AI Keyword Research and Search Intent Clustering: A Practical Workflow

By Naveen Sehgal, Zylous Digital.

Anyone can export a list of thousands of phrases in a few minutes. The hard part comes after: deciding which of them deserve a page, what each searcher is hoping to find, and which phrases can share a page. That is slow, repetitive work, and it is where a language model earns its place. Used with care, AI keyword research takes over the tedious jobs, such as tidying exports, expanding seed terms, tagging the intent behind each phrase and grouping phrases that belong together.

Below is a workflow you can run with a spreadsheet and any general-purpose AI assistant. We start with where the data comes from, move through labelling, grouping and checking, and finish with the places where the method falls short.

Start with the reader's question, then the phrase

Google's guidance on helpful content asks whether a page was made primarily for people who visit directly or primarily to attract search traffic. Chasing trends with no real audience interest is on its list of warning signs. So a high-volume phrase raises a question: who is searching, and what do they need? Volume on its own is no reason to publish a page.

Google's SEO Starter Guide makes a related point: experts and beginners often describe the same thing differently, so it is worth thinking about the words a reader would actually type. It adds that repeating the same words over and over wears readers out, and that keyword stuffing breaks Google's spam policies. A good term list captures the different ways your audience phrases things, which lets each page be written in natural language.

Where the data should come from

A language model does not know how many people search for a phrase. It can suggest phrases that sound plausible, but it has no access to current search counts, and it can invent terms nobody uses. Search counts have to come from tools that actually measure them. The input file can draw on these sources:

Google Keyword Planner: ideas generated from seed words or a website, search volume estimates and forecasts. Google notes that very low volume terms and some sensitive categories, such as health and finance, may return no ideas or forecasts, and that forecasts are estimates affected by bids, budgets and ad quality.
Google Search Console: the Performance report lists the queries that already show your site, with clicks, impressions, click-through rate and average position. Google states that tables show a maximum of 1,000 rows and that rare and sensitive queries are not shown, so use the API or several filtered exports for larger sites.
Your own material: sales call notes, support tickets, site search logs and the questions clients ask. Customers often phrase things in ways no tool lists.
The search results themselves: autocomplete suggestions, People Also Ask boxes and related searches, worth recording by hand for the few topics that matter most.

Combine these into one sheet with a column for the phrase, a column for the source and, where available, a column for volume. Trust the volume and impression figures. Anything the model adds is only a suggestion until you have verified it.

What the AI is good for, and what it is not

Language models are well suited to work with text patterns. They can normalise spelling variants, remove near-duplicates, expand a seed topic into the questions a buyer might ask and sort a long list into categories. They are poor at facts that need a source. Anthropic's documentation on reducing hallucinations notes that models can produce text that is factually wrong and recommends giving the model explicit permission to say it does not know, and checking claims against quotes. Both habits apply directly to term lists: tell the model to mark any phrase it is unsure about, and check a sample against live results.

Google's own advice on generative AI content adds that people should review AI output for accuracy before publishing, titles, meta descriptions and image alt text included. Apply the same care at the planning stage. A wrong assumption made while clustering gets copied into every page built on it.

The workflow, step by step

Collect and merge. Export from your sources into one sheet and remove exact duplicates. Add a column recording where each phrase came from.
Clean. Ask the model to standardise spelling and plurals and to flag near-duplicates for you to confirm. Have it return a table and keep the original wording in a separate column so nothing is lost.
Expand with care. For each core topic, ask the model what a customer might want to know at each stage, from first learning about the subject to comparing suppliers. Label those rows as model-suggested, and test them against autocomplete or a volume tool before they shape the plan.
Label the intent of each phrase, using the categories in the next section.
Cluster the phrases into groups that one page could answer.
Validate the clusters against the live results for a sample of phrases.
Assign each cluster to a page: an existing page to improve, or a new page to write. Record the primary phrase and the supporting phrases for each.

Labelling intent

An intent label records what the searcher is trying to do. The four below are common industry shorthand, and other teams use different names. Whatever scheme you pick, apply it the same way every time.

