Using AI for keyword research: what it does well and where you still need real data

Use AI for keyword research by asking it to generate topic ideas, group keywords by search intent and map them to pages, then check every shortlisted term against real data such as Google Search Console or a keyword tool. As of 2026, AI is excellent at ideas and organisation, but it cannot reliably tell you how many people in the UK actually search for a phrase.

Most small businesses end up on one of two routes. The first is to type "give me 50 keywords for my plumbing business" into a chatbot and publish whatever comes back. The second is to use AI as a fast thinking partner, then validate its output against real search data. The first route is quick and often produces content nobody searches for. The second takes a little longer and produces a keyword list you can trust. What decides between them is simple: whether you treat AI's output as a hypothesis or as a fact.

What AI can and can't do in keyword research

Large language models are trained to predict plausible text. That makes them very good at language tasks and weak at measurement. Keyword research needs both, so it helps to split the job clearly.

Where AI genuinely saves time

• Generating seed ideas. Give it your services, location and customer types, and it will suggest dozens of phrasings you might not have considered, including UK-specific wording ("solicitor" rather than "attorney", "trainers" rather than "sneakers").
• Clustering. Paste in a list of 300 keywords and ask it to group them by topic. This is tedious by hand and quick with AI.
• Classifying intent. It can label each term as informational, commercial, transactional or local, which tells you whether the page should be a guide, a service page or a location page.
• Spotting gaps. Describe your existing pages and ask what questions a customer would ask that you haven't answered.

Where AI falls short

• Search volume. A chatbot may state a monthly volume with total confidence. Treat any such figure as invented unless it comes from a connected data source.
• Keyword difficulty. It cannot see who currently ranks or how strong those pages are.
• Freshness. Search behaviour shifts, and a model's knowledge may lag behind what people search for now.
• Local nuance. It may blur UK and US spelling or miss regional terms unless you tell it where you operate.

Google's own Search Central documentation is clear that content should be created to help people rather than to manipulate rankings, and that automation is acceptable when it isn't used mainly to game search. Keyword research that starts with real customer questions fits that principle well. Our piece on what separates good content automation from a Google penalty waiting to happen covers the line in more detail.

The steps, in the order that matters

1. Write a one-paragraph brief for the AI. Include what you sell, where you sell it (name the UK towns or regions), who buys it and what you want them to do. State "UK English and UK search terms" explicitly.
2. Ask for seed topics, not final keywords. Request 30 to 50 customer problems, questions and service descriptions. These are raw material.
3. Pull real data from Google Search Console. Open the Performance report and look at queries already earning impressions. Google provides this data directly, and it reflects your actual site rather than a model's guess. Pay attention to queries where you sit on page two, as these are often the quickest wins.
4. Run your seed list through a keyword tool. Use it to get volume ranges and difficulty indicators. An all-in-one platform such as Illumae includes keyword research alongside tracking and audits, which saves exporting between tools. Whichever you use, the aim is the same: replace AI guesses with measured numbers.
5. Cluster the validated list with AI. Now that the terms are real, ask AI to group them by topic and intent, and to suggest one primary keyword per page.
6. Map clusters to pages. Decide which clusters belong on existing pages and which need new ones. Avoid two pages chasing the same phrase.
7. Prioritise. Rank by a blend of relevance to revenue, realistic difficulty and current visibility. A modest term that brings buyers beats a high-volume one that brings browsers.
8. Set up tracking before you publish. Record starting positions so you can tell whether the work paid off. See how to use AI for keyword tracking for the next stage.

Prompts that produce usable output

The quality of AI keyword work depends heavily on how you ask. Vague prompts return generic lists. Specific prompts return something you can act on.

A weak prompt: "Keywords for my bakery."

A stronger prompt: "I run a family bakery in Leeds selling celebration cakes and wedding cakes. Customers are mostly local, and we want more enquiries, not just visitors. Suggest 40 search phrases UK customers might use, grouped by intent, using UK spelling. Include questions people ask before ordering."

Intent matters more than volume

A phrase like "wedding cake prices" signals someone comparing options, so a pricing guide suits it. "Wedding cake Leeds" signals someone ready to enquire, so a local service page suits it. AI is good at making this distinction quickly, and getting it right stops you writing a blog post for a search that really wants a service page.

Competitor angles

You can also paste in the headings from a competitor's page and ask what subtopics they cover that you don't. This is a legitimate way to find gaps, provided you write your own original material rather than copying. For a fuller treatment, see AI tools for competitive analysis.

How a shortlist is validated, step by step

This is an illustration using assumed figures, not real data. Assume a Manchester accountancy practice asks AI for keyword ideas and gets 40 suggestions. The owner then checks them against a keyword tool and Search Console.

Assumptions: of the 40 AI suggestions, 10 show no measurable search demand at all. Of the remaining 30, 12 are informational questions better suited to blog posts, and the rest are commercial or local terms.

1. Terms with no demand: 40 − 10 = 30 usable terms remain.
2. Commercial or local terms: 30 − 12 = 18 terms suit service or location pages.
3. Share lost to guesswork: 10 ÷ 40 = 25%.

So in this scenario a quarter of the AI's output would have been wasted effort if published unchecked. The proportion will vary with the prompt and the niche, so don't treat 25% as a benchmark. The point is that some share of unvalidated output will usually be unusable, and only real data tells you which share.

The owner now has 12 blog topics and 18 page targets, grouped and ready to schedule. Agencies repeating this across clients can read about managing multiple client domains without the process becoming unwieldy.

Where AI keyword research stops being reliable

Be honest about the edges of this approach, because they are where money gets wasted.

• Invented numbers. If an AI tool gives you volumes without naming a data source, assume they are estimates at best. Verify before committing budget.
• No ranking guarantee. A well-chosen keyword improves your odds, but rankings depend on content quality, links, site health and competitors. Nobody, including Google's documentation, promises a position.
• Small, new sites face real difficulty. Terms dominated by large UK publishers or national brands may be unrealistic for a young domain, however attractive the volume looks.
• Very niche terms. Low-volume tools often show zero for genuine buying phrases. Zero in a tool does not always mean zero demand, so use judgement and check Search Console.
• Privacy of your data. Before pasting customer data, sales figures or client details into a public AI tool, consider your obligations under UK GDPR and the ICO's guidance on AI and data protection. Keyword lists rarely contain personal data, but exported customer queries or CRM notes might.

The one thing to verify before you act on any keyword list is where each number came from. If a volume or difficulty score can't be traced to Search Console or a named keyword tool, treat it as a guess and check it before you build a page around it.