ChatGPT for keyword research feels like cheating until you read the output. Most of the keywords are made up, the volumes are pure invention, and the suggestions are 80% obvious. There is a way to get genuinely useful keyword research out of ChatGPT, but it needs a structured workflow and a filter. This guide shows the 4-step framework, ten production-ready prompts, and a worked example for a Birmingham accountant.
Why ChatGPT keyword research fails for most people
Ask ChatGPT for "the best keywords for my plumbing business" and you will get 30 keywords back. About six of them are useful. The rest are either keywords with zero search volume, keywords your competitors have locked, or completely invented terms that nobody searches. This is not ChatGPT being bad at SEO. It is ChatGPT doing exactly what it does: pattern-matching plausible words. Search volume requires actual SERP data, which ChatGPT does not have natively.
The teams that get value from ChatGPT for keyword research treat it as an ideation engine, not a data source. ChatGPT is excellent at expanding seed keywords into hundreds of variations, identifying intent and topical clusters, and writing long-tail questions. It is terrible at telling you which of those are worth pursuing. You bring the volume data and the judgment. ChatGPT brings the breadth.
The 4-step framework
- Seed — give ChatGPT 5–10 seed keywords you already know matter to your business.
- Expand — use prompts to multiply each seed into clusters, long-tails, questions, and intent variations.
- Filter — manually or with a separate tool, remove keywords with zero search volume, obvious hallucinations and intent mismatches.
- Validate — for the remaining 50–100 keywords, run them through Ahrefs, Semrush or Google Keyword Planner for actual volume and difficulty.
Without steps 3 and 4, you have a list of 500 keywords that look impressive in a slide deck but rank for nothing. With them, you have 30 high-intent keywords that actually map to revenue.
10 production-ready ChatGPT prompts for keyword research
Copy these verbatim. Replace the placeholders in [brackets] with your own data.
Prompt 1 — Seed expansion
"I run a [type of business] in [city]. My main service is [service]. List 30 keyword variations a customer might use to find this service. Group them by buying intent: informational, commercial-investigation, transactional."
Prompt 2 — Long-tail mining
"For the keyword [primary keyword], list 20 long-tail variations of 4-7 words each. Include question-format variations. Prioritise variations that suggest the searcher is ready to buy."
Prompt 3 — Question harvesting (PAA-style)
"List 25 questions a [target customer] might type into Google about [topic]. Cover: how-to, definition, comparison, troubleshooting, and decision-making questions."
Prompt 4 — Topical cluster mapping
"For the topic [topic], build a pillar-and-cluster keyword map. One pillar keyword, then 6-8 cluster topics, then 3-5 supporting keywords per cluster. Output as a markdown table."
Prompt 5 — Competitor angle gap
"Here are the top 3 ranking pages for [keyword]: [URL1], [URL2], [URL3]. Based on titles and meta descriptions, identify 5 content angles or sub-topics that none of them are covering. Frame each as a potential keyword."
Prompt 6 — Local keyword localisation
"Take this keyword: [keyword]. Generate 20 location-modified variations for UK cities. Mix [primary city], [city] + service, and 'near me' variations."
Prompt 7 — Intent classification
"Here are 30 keywords: [paste list]. For each one, classify the search intent as: I (informational), C (commercial-investigation), T (transactional), or N (navigational). Output as a markdown table with one row per keyword."
Prompt 8 — Negative keyword discovery
"I sell [premium service for B2B clients]. What 30 keywords might surface my ads or content but bring the wrong audience (e.g. DIY, hobbyists, students, free-seekers)? These will become my negative keyword list."
Prompt 9 — Featured snippet target identification
"Of these 20 keywords [paste list], which ones have a high probability of triggering a Google featured snippet? Look for question-format, list-format and definition-format intent."
Prompt 10 — Content brief sketch
"For the keyword [keyword], propose a content brief covering: working title, H1, 6-8 H2 sections, 3 unique angles competitors miss, and 5 internal-link suggestions. Word count: [target]."