Informational: the searcher wants to learn something, for example a definition, a method or a comparison of approaches.
Commercial investigation: the searcher is weighing options before deciding, for example by reading reviews, looking at alternatives or comparing prices.
Transactional: the searcher is ready to act, for example to buy, book or request a quote.
Navigational: the searcher is trying to reach a particular site or page, and the query is usually a brand name.

Many phrases fit more than one label, and a short phrase can mean several things. Ask the model to give a primary label, a secondary label where one applies and a one-line reason. The reason column makes the output easy to review, since weak logic shows at a glance. Look hardest at short head terms, because they are the most ambiguous.

Clustering by what the results show

Grouping by meaning alone pulls together phrases that merely sound alike. Grouping by intent, and by the results each phrase triggers, works better. Two phrases that bring up the same kind of page, with largely the same sites ranking, can usually share one page. Two phrases with similar words but different result types, such as a guide for one and a product listing for the other, usually need separate pages.

In practice, the model proposes groups from the text, and you then test the proposals. Search the main phrase for each group plus two or three of its supporting phrases in a private browser window, and note the type of page that ranks. When the results differ, split the group. This is a working rule from practice, not a measured threshold, so adjust it to your market. Keep in mind that results vary by location, device and time, which is also why Google warns that a query shown in Search Console may look different when you search it yourself.

Example prompt for a first-pass grouping

The prompt below is a template for the grouping step. Replace the text in square brackets and paste in no more than a few hundred rows at a time, so you can review each batch.

You are an SEO analyst. Task: group the phrases below into clusters that one web page could answer. For each cluster give a proposed page title, a primary label and secondary label from this list: informational, commercial investigation, transactional, navigational. Add one sentence explaining the grouping. Context: the site belongs to [business type] and serves [audience] in [region]. Format: a table with the columns cluster name, phrases, intent labels, reason. If you are unsure about a phrase, put it in a separate list called Unsure and do not guess. Phrases: [paste rows here].

Checking the output before you act on it

Open a random sample of ten to twenty phrases and confirm the intent label against the live results.
Confirm that every phrase in the Unsure list got a human decision.
Check that each cluster has a single main purpose. If you cannot describe the page in one sentence, split it.
Compare the clusters with pages that already exist, so you improve a page that is ranking before you write a competing one.
Run the same batch twice. Where the two runs group phrases differently, the boundary is unclear, so settle that one by hand.

Limits to be aware of

Ask a language model the same question twice and you may get two different answers, so cluster boundaries can shift from run to run. Its training data stops at a cutoff date, which means newer products and recent terminology can be missing or misread. Local, medical, legal and financial topics deserve extra manual checking, and Google's own tools may return nothing for very rare or sensitive terms. Remember too that a neat cluster map says nothing about ranking. Content quality, links, site authority and competition still decide that, and no workflow removes the uncertainty.

Common mistakes

Taking a model's suggested phrases as proof of demand before checking a volume source.
Building one page per phrase. You end up with thin, overlapping pages, and Google's spam policies warn against producing large numbers of pages with little added value.
Running the process once and leaving it alone. Query wording and results shift over time, so review the map every quarter.
Pasting client data, such as private analytics exports, into a public AI tool before reading its data terms.

Where this fits in a wider AI workflow

Term research and clustering sit at the planning stage of a content programme. Everything after it, from briefs and drafts to reviews and reporting, works best with the same split: AI handles the repetitive text work, and a person stays responsible for the decisions. You can read how Zylous Digital approaches this on the AI-first approach page.

Frequently asked questions

Can AI replace a keyword tool?

No. A language model can't tell you how often people search for a phrase. Use it to clean, label and group data that comes from Keyword Planner, Search Console and your own records.

How do I know two phrases belong on the same page?

Search both and compare what comes up. If largely the same sites and the same type of page appear, one page can usually cover both. If the result types differ, each phrase needs its own page.

How often should the clusters be reviewed?

Quarterly is a sensible starting point. Also look again whenever Search Console shows new queries that fit no cluster, and adjust the pace to how fast your market moves.