How to Automate Keyword Research With AI (Step by Step)

A founder I work with pasted a list of 312 keywords into a spreadsheet, sorted by search volume, and started writing posts from the top down. Three months later the top terms had brought in nothing, because every one of them was a query that big brands already owned and his new domain had no chance of ranking for. The list was not wrong. The order was. Keyword research that starts from volume and ends in a spreadsheet is the most common way founders waste their first quarter of SEO.

By the end of this guide you will have a short list of queries your real buyers search, that you can plausibly rank for, ready to hand to drafting. AI does the heavy lifting at every stage. You keep the two judgments it cannot make: does this query match a buyer, and can you actually win it. You need two things before you start: a live site, and Google Search Console connected to it. The first gives you something to rank; the second gives you the only honest demand data you will get for free.

Step 1: Seed from what your buyers actually search

Open a blank doc and write down the words your customers use, not the words your marketing page uses. These are different. Your homepage might say "revenue intelligence platform" while your buyers type "how to track which marketing actually drives sales." The seed list is the second one.

Pull seeds from real sources: your sales call notes, the questions in your support inbox, the subject lines of emails people send you, the subreddits where your buyers complain. If you sell project software to agencies, seeds look like "client approval workflow," "agency time tracking," "scope creep template." Aim for 15 to 25 honest seeds. AI can help here too: paste your product description and ask it to list the problems a buyer would search before they know your category exists. Treat that output as a draft, then cut anything that sounds like you and add anything that sounds like them. This buyer-language step is the same instinct behind good SEO for startups: rank for the problem, not the product name.

Step 2: Use AI to expand seeds into long-tail variants

Now take each seed and have AI fan it out into the specific, lower-volume phrasings people actually type. One seed like "client approval workflow" expands into "how to get client sign-off faster," "client approval process for design agencies," "approval workflow template for freelancers," and forty more.

This is the step AI genuinely transforms. By hand, brainstorming long-tail variants is slow and you run out of imagination after a dozen. AI does not get tired, so it surfaces the question forms, the "for [audience]" variants, the "vs" comparisons, and the "best [thing] for [use case]" phrasings in seconds. Long-tail queries matter for a new site because they carry clear intent and far less competition than the head term. Nobody fights over "approval workflow template for freelancers" the way they fight over "project management." Ask AI to generate 30 to 50 variants per seed, then move on. You will cut most of them later, and that is fine. The point of this step is coverage, not quality. Good keyword research is more about using AI for marketing to widen the funnel of candidates than about polishing any single one.

Step 3: Pull demand signals from Search Console and real queries

A long list of plausible-sounding phrases is still a guess until you attach demand to it. Two sources turn the guess into evidence: a keyword tool for raw volume, and Google Search Console for queries you already touch.

Search Console is the underused one. Its performance report shows the exact searches that already bring impressions to your site, including phrasings you never deliberately targeted. A query sitting at position 14 with 200 monthly impressions is gold: people are searching it, Google already thinks you are relevant, and a focused post can move you onto page one. Export that report and fold those real queries into your candidate list. For raw volume on terms you do not yet rank for, a keyword tool fills the gap. Attach a rough monthly volume to every candidate so you can tell a term twelve people search from one twelve thousand search. Be honest that volume tools estimate; treat the numbers as order-of-magnitude, not precise. The signal you trust most is your own Search Console data, because it is measured, not modeled.

Step 4: Cluster by intent and dedupe against existing posts

You now have a messy list of hundreds of candidates with volume attached. Most of them are near-duplicates of each other. Group them.

Ask AI to cluster the list by search intent: which queries are the same question wearing different words. "How to get client sign-off faster," "speed up client approval," and "reduce client approval time" are one cluster, one future post, not three. Clustering matters because a single strong page ranks for the whole group, while three thin pages on the same intent compete with each other and none of them win. While you cluster, dedupe against what you have already published. Run each cluster against your live URLs and drop any intent you already cover, because a second post on the same intent splits your own authority. AI is fast at proposing clusters, but read them. It will sometimes lump "approval workflow software" (someone shopping for a tool) with "approval workflow template" (someone wanting a free doc), and those are different buyers at different stages. Splitting that is your call, not the model's.

Step 5: Pick the few winnable queries (intent plus low competition)

Here is where most automated keyword research quietly fails, and where a human has to step in. You have clean clusters with volume. Now pick the handful worth writing, judged on two axes the volume number does not show: does the query match a buyer, and can a new site rank for it.

Intent first. A high-volume cluster that attracts students, job seekers, or the merely curious is traffic you cannot convert. "Project management certification" has huge volume and zero relevance to a project software you sell. Cut it. Competition second. For each surviving cluster, open the actual first page of Google results and look at who ranks. If it is a wall of established brands with deep pages, skip it for now; a new domain will not break in. If you see thin posts, forum answers, or off-topic results, that is a gap you can take. AI can flag an estimated difficulty score to triage the list, but the deciding read is your eyes on the real results page. Aim to leave this step with five to ten clusters that are relevant and winnable. A short list of beatable queries beats a long list of impossible ones, the same lesson the founder with 312 keywords learned the slow way.

Step 6: Hand the shortlist to drafting

The output of all this is small on purpose: five to ten clusters, each with a primary query, its supporting long-tail variants, and a one-line note on the buyer it serves. That note is what makes the draft good, because it tells the writer who is searching and what they need to read.

Hand that shortlist to whoever writes, with the cluster intact so one post can earn the whole group of searches. Each post targets the primary query, works in the variants naturally, and answers the specific buyer question behind the cluster. This is also where the research connects to distribution beyond Google: a page built around a clear buyer question is the same kind of page that gets recommended by ChatGPT when someone asks an assistant the same thing. Research and drafting are not separate projects. The shortlist is the bridge.

What to do next

Run the six steps once, by hand, on your real product this week. Seed from buyer language, expand with AI, pull Search Console, cluster, judge intent and competition, ship a shortlist. You will feel exactly where the work is slow and where it is judgment.

That split is the whole reason Revnu exists. Revnu runs keyword research as one lane of a cross-channel growth agent: it pulls your Search Console data, expands seeds into long-tail clusters, and proposes a shortlist of winnable queries scored on intent and competition. Then it stops and waits for you. You approve the queries that match a real buyer, the agent drafts the posts, and nothing publishes without your sign-off. The model does the fan-out and the clustering you would dread doing in a spreadsheet. You keep the two calls that decide whether the work pays off. See how the lanes fit together with AI growth agents.