AI keyword research: how to use AI for keyword research properly
Where AI genuinely improves keyword research — intent clustering, gap analysis, question expansion — and where it invents volumes you should not trust.
AI has changed keyword research more than almost any other SEO task, but not in the way the marketing suggests. This article covers where models genuinely help, where they produce confident nonsense, and how to combine them with real data.
The critical distinction
Models are good at language and bad at facts they were never given. Keyword research contains both kinds of work, and the split is clean:
Language work — generating phrasings, grouping terms by intent, mapping keywords to pages, spotting missing subtopics, inferring what a searcher wants — is exactly what models do well.
Factual work — search volume, keyword difficulty, cost per click, actual ranking positions — must come from a data source. A model asked for search volumes will produce plausible numbers that are simply invented. This is the single most common misuse, and the output is worse than useless because it looks credible.
The rule is straightforward: use AI for structure and language, use a keyword data provider for numbers.
Where AI genuinely improves the work
Intent clustering at scale. Give a model a few thousand queries and it will group them by what the searcher wants far more accurately than string-similarity clustering, because it understands that two phrasings with no shared words can be the same request.
Question expansion. Models are very good at generating the realistic ways people phrase a question, which matters more now that engines fan a single query out into many retrievals. Feeding a model your core topic and asking for the fifty ways a buyer might ask about it produces a usable prompt and coverage list.
Mapping keywords to existing pages. Supply your URL list with titles and a query set, and a model will produce a sensible mapping, including flagging queries with no good home and pages competing for the same intent.
Gap analysis against competitors. Give a model the subtopics covered by the top-ranking pages and your own page, and ask what is missing. This is summarisation, which is reliable.
Prompt-set construction. For AI search work you need questions rather than keywords, and models are good at converting a keyword list into the natural-language questions a person would actually type into an assistant.
How to run it
- Pull real data first — queries from Search Console, plus a keyword provider for volume and difficulty. This is your factual base.
- Give the model the actual list rather than asking it to generate one from scratch. Grounded tasks produce grounded output.
- Ask for clustering by intent, with each cluster labelled by the job the searcher is trying to do.
- Ask it to map clusters to your existing URLs and flag conflicts and gaps.
- Ask it to expand each priority cluster into the natural-language questions people ask assistants.
- Validate the volumes for anything new against your data provider before committing to it.
- Keep the output in a spreadsheet you own, not locked inside a tool.
Common failure modes
- Accepting invented metrics. If a number did not come from a data source, it is not a number.
- Asking for keyword lists with no input. The result is generic category vocabulary with no relationship to your market.
- Over-clustering. Models will happily produce two hundred clusters from a list that supports thirty. Specify the target granularity.
- Ignoring your own data. Search Console shows queries you already receive impressions for, which is far more actionable than a database estimate.
- Treating the output as a plan. Clustering produces a map; deciding what to publish is still a judgement call.
Keywords versus questions
One genuine shift is worth naming. Traditional keyword research optimises for terms people type into a search box. AI search requires the questions people ask assistants, which are longer, more conversational, and frequently include context about their situation. Both lists matter, and they are not interchangeable. Building the second list is a task models are unusually well suited to, since it is pure language generation grounded in a topic you supply.
Common questions about AI keyword research
- Can AI give me search volumes? No, and this is the most damaging misuse. A model asked for volumes will produce plausible invented numbers. Volume, difficulty, and cost per click must come from a data provider.
- How do I do keyword research using AI properly? Pull real data first, then give the model the actual list and ask it to cluster by intent, map clusters to your existing URLs, flag gaps and conflicts, and expand each cluster into natural-language questions.
- What are the best AI tools for SEO keyword research? Those that work from your own Search Console and crawl data rather than a generic database, let you see and adjust the clustering logic, and export cleanly.
- Is AI better than traditional keyword tools? It is better at language tasks — intent grouping, phrasing expansion, page mapping. It is worse at facts. The two are complementary, not substitutes.
- Do I still need keywords if buyers use assistants? You need both lists. Keywords are what people type into a search box; questions are what they ask an assistant, and they are longer and more contextual. Converting one into the other is a task models do well.
- How do I stop the model over-clustering? Specify the target granularity in the prompt. Left unconstrained, models will happily produce two hundred clusters from a list that supports thirty.
Tip: keep the clustering prompt and reuse it. The value compounds when the same logic is applied to every refresh, and inconsistent clustering between quarters makes trends impossible to read.
Still stuck? We typically reply within 1 business day.
Contact support