ChatGPT is genuinely useful for keyword research in exactly one way: it widens your thinking. It knows how people phrase problems, and it never gets tired of generating angles.
It is also confidently wrong in one specific way: numbers. This guide shows how to use the first part and defuse the second.
What ChatGPT knows and what it invents
The line is simple once you see it. Language and topics: real knowledge. Search data: invented on the spot.
Ask it for volumes and it will answer fluently, with specific numbers, and they are fabricated. Not rounded, not outdated. Made up.
The prompts that actually work
Generic prompts get generic keywords. The wins come from persona, format and constraint.
| Job | Prompt pattern | Why it works |
|---|---|---|
| Widen topics | "List 30 subtopics of [niche] a beginner asks about, grouped" | Coverage you did not think of |
| Real phrasing | "Act as [persona]. Write 20 questions you would type into Google about [topic]" | Personas surface real language |
| Long-tail angles | "Give 15 comparisons and 15 problems people have with [topic]" | Maps vs and how-to keywords |
| Grouping | "Cluster this keyword list into topics, one page per cluster" + paste | Fast first-pass clustering |
| Intent check | "For each keyword, label the likely intent and page type" | Ten-minute job done in one |
The workflow: widen, then verify
Treat ChatGPT as the top of your funnel, never the bottom. The safe loop takes three steps.
Step 2 is where the keyword tool on this site fits: paste ChatGPT's best candidates as seeds and see which ones autocomplete actually recognizes. If Google never suggests it, nobody types it.
Step 3 is Google Keyword Planner or any tool with real volume data. Our free tools roundup lists the options.
Where it falls apart in practice
The confident volume trap. Ask for "keywords with search volumes" and you get a tidy fabricated table. If a number did not come from a real tool, it does not exist.
The consensus trap. ChatGPT suggests what is commonly written about, which skews toward crowded topics. Autocomplete and People Also Ask surface what is actually searched, including the underserved corners.
The sameness trap. Everyone prompts "keyword ideas for [niche]", so everyone gets similar lists. Persona prompts and your own customer language are the difference.
A worked example
For a cold brew gear site: ChatGPT's persona prompt produced "can I make cold brew without a special maker", "why is my cold brew bitter" and "cold brew ratio for a half gallon".
Autocomplete confirmed all three phrasings exist. Keyword Planner showed the ratio query with real volume, and the bitter question small but nonzero.
Two pages got built. The third idea, "cold brew flavor wheel", autocomplete had never heard of, and it died there instead of becoming a page nobody visits.
The one-line takeaway: ChatGPT is a brilliant brainstorming partner and a habitual liar about data. Widen with it, verify phrasing in autocomplete, and let real volume numbers make the final call.