How to Do Keyword Research with AI (Step-by-Step)

AI is genuinely good at the language parts of keyword research: generating seed keywords at scale, clustering semantically related terms, labeling search intent from the titles on page one, extracting questions from People Also Ask boxes, and drafting content briefs. These are pattern-matching tasks over text, and modern LLMs do them faster and cheaper than a human analyst. A well-prompted model can hand you 60 raw seeds for a topic in two minutes — the same list that used to take an afternoon of spreadsheet staring.

AI is genuinely bad at the measurement parts: real search volume, keyword difficulty, click-through rates, and the current composition of the SERP. An LLM does not see Google’s index. It will confidently produce numbers that look like monthly volume but are estimates reconstructed from training data — in practice, hallucinations. This is not a prompt-engineering problem; it is a data-access problem. The only tools that know what people actually search are the ones with clickstream and crawl data, like Ahrefs, Semrush, and Surfer SEO.

The practical rule: treat LLM output as hypotheses and keyword-tool output as evidence. Generate with ChatGPT or Claude, validate every number with Ahrefs or Semrush, and use Surfer SEO for the SERP-level checks. The workflow below is built exactly this way, so no step depends on a number an AI invented.

Run these six steps in order, and keep the artifacts from each step — they feed the next one. Budget two to four hours for your first topic cluster; it drops to about an hour per cluster once the routine is set.

Scenario (illustrative example): Northwind, a B2B SaaS selling AI support agents, wants one new blog post. In Step 1, Claude returned 60 seeds around “AI customer support.” In Step 2, pasted into Ahrefs Keywords Explorer with a US volume floor of 150 and difficulty cap of 40, eleven seeds survived. Three carried the post: “AI customer support software” (commercial, difficulty 38), “AI customer support statistics” (informational, difficulty 29), and “how to implement AI customer support” (informational, difficulty 22). The seed “AI support agent cost per ticket” — which sounded perfectly plausible — returned no measurable volume in the tool and was deleted.

In Step 3, the live SERP for “AI customer support statistics” showed an AI Overview summarizing headline figures at the top. The team kept the keyword anyway, because searchers hunting for statistics still click for a citable list with sources — the overview is partial, not complete. Step 4 confirmed the intent labels, Step 5 clustered the three keywords into one hub page (the software comparison) plus this post, and Step 6 produced a brief whose H2s came from the People Also Ask questions and the top-10 titles, not from the model’s imagination.

The split of labor is the point of the example: the LLM contributed the raw seed list and the first draft of the brief, the keyword tool eliminated roughly a third of the seeds and supplied every number, and the human made the three decisions that set direction — which cluster to target, which intent to write for, and which SERP was winnable. Remove any one of the three and the post would have targeted something nobody searches.

Every figure below comes from official pricing pages fetched on 31 August 2026. Where a page rendered its prices in JavaScript, we say so and print no numbers — the same standard we applied in our small business tools ranking. Vendor claims are flagged as claims.

The teams that get the least from AI keyword research repeat the same five mistakes. Each one is a workflow failure, not a tool failure — and each is avoidable.

Q: Can ChatGPT do keyword research? A: It can generate seeds, cluster terms, label intent from titles, and draft briefs — the language parts. It cannot measure search volume or see the live SERP, so every number it produces must be validated in Ahrefs, Semrush, or Surfer before you target anything. Use it as the generation layer in a workflow, not as the referee.

Q: What is the best AI tool for keyword research in 2026? A: For validation: Ahrefs (entry plans “from $199/mo” per its pricing page, fetched 31 Aug 2026) or Semrush (free tools at $0; paid tiers not statically verifiable on the same date). For SERP-level optimization and AI-answer tracking: Surfer SEO from $49/month billed yearly. For generation: ChatGPT or Claude Pro at $20/month. The workflow beats any single tool.

Q: Is keyword research still worth it when AI Overviews answer queries directly? A: Yes, but the target changed. You want queries where the AI answer is partial — statistics roundups, tool comparisons, how-to sequences — because searchers still click for the source. Queries the overview fully answers are zero-click by design; skip them and let competitors rank for zero visits.

Q: How much does an AI keyword research stack cost? A: A free start is real: Semrush’s free tools plus ChatGPT’s free tier cover Steps 1–2 for small lists. A paid setup runs about $20/month for Claude Pro plus one keyword tool: Surfer Discovery at $49/month (billed yearly) is the cheapest verified entry, and Ahrefs entry plans start “from $199/mo” per its page. Most solo operators need nothing between those two.

Q: How long does AI keyword research take? A: Two to four hours for the first topic cluster, about an hour per cluster once the routine is set. The LLM steps take minutes; the SERP checks and intent decisions take the rest — and they are exactly the steps that decide whether the content joins the 3.45% of pages that get traffic.