Keyword research with ChatGPT and Claude means using a language model to brainstorm, group and sort search phrases, then confirming that people really search for them. The model supplies ideas quickly. It cannot supply data, because it has no access to Google’s search logs.
What can an AI model do for keyword research, and what can it not?
| Good at | Bad at |
|---|---|
| Brainstorming dozens of phrases from a product description | Search volumes: any number it gives is a guess |
| Grouping keywords into topics and intents | Knowing which phrases are growing or fading |
| Rewriting a seed term the way different customers say it | Showing who ranks and how hard it is to beat them |
| Spotting questions and objections a buyer has | Telling you which country or language people search in |
| Drafting a content plan from a clean keyword list | Being the same on every run |
Models also invent plausible phrases that nobody types. That is why the last step of every workflow below is a check against real data.
Why does it matter?
Founders without a keyword tool budget get a free first draft of a topic map in 20 minutes. Experienced SEOs use the same trick to widen a list they already built with classic keyword research, and to sort thousands of phrases by intent. Used this way it saves hours. Used as the whole method, it fills your site plan with phrases that have zero searches.
How do you do keyword research with ChatGPT or Claude, step by step?
- Give the model context. Describe your product, customer, country, language and price level in 5 to 8 lines. Add 3 competitor names and the problem you solve. A vague prompt returns generic phrases.
- Ask for ideas in batches by angle. Request separate lists: problems customers describe, questions they ask, comparisons (“X vs Y”), jobs to be done, and phrases with a buying signal such as “pricing” or “near me”. Ask for the wording a customer would use, not a marketer.
- Collect real phrasing next to it. Copy questions from Google autocomplete, People also ask, support tickets, sales calls and Reddit threads. Paste them into the same sheet, so the model’s ideas sit beside real language.
- Cluster by search intent. Paste the list and ask the model to group phrases that the same page can answer, and to label each group informational, commercial, transactional or navigational. Then check the top results for 2 or 3 phrases per group yourself: if different pages rank for two phrases, they need separate pages (see search intent).
- Check demand in a real tool. Paste the list into Google Keyword Planner with “Get search volume and forecasts”. It returns average monthly searches in ranges, which is enough to separate 0 from 10 from 1,000. A free Google Ads account is required. Drop phrases with no volume, unless they are very specific questions you can answer in one section.
- Check what you already get. In Google Search Console, open Performance and sort the Queries by impressions. Queries with impressions and a position of 8 to 30 are the quickest wins, and the model can help you turn them into headings and new sections.
- Find gaps. Give the model your list of existing page titles and ask which clusters have no page. Then confirm each gap in a search result (see content gap analysis).
- Build the plan. Assign one main phrase and 3 to 6 related phrases per page, and group pages into topic clusters. Keep the spreadsheet as the source of truth, not the chat.
Which prompts work well?
Use prompts that ask for reasoning and structure, not numbers:
- “Here is my product and customer. List 40 questions a first-time buyer types into Google before choosing, grouped by stage of the decision.”
- “Take these 80 keywords. Group them into clusters that one page can answer. Name each cluster and label its intent.”
- “Here are my 25 page titles. Which common questions about [topic] do they not answer?”
- “Rewrite this seed phrase the way a beginner, a developer and a procurement manager would search it.”
Never ask “what is the monthly search volume of X”. The answer is made up.
How do you measure the result?
Keep the research honest by tracking it afterwards. For pages built from this workflow, watch impressions and clicks per target query in Search Console after 4 to 8 weeks, and compare the hit rate of the model’s ideas with the share that Keyword Planner confirmed as real. If fewer than half of the AI suggestions have any measurable volume, tighten the prompts with more customer language.
What mistakes should you avoid?
- Trusting volumes, difficulty scores or trends that the model states. None of them come from data.
- Skipping the intent check. Two phrases that look alike can need different pages.
- Mass-producing a page per phrase. Google’s spam policies cover scaled content abuse, and Google’s guide on AI features says writing variations of every long-tail query is not needed. One page answers a cluster.
- Feeding private data into a public chat. Strip customer names and revenue figures from what you paste.
- Forgetting local and language context. Say the country and language in the prompt, then check in a tool set to that market.
- Stopping at keywords. The best pages answer the question better than the current top results, with first-hand detail.
The free blog ideas generator from Fe/male Switch is another quick way to get topic starters. Next steps: turn the confirmed list into pages with long-tail keywords and write them with the AI-assisted content workflow.