How to Do Keyword Research with AI (Step by Step)
A practical, step-by-step process for keyword research with AI: prompts to paste, how to cluster and prioritize terms, and where AI gets search data wrong.
For Marketing

How do you do keyword research with AI?
Use AI for the language work and real SEO tools for the numbers. Ask AI to brainstorm seed topics and phrasings, expand them into candidate keywords, group them into clusters, label the search intent behind each one, and draft a content plan. Then get search volume and difficulty from a tool like Google Keyword Planner, Semrush or Ahrefs, because AI chat tools cannot reliably tell you how many people search a term. Finally, check the live results page yourself before you commit to writing anything.
That split is the whole method. AI is fast at sorting, grouping and rewording. It is unreliable at facts that live in a database it cannot see.
TL;DR
- AI shortens the grunt work: brainstorming phrasings, clustering hundreds of terms and labeling intent can take minutes instead of days.
- Never trust AI for search volume or difficulty. Those numbers come from keyword tools or Google Search Console, and a chatbot will happily invent them.
- Start from your customers' words, not yours. Paste in sales call notes, support tickets and survey answers and let AI pull out the phrases.
- Cluster before you prioritize. One page should target one group of related keywords, not one keyword.
- Look at the actual results page for your top picks. AI can summarize what ranks, but only if you paste in what you see.
- Keep the output in one sheet with intent, cluster, volume, difficulty, owner and status, so the research turns into a plan someone can act on.
What do you need before you start?
You need four things, and none of them require a paid SEO suite to begin with.
- A goal. "More organic signups for our expense software" is a goal. "Rank for stuff" is not. The goal decides which keywords matter.
- A real AI assistant. Any current general-purpose chat tool works (ChatGPT, Claude, Gemini, Copilot). One that can read an uploaded spreadsheet is more useful.
- A source of real search data. Google Keyword Planner is free with a Google Ads account. Google Search Console is free and shows what your site already appears for. Semrush and Ahrefs are paid and go deeper. Check each vendor's pricing page for current plans.
- A spreadsheet. Everything ends up here.
If you work at a large company, check your policy before pasting anything into a public AI tool. Keyword lists are low risk. Customer emails, call transcripts and unreleased product names are not. Use your company's approved AI tool for those.
Step 1: How do you pick seed topics?
Seeds are the broad topics your audience cares about. Weak seeds produce weak keyword lists, so spend time here.
Good raw material, roughly in order of value:
- Phrases customers use in support tickets, sales calls and reviews
- Your own product and category pages
- Questions from your sales and customer success teams
- Terms you already get impressions for in Search Console
- Competitor page titles and headings
Paste a sample of that material into AI and ask it to extract the language.
Prompt: extract seed topics
I work on marketing for [company/product] which helps [audience] do [job].
Below are [support tickets / call notes / survey answers].
1. List the 15 most common problems or questions, in the customer's own words.
2. Group them into 5 to 8 broad topics.
3. For each topic, give 3 ways a person might type it into Google,
including a casual phrasing and a more specific one.
Do not invent search volumes. Do not add topics that are not supported by the text.
[PASTE TEXT]
What to check: Every topic should trace back to something in your text. If a topic appears that nobody mentioned, AI filled in a generic answer. Delete it or flag it as a guess.
Step 2: How do you expand seeds into a keyword list?
Now you widen the net. Ask AI for variations along several angles, because people search the same need in many ways.
Prompt: expand a seed
Seed topic: "[topic]"
Audience: [who]
Their goal: [what they want]
Generate 60 search queries a real person might type about this, covering:
- questions (how, what, why, can, should)
- comparisons ("X vs Y", "alternatives to X")
- problems and error-style searches
- "best" and "for [specific role or industry]" searches
- beginner and expert phrasings
- templates, examples and checklists
Output as a table: query | angle | likely stage (learning, comparing, ready to buy).
No duplicates. No volume estimates.
Run this for each seed. You will end up with several hundred candidates, many of them near-duplicates. That is fine. Clustering handles it in Step 4.
You can also ask AI to mine your competitors. Paste in the headings from a competitor's blog index or their sitemap page titles and ask: "What topics do these cover that my list does not?" This works well because you supply the source text.
Step 3: How do you get real search volume and difficulty?
This is where most "keyword research with AI" advice goes wrong. A chatbot does not have access to Google's search data. If you ask "what is the monthly volume for 'project budget template'?", it will give you a confident number that is made up or badly out of date.
