The typical keyword set handed over with a project is a 4,000-row export sorted by search volume. At the top sits "website" with tens of thousands of impressions; near the bottom, "order a landing page for a dental clinic" with thirty. The second one pays the bills, and the budget almost always goes to the first. What follows is how to harvest queries by intent, how to group them into clusters that map onto real pages, and how to decide what to build first.

Why sorting by search volume leads you astray

Volume tells you how many people typed a phrase. It says nothing about whether any of them wants to buy, or whether you stand a chance of ranking. Three reasons the impressions column is the worst possible sort order for a keyword set.

• High-volume queries are nearly always informational. People searching "website" want definitions, examples and builders, not a vendor. Google fills those results with articles, and a service page will not get in however well optimised.
• The top of broad queries already belongs to aggregators. Directories, marketplaces and large publishers hold link profiles a small business will not match within a year. Compete where the top ten contains sites your own size.
• Low volume does not mean low demand. Thirty impressions a month converting at 5% is one and a half enquiries; ten such queries make fifteen. Tools also underestimate long phrases, so the real number is usually higher.

Where to find queries besides the keyword planner

Google’s Keyword Planner is a poor source of ideas: it only returns what you already thought to ask about, and it merges close variants into one estimate. The most valuable queries come from places where people describe their problem in their own words.

A reasonable output from stage one is 300–800 unique phrases for a service business, or a few thousand for a store once product queries are counted. More is wasted effort — anything that does not land in a cluster gets discarded at the next step anyway.

Classifying by intent: the step that decides everything

Intent determines what kind of page Google is willing to show at all. Getting it wrong is expensive: a service page targeting an informational query can sit unranked for years regardless of how many links point at it.

Clustering and the page map

A cluster is a group of queries one page can satisfy while ranking for all of them. Grouping by word similarity is the classic error: "website development" and "website creation" really are one thing, whereas "website cost" and "website development stages" are not, despite the shared words.

1. 1

   Compare the results, not the wording

   Pull the top ten for two queries. Three or more shared URLs means one cluster and one page. One or two means Google treats them as separate topics, and a single page will not hold both.

2. 2

   Pick the cluster’s primary query

   The one that describes the intent most precisely, not the highest-volume one. It becomes the basis of the title and H1, while the remaining phrases distribute naturally across subheadings and body copy.

3. 3

   Discard clusters with no realistic path

   If the top ten is entirely aggregators with thousands of referring domains, shelve that cluster until year two. That is not defeat — it saves months of work that would have produced nothing.

4. 4

   Build the page map

   A four-column table: URL, primary query, intent type, status (exists / to create / to rewrite). That document, not the export, is the actual deliverable of keyword research.

5. 5

   Check for cannibalisation

   Two pages sharing a primary query compete with each other and both slide. Where the map shows duplicated intent, merge the pages and 301 the weaker one into the survivor.

Where to start when you have sixty clusters

Spreading effort evenly is the main reason SEO "does not work". Sixty clusters at equal priority produce sixty mediocre pages. The order that pays out fastest:

1. 1Queries where you already rank 8–20. Reworking an existing page takes a few hours and moves within two to four weeks, not six months.
2. 2Commercial clusters with modest demand and a weak top ten — where SEO revenue actually comes from, and where long pages with prices and examples earn their keep.
3. 3Local clusters, if you have a physical address. Least competition, best conversion into phone calls.
4. 4Informational articles — not indiscriminately, but those that lead logically to a specific commercial page. How to wire those connections is in our guide to internal linking.
5. 5Broad high-volume queries: not before your first year is behind you. Until then they only consume budget.

From there the keyword set becomes a content plan: every cluster gets a deadline, an owner and a format. How to keep that from degrading into text for its own sake is covered in our content marketing for SEO guide. The technical prerequisites, without which even perfect keyword work fails, are in the SEO checklist for 2026.

Frequently asked questions about keyword research

In short

• Sort by intent and realistic winnability, not by volume. High-volume queries are usually informational and already held by aggregators.
• The sharpest phrasing comes from Search Console, Google autocomplete and your sales conversations — not from a keyword planner.
• Clusters are defined by overlapping results: three shared URLs in the top ten means one page, one or two means separate topics.
• The deliverable of keyword research is a page map with URLs and primary queries, not a spreadsheet of phrases.
• Start with positions 8–20 and commercial clusters. Postpone broad high-volume queries by at least a year.