Keyword research looks mechanical from the outside, but every step is a judgement call: what is relevant, what fits the searcher's intent, what can rank and what is worth turning into content. ChatGPT does not remove the need for judgement. It changes which parts are worth automating. This is the three-step workflow our SEO team presented at SEO Vibes Tour Athens.
Intelligent keyword research organisation
Before AI comes into play, the research plan needs structure. A typical e-shop or B2B site has dozens of intent clusters: buying, informational, comparison, brand and non-brand, plus long-tail variants per persona, use case and funnel stage. Without that structure, AI produces volume, not coherence.
This step is human work: define the clusters, map the stages of the customer journey the keywords must cover, and set boundaries for what is in scope and what is a different project (another product, market or intent). AI can help with first hypotheses, but the final structure is a judgement call and becomes the basis for on-page SEO and content marketing.
Use ChatGPT to automate and optimise the process
Once the cluster architecture is set, ChatGPT becomes genuinely useful as a volume generator. The pattern we use:
- Expand seed keywords into long-tail variants across each cluster, orders of magnitude faster than manual brainstorming.
- Infer search intent per keyword by prompting for the underlying question a user is likely asking.
- Generate topical outlines for cluster-level content briefs, which the SEO team then reviews and refines.
- Cross-check competitor coverage: prompt the model with competitor URLs and ask what keyword gaps are visible.
Critically, none of this output ships as-is. It is raw material for the SEO team to validate with classic tools (Search Console, keyword volume data, SERP inspection) and with their own judgement.
Iterate based on real results
The third step is where most AI-assisted keyword workflows break down. Teams prompt, get output, ship, and never close the loop. The discipline is to feed real performance data back into the prompting workflow — which keyword clusters actually drove traffic, which content briefs converted, which terms the AI over- or under-indexed on.
After a few iteration cycles, the prompting gets sharper, the output gets more aligned with what actually works in this specific site and vertical, and the time saved compounds. The key to success is maximising time efficiency and human productivity — not replacing the SEO practitioner, but removing the tedious volume work so the practitioner can spend time on the judgement calls that matter.