Instead, take your candidate list and run it through a tool that has the data:
- Google Keyword Planner gives volume ranges and a competition indicator aimed at advertisers. Good for a rough sense of scale.
- Google Search Console shows queries your site already gets impressions and clicks for. This is the most honest data you will find, and it often reveals easy wins: terms where you rank on page two.
- Semrush, Ahrefs and similar tools provide volume estimates, keyword difficulty scores and the pages that rank. Their numbers are estimates, and they differ between tools, so use them to compare keywords against each other rather than as exact truth.
Export the results as a CSV and combine them with your AI-generated list. Many tools also have a built-in keyword suggestion feature. Run your best seeds through it and add what it finds, since it is based on real queries.
Step 4: How do you cluster keywords with AI?
Clustering is where AI saves the most time. Grouping 500 keywords by hand takes hours. The principle: keywords that mean the same thing, or that one page could satisfy, belong in the same cluster.
Prompt: cluster keywords
Below is a list of keywords with monthly search volume.
Group them into clusters where one web page could reasonably target the whole group.
For each cluster, give:
- Cluster name (a plain-English topic)
- Primary keyword (the best single target, usually the highest-volume one that still reads naturally)
- Supporting keywords (the rest of the group)
- Total volume of the cluster (add up the numbers I gave you)
Rules:
- Only use keywords from my list.
- If a keyword does not fit anywhere, put it under "Unassigned".
- Keep clusters small enough for a single article.
[PASTE KEYWORDS AND VOLUMES, or upload the CSV]
What to check:
- Spot-check the math. AI is not dependable at adding long columns of numbers. Verify a few cluster totals in your spreadsheet, or ask it to leave totals out and add them with a formula.
- Look for over-merging. "Best CRM for small business" and "what is a CRM" are different needs and different pages, even though both mention CRM.
- Look for over-splitting. "How to write a budget" and "how do you write a budget" belong together.
- Read the "Unassigned" list. Sometimes it holds a whole missed topic.
Step 5: How do you work out search intent?
Search intent is what the person wants when they type the query. It decides what kind of page can rank. The four common types:
| Intent | What the searcher wants | Typical page |
|---|---|---|
| Informational | To learn something | Guide, explainer, how-to |
| Commercial | To compare options before choosing | Comparison, "best of", review |
| Transactional | To buy or sign up | Product or pricing page |
| Navigational | To reach a specific site | Brand or login page |
Ask AI to label every cluster.
Prompt: label intent
For each cluster below, label the dominant search intent
(informational, commercial, transactional, navigational)
and name the page format that best fits.
If a cluster mixes intents, say so and suggest splitting it.
[PASTE CLUSTERS]
Treat the labels as a first draft. The real test of intent is what Google already shows, which is the next step.
Step 6: How do you check the results page?
Take your 10 to 20 most promising clusters and search the primary keyword yourself, in a private browser window. Note:
- What formats rank (guides, lists, tools, videos, product pages)?
- Who ranks (big brands, small blogs, forums, Wikipedia)?
- Are there features like featured snippets, "People also ask" boxes or shopping results?
- Do the top pages look like something you could beat, or are they from sites with far more authority?
If the first page is all product pages and you were planning a blog post, your intent label was wrong. If it is all giant publishers, that keyword is probably out of reach for now.
Then use AI to compare what you found. Copy the titles and headings of the top five pages into a chat.
Prompt: analyze what ranks
Here are the titles and headings of the top 5 pages for the query "[keyword]".
1. What do they all cover?
2. What do only one or two cover?
3. What questions or angles appear to be missing from all of them?
4. What would a page need to include to be at least as complete as these?
Base this only on the text I pasted.
[PASTE TITLES AND HEADINGS]
This gives you a content outline grounded in what actually ranks, plus a gap to fill.
Step 7: How do you prioritize which keywords to target?
You will have more opportunities than time. Score each cluster on a few simple factors:
- Business value: how close is this searcher to buying or doing what you want? (1 to 5)
- Volume: use your tool's number, or a relative scale.
- Difficulty: from your tool, adjusted by what you saw on the results page. (Low, medium, high)
- Fit: can you credibly speak on this topic with something original to say? (1 to 5)
- Existing position: if Search Console shows you already rank on page two, improving that page is often faster than writing a new one.
