail | Informational | Prepare a reliable count | Low: narrow procedural need | /guides/stocktake/ — CANNIBAL RISK | 1 | Low | | bike shop reorder point calculation | Long-tail | Informational | Avoid replenishing too late | Medium: generic inventory guides compete | NEW: reorder guide | 1 | Low | | bike shop inventory spreadsheet | Mid-tail | Informational | Organise stock cheaply | Medium: downloadable templates compete | NEW: spreadsheet resource | 1 | Medium | | slow moving stock in bike shops | Long-tail | Informational | Identify cash tied up | Low: specific retail problem | NEW: ageing-stock guide | 1 | Low | | bike shop inventory software pricing | Long-tail | Commercial | Check affordability | Medium: vendor pricing pages compete | /inventory-software/ — CANNIBAL RISK | 2 | Low | | bike shop supplier catalogue import | Long-tail | Commercial | Reduce manual data entry | Low: narrow feature requirement | NEW: import feature page | 2 | Low |
Production judgement
Refresh the existing software page rather than creating separate pages for its three overlapping query groups. Confirm that it explains bicycle-retail workflows before targeting the broader software term. That term may imply capabilities this product lacks, so it is not a first-wave target.
Treat “CANNIBAL RISK” as a mapping warning, not proof that two pages currently compete. The stocktake keyword belongs on the existing guide. Exclude workshop-booking and payroll queries entirely. Validate spreadsheet demand before commissioning a download; publish only if someone can maintain a genuinely usable file.
First five keywords to publish or refresh
- Bike shop inventory software: refresh the closest commercial page first.
- Bicycle shop stocktake checklist: improve the existing guide with a printable counting sequence.
- Bike shop reorder point calculation: demonstrate a supported replenishment workflow.
- Slow moving stock in bike shops: address a distinct stock-management problem.
- Bike shop inventory spreadsheet: test an entry-level resource once demand and maintenance ownership are confirmed.
Under the hood
Why this prompt works
The prompt separates the two halves of keyword research that people usually blur together. Structure and intent are language problems, which models handle well; volume and difficulty are data problems, which they fabricate. By banning numeric volume claims and forcing band estimates instead, the output stays honest. Passing in existing URLs turns the exercise into a gap analysis rather than a blank-page list, and the cannibalisation flag catches the most common outcome of AI-generated keyword lists: three new pages competing with one you already have.
Model fit
Best AI models for this prompt
ChatGPT
Holds a seven-column table across 40+ rows without drifting. The default choice for volume of output.
Claude
The best judgement layer. It will tell you a keyword is not worth targeting, which the others rarely do.
Gemini
Strongest when you paste competitor page copy alongside the seed topic and ask it to find the gaps.
When to use
- Before opening a new content area, to test whether one seed contains several serviceable search intents.
- Between discovery and writer briefing, to assign provisional page targets and production priorities.
- During a content audit, to separate refresh opportunities from genuinely missing pages.
- Before keyword-tool research, to build a candidate list worth validating.
When not to use
- When forecasting traffic or revenue; use measured demand, ranking evidence and explicit conversion assumptions.
- When a validated keyword export already exists; analyse that dataset instead of generating replacements.
- For branded keyword tracking, which needs recurring rank and search-performance data.
- When deciding difficult page splits; inspect live results and involve an SEO reviewer before changing architecture.
Get more from it
Pro tips
- 1
Supply existing URLs with page titles and one-line purposes; a URL alone may hide an overlapping target.
- 2
Reject unsupported exact search volumes. Keep bands explicitly estimated, and ask what makes one query relatively broader than another.
- 3
Validate shortlisted terms in a keyword tool using the same country setting. Replace estimated bands with sourced evidence before committing production budget.
- 4
Specify country and language in {{market}}, including preferred local terminology; do not merge translated queries into one demand estimate.
- 5
Describe unsupported products, services and audiences in {{business_description}} so attractive but unserviceable queries are excluded.
- 6
On a second pass, challenge every NEW recommendation against {{existing_pages}}. Ask which intent requires a separate page rather than another section.
- 7
Audit near-duplicate rows before accepting the 40-keyword minimum. Request additional audience problems, not pluralisations or reordered phrases, when coverage is thin.
Don't ship this
Common mistakes
✗ Accepting the volume estimates as data.
Fix — Treat bands as hypotheses and validate every priority-1 keyword in a real tool.
✗ Omitting existing pages from the context.
Fix — Paste your sitemap URLs so the model can flag cannibalisation.
✗ Chasing every head term in the output.
Fix — Ship the long-tail rows first — they are the ones a new site can actually win.
People also ask
Frequently asked questions
Q.Can this prompt give me real search volumes?
No, and it is instructed not to try. Language models fabricate volume figures that look plausible. The prompt returns volume bands and intent judgement instead, which you then validate in a real keyword tool before committing to a content plan.
Q.How many keywords should I expect back?
At least 40, with no more than 8 head terms. The bulk of the value is in the mid-tail and long-tail rows, which are the ones a newer site can realistically rank for within a few months.
Q.What is the cannibalisation flag for?
It marks keywords that overlap with a page you already have. Acting on those rows first is usually cheaper than writing new content, because you are improving a page that already has some ranking history.