When ChatGPT beats Ahrefs (and when it absolutely doesn't)
ChatGPT wins for:
- Brainstorming long-tail variations Ahrefs misses because their volume is too low to index
- Generating PAA-style questions for FAQ sections
- Mapping topical clusters quickly when starting a new content programme
- Writing dozens of variations on the same intent for ad-copy testing
- Classifying intent across a list of 100+ keywords in 30 seconds
Ahrefs (or Semrush) wins for:
- Actual search volume, KD, CPC and SERP analysis
- Identifying which keywords your competitors rank for that you do not
- Tracking keyword rankings over time
- Backlink data and authority signals
- Anything where being right matters more than being fast
The right workflow uses both. ChatGPT for the wide-net ideation, Ahrefs (or another paid tool) for the narrow-funnel validation. Eight other AI workflows that save serious time in marketing walks through what we automate at agency scale.
Cleaning the list: filtering ChatGPT's keyword hallucinations
After expansion, your raw list will be 200-500 keywords. Filter in this order:
- Run the list through Google's Keyword Planner or Ahrefs Keywords Explorer in bulk. Drop anything with zero monthly volume. This removes most hallucinations in one pass.
- Drop near-duplicates. "best plumber in birmingham" and "best birmingham plumber" target the same intent. Keep one.
- Remove off-intent keywords. If you sell B2B services and a keyword screams DIY consumer, cut it.
- Tag remaining keywords by funnel stage. Awareness, consideration, decision. This guides the content type.
- Re-cluster the survivors into pillar topics. You will end with 30-80 high-quality keywords mapped to a content roadmap.
Worked example: keyword research for a Birmingham accountant
Seeds: accountant birmingham, small business accountant, tax return help.
After Prompts 1, 2, 3 and 6, we get 137 raw keywords. After filtering through volume data, we are left with 42 keywords that have real UK search volume. Top picks:
- "small business accountant birmingham" — 320 vol, KD 22, transactional
- "accountant for limited company birmingham" — 90 vol, KD 18, transactional
- "how much does an accountant cost uk" — 1,400 vol, KD 32, commercial-investigation
- "do i need an accountant for self assessment" — 720 vol, KD 24, informational
- "birmingham bookkeeping services" — 210 vol, KD 19, transactional
The pillar piece writes itself: "How Much Does an Accountant Cost in the UK in 2026?" — a definitive, well-researched answer to the 1,400-volume informational query, with internal links to the Birmingham-specific transactional pages. This is exactly the topical authority approach that lets smaller sites outrank bigger competitors with less content.
Want to see whether your own site has the on-page foundations to rank for the keywords you find? Our free 10-point SEO audit tool gives an instant scorecard with the top 3 fixes.
Frequently asked questions
Can I use ChatGPT instead of paying for Ahrefs?
Not for keyword volume or SERP data. ChatGPT does not have search data. You can replace some Ahrefs use cases (ideation, intent classification, brief generation), but you will still need a real keyword tool for volumes. Google Keyword Planner is free and covers the basics.
How do I stop ChatGPT inventing fake keywords?
You cannot stop it entirely. The framework above filters them out at step 3. If you want to reduce hallucinations at source, give ChatGPT real data to work from: paste a Google autocomplete list, an Ahrefs export, or 10 competitor URLs as context. It will then generate variations on real keywords rather than inventing new ones.
Which is better for keyword research, ChatGPT or Claude?
Claude tends to produce better intent classification and cluster mapping. ChatGPT tends to produce more raw variations. For pure ideation volume, ChatGPT. For thoughtful clustering, Claude. Try both for the same task once and pick.
Does ChatGPT see real-time search data with web browsing turned on?
Yes, but with significant lag. ChatGPT's browsing returns titles and snippets, not search volumes. It is useful for fast competitor research but not a replacement for keyword tools.
How long should the full keyword research process take for one site?
For a UK small business: 2-3 hours including filtering. For an established business in a competitive niche: 6-10 hours over two sessions to cover the full topical map. Anything less and the analysis is too shallow to rank.
Where to go next
If you are running a small marketing team and ChatGPT-led keyword research still leaves you short on capacity, our AI automation services build custom keyword research workflows tailored to your business. Or and we will audit your current keyword set and identify the 10 highest-impact terms to target first.