Business value and fit are your judgment, not data. AI can propose scores, but you should set them, because it does not know your margins, your product strengths or your sales priorities. A cluster with modest volume and strong buying intent often beats a high-volume curiosity query.
What does the finished sheet look like?
Keep one row per cluster. Copy this template into a spreadsheet:
| Cluster | Primary keyword | Supporting keywords | Intent | Volume (source, date) | Difficulty | Business value (1-5) | Fit (1-5) | Page type | Existing URL or "new" | Owner | Status |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Expense policy basics | expense policy template | company expense policy, travel expense policy sample | Informational | from tool, month/year | Medium | 4 | 5 | Template page | New | Priya | Brief drafted |
Note the source and date next to every volume figure. Numbers from different tools and different months are not comparable, and you will want to know where each one came from in six months.
Once rows are prioritized, AI can turn each into a content brief.
Prompt: write a content brief
Create a content brief for this keyword cluster.
Primary keyword: [..]
Supporting keywords: [..]
Intent: [..]
Audience: [..]
What top-ranking pages cover (from my notes): [..]
Gap we can fill: [..]
Include: suggested title (under 60 characters), meta description (under 160),
H2 outline, questions to answer directly, what original element we should add
(example, data, template, expert quote from our team), and internal links to suggest.
Do not invent statistics or sources. Mark anything I need to research as [NEEDS SOURCE].
Where should you not trust AI in keyword research?
- Search volume, difficulty and trend data. Always from a tool.
- "What ranks" claims. Unless you pasted the results in, it is guessing or remembering old data.
- Competitor facts. It may describe a competitor's content that does not exist.
- Fresh topics. New products, new regulations and recent events fall outside what a model reliably knows.
- Exact-match assumptions. AI often treats near-synonyms as identical. Google sometimes does too, but not always.
- Your business priorities. It will optimize for what is searchable, not what is profitable for you.
A practical rule: if a claim can be checked in a tool or on the live results page, check it before it goes into the plan.
How often should you redo keyword research?
Revisit the sheet every quarter, and check Search Console monthly. Queries shift, competitors publish, and your own pages start ranking for terms you did not plan for. Those unplanned queries are often the best new ideas. Export them, paste them into the clustering prompt from Step 4, and see which clusters grow.
Could you build your own keyword tracker instead of managing a spreadsheet?
A shared spreadsheet works until several people edit it, statuses drift and nobody knows which brief belongs to which row. If that is where you are, a small internal tool is a reasonable next step. With Remy, you can describe a keyword and content tracker in plain English: clusters, primary keywords, intent, volume with its source and date, owner, status, and a page where an editor can see what is in progress. Remy builds it as a full-stack app with a database and sign-in, and deploys it to a live URL, so your team uses one place instead of passing files around. You own the code, and it is a tracker shaped to your process, not a subscription you adapt to.
Frequently Asked Questions
Can AI replace keyword research tools like Semrush or Ahrefs?
No. AI chat tools do not have access to current search volume or difficulty data, and they will produce plausible-looking numbers if asked. Use AI to generate ideas, cluster and plan, and use a keyword tool or Search Console for the data. Some SEO platforms now include AI features built on their own data, which is a different thing from asking a general chatbot.
Is keyword research with AI good enough for a beginner?
Yes, and it helps beginners in particular, because AI explains concepts like search intent and suggests phrasings you might not think of. The risk is accepting its output without checking. Follow the steps above, pull real data from Google Keyword Planner or Search Console, and look at the live results before choosing a target.
How many keywords should I research for one article?
Aim for one cluster per article: a primary keyword plus a handful of closely related phrases that the same page can satisfy. You do not need to include every variation. Cover the topic well, and the page will usually pick up related searches on its own.
Is it safe to paste company data into an AI tool for keyword research?
Keyword lists and public page text are generally low risk. Customer conversations, internal documents and unreleased product details are different. Follow your company's AI policy and use an approved tool for anything sensitive, or remove names and identifying details before pasting.
Will Google penalize content planned with AI?
Choosing topics with AI is not the issue. What matters is whether the finished page is useful, accurate and written for the searcher. Use AI for research and planning, then make sure each page has something real to offer: original examples, firsthand experience, correct facts.
Useful for
How we source, write, and correct our articles. Editorial standards
Related articles

Content Calendar Template (Plus an AI Prompt to Fill a Quarter)

How a Grok Bot GTA 6 Clipping Bot Works (and Its Results)

How Much Does an AI Short Film Cost? The $102,000 Breakdown
