How to Do Keyword Research with AI [Free Stack]

Why does ChatGPT keep handing me keywords nobody actually searches?

You’ve typed “give me keyword ideas for [your niche]” into ChatGPT, got a clean list of long-tails that sound perfect, and then watched Google Keyword Planner or a free checker return “no data” or a sad little range. That’s the exact moment most beginners either quit or publish into a void.

Publishing into that void feels productive for a week. Then the calendar is full of pages that never earned a single qualified visit – not because you wrote badly, but because the phrases were cosplay search demand. This walkthrough is the filter that stops that tax. Free stack first. Paid only when the free path hits a real wall.

AI expands. Data decides.

AI is great at plausible language and grouping ideas by meaning. It does not see live volume, organic difficulty, or today’s SERP features unless you paste a real export or wire up a database.

Unconnected chatbots still emit volumes, KD scores, and trend calls anyway. That isn’t “estimate” energy – it’s training-data fiction. Ahrefs’ AI keyword research guide spells out the failure mode: no live keyword DB, no trustworthy metrics.

Filter job, in one line: real demand signals + business fit + a page type you can own. Skip it and you write for sentences that sound searchable, not queries people type or ask assistants.

Free stack: one niche, end to end

Finish this on a single topic before you open a paid SEO suite.

1. Seed with ruthless context

ChatGPT free tier is enough to start. Plus is $20/mo as of mid-2026 if you need higher limits. Paste this and swap the brackets – notice what it forbids:

I run a [small local bakery / indie tool for freelancers / whatever]. Audience: [busy parents / solo designers who hate admin]. Goal: [email signups / product demos], not vanity traffic.

List 40 real search queries and full-sentence questions this audience would type into Google or ask ChatGPT/Perplexity about [core problem, e.g. "meal prep for picky kids" or "invoicing without accountants"].
Group by rough intent: learn, compare, buy/do now.
Do NOT invent monthly search volumes, difficulty scores, or "high opportunity" labels. Just the phrases.

Copy the plain list. Half of it dies in the next step. That’s success.

2. Validate demand without paying

Google Keyword Planner needs a Google Ads account. You do not need to spend.

No-spend views give ranges (100-1K, 1K-10K), not exacts. Two terms in the same bucket can sit several times apart in real demand – treat a shared range as “maybe,” not a tie. The Competition column is advertiser heat, not organic rank difficulty. Keep non-zero ranges that still sound like your audience. Delete phrases the tool never recognizes.

Direction check: Google Trends (relative 0-100, rising/breakout, related queries). A rising mid-range phrase beats a flat head term sliding down. Gemini-related suggestions inside Trends (as of 2026 product updates) are extra variants, not volume proof.

3. Mine questions and assistant phrasing

Strongest survivors go into free AlsoAsked (tight daily cap) or manual People Also Ask expansion. Then ask Perplexity or ChatGPT the same problems and note which URLs get cited.

Those citations plus the long conversational prompts buyers actually use are the modern long-tail pile – often uglier than classic keyword tools admit.

Volume is a proxy. If sales calls and AI answers keep surfacing a near-zero phrase, it can still deserve a page. Just don’t expect a classic organic traffic hockey stick.

4. One more AI pass: cluster and map

Paste only the cleaned list:

Group these validated keywords into 5-8 topic clusters. For each cluster give: primary phrase, 3-5 supporting questions, dominant intent, and the single page type that should own it (guide, comparison, landing, FAQ). Flag any that feel off-brand for [your business].

Sheet time. One cluster → one URL (or a tight internal-link set). If you already have Search Console data, pull those queries first – free, first-party, already proven.

Ever notice how “perfect” keyword spreadsheets calm the anxiety and still don’t move revenue? That’s the tell you’re scoring fiction. Demand you can defend beats a prettier grid.

Pitfalls that burn weeks

Force the model to refuse metrics in the prompt. Validate outside. A content calendar built on “2,400 searches, KD 18” that never existed is how quarters disappear.

Same GKP bucket ≠ same opportunity. Before you write, skim Trends direction and the live SERP: what format ranks, is an AI Overview present, who actually owns the intent?

Your buyer asks ChatGPT a full sentence. Classic tools shrug “zero.” Community SEO threads and 2026 practice keep showing those conversational lines get cited inside assistants anyway. Write the ones that show up in real buyer language – and separate “AI citation value” from “classic rank-and-traffic value” in your own head.

Meaning-based clusters can still misread the SERP. Open the top few results for the primary term. If Google rewards listicles and your draft is a product pitch, fix the page type before you fix the adjectives.

When the free stack is enough – and when it isn’t

Stay free until exact prioritization across many clusters (or bulk competitor gaps) is the blocker. Prices shift – confirm on Ahrefs pricing and Semrush pricing before you buy.

FAQ

Can I do real keyword research with only free AI and Google tools?

Yes for the first 1-3 topic clusters: ideation, rough demand filters, Trends direction, question mining. You will not get precise organic difficulty or bulk competitor gap views. That absence is the upgrade trigger – not a reason to skip the free pass.

What if AI keeps suggesting keywords my tool says have zero volume?

Keep it only if a real person would ask an assistant that way and a citation or sales conversation would still help. Else delete. Marketing-speak inventions go first.

Should I research “prompts” the same way as keywords now?

Yes – with limits. People dump longer, messier questions into ChatGPT, Perplexity, and Google AI Mode. Expand your best commercial and informational terms into full buyer sentences, then watch which sources get cited. Prompt databases are showing up inside major SEO suites for that layer.

What it is not: a full replacement for classic demand data. Volume still shows where traditional opportunity is fat. Prompt research shows where assistant-shaped demand hides – especially the awkward long-tails tools used to ignore.

Pick one niche you already know. Run the seed prompt, validate ten phrases in Keyword Planner + Trends, map one cluster today. Ship the first page from that map next. That single loop beats another hour of unread prompt galleries.