Keyword research: how to pick the right phrases

Keyword research is the study of the language customers use to describe a problem, then mapping those questions to specific pages. Good research doesn't end with a spreadsheet of a thousand phrases. It ends with a decision: what to write about, what to skip, and which URL should answer a given query.

This matters because two pages targeting the same question don't automatically reinforce each other. They often compete, dilute internal linking, and make it harder for search engines to pick the right answer. That's why phrase research is part of content, offer, and site structure planning — not a checkbox to tick before publishing.

A keyword is not a command to the search engine

A phrase shows how someone formulates a need. It doesn't yet say what should be on the page. The same word can hide several intents:

• "RAG" — quick definition or attempt to understand the concept;
• "RAG in business" — question about application and risk;
• "RAG implementation price" — comparing solutions and vendors;
• "RAG vs fine-tuning" — deciding between approaches.

A shared root doesn't justify one overloaded page. First, check whether the user expects a definition, instructions, a comparison, a product, or a local service. Only then decide if an existing article meets that need, needs expansion, or a new page should be created.

Start from customer problems, not from the tool

The best sources for your initial list are usually already at hand. Gather them before opening any keyword planner:

1. questions from sales calls, briefs, and emails;
2. names of processes, products, and documents the client uses;
3. Search Console queries leading to existing pages;
4. questions the team answers repeatedly during implementations;
5. language of competitor offers — not to copy, but to spot missing questions.

This list is valuable even when some phrases have low volume. Someone asking in detail may be closer to a decision than a thousand users typing a generic term. In B2B, it often pays more to answer one real question precisely than to chase a broad phrase with no clear intent.

Group phrases into clusters, not a flat list

A cluster ties questions around one problem and separates page roles. This lets the reader move from basics to a decision, while each publication has its own job.

User question	Content role	Target page type
"what is SEO"	understand basics	pillar article
"keyword research"	do a specific job	satellite guide
"SEO vs Google Ads"	choose a channel	comparison / decision guide
"organic positioning for business"	evaluate support	service page

In this silo, What is SEO is the pillar article. This text expands on keyword research, while SEM and the differences between SEO and PPC answers the question about organic vs paid traffic roles. Such division matters more than publication count.

How to judge if a phrase is worth the effort

Popularity helps, but isn't the only criterion. For each candidate, answer five questions:

Criterion	Question	Warning signal
Fit with offer	Does answering this question demonstrate the competence the client needs?	phrase brings traffic but leads to a topic outside the company profile
Intent	What does the reader want to do after reading?	one page tries to define, compare, and sell simultaneously
Business value	Does the topic attract the right audience or strengthen a key pillar?	decision based solely on volume
Ability to add something original	Do we have experience, data, or a useful way to explain?	text would only summarize search results
Page owner	Which page best answers this need?	two URLs have nearly the same title and promise

The last point protects against cannibalization. When an existing page already has the right intent, it's usually better to update and strengthen it with links than to create a competing article from scratch.

Which data source answers which question

Tools aren't interchangeable. Each sees a different slice of demand, so sound research combines at least first-party data with observation of search results.

Source	Use case	What you won't learn from it
Conversations, briefs, site search	customer language and near-purchase problems	full market picture
Google Search Console	queries and pages already getting impressions	all searches or exact demand outside your site
Keyword Planner	variants, seasonality, signals useful for ads	guaranteed traffic or organic rankings
Search results	dominant intent and answer format	profitability of the topic for your business
Third-party SEO tool	topic comparison and rough estimates	certain number of future customers

If two tools show different numbers, don't pick the convenient one. Check the metric definition, country, language, device, and period. For editorial decisions, direction and intent alignment usually suffice; false precision won't fix a poorly chosen topic.

Keyword research process in five steps

1. Define the decision scope

Write down which product, service, or problem you're working on. "SEO" is too broad. "How a service company builds organic traffic without a fixed ad budget" already sets a topic boundary and lets you discard peripheral phrases.

2. Build the input list

Combine customer language, your own pages, Search Console data, and problem names. Don't filter too early. At this stage you're collecting variants, questions, comparisons, abbreviations, and natural phrasings — including those that don't sound like "ideal" marketing phrases.

3. Expand and organize the list

Google Keyword Planner may suggest related queries, but its data is designed primarily for ad campaigns. Treat it as a hint about language and seasonality, not a forecast of SEO sales. Google explains that forecasts rely on historical data, budget, bids, and ad quality. Keyword Planner forecasts documentation

Also add queries that already show your page. Search Console reveals both expected and unexpected phrases, though some data is anonymized or truncated — don't treat the report as a complete picture of all searches. Search Console query data limitations

4. Check results, not just numbers

Type the most important phrases into the search engine and examine the dominant answer types. Are they guides, product cards, local businesses, videos, comparisons, or definitions? This is a quick intent test. If results answer a completely different question than your planned page, change the topic or format instead of trying to "convince" a phrase with a mismatched text.

5. Assign the phrase to a specific URL

The final artifact isn't a list, but a map: topic → intent → target URL → proof material → internal links. Only then can you create an article brief or improve an existing page without duplication risk.

An example row of such a map might look like this:

Audience problem	Intent	Answer owner	Proof and next step
"How to validate an idea before building an app?"	informational, nearing decision	/blog/discovery-produktu-przed-budowa-aplikacji	process description → /oferta/projektowanie-planowanie

This record immediately exposes three common mistakes: creating a second article for the same need, sending an educational question straight to an aggressive sales page, and publishing without proof or a next step. In a small site, a simple spreadsheet map is enough; what matters is one address owning one primary intent.

Long tail, branded phrases, and questions - how to use them

The split into head terms, long tail, and branded phrases is still useful, but don't treat it as a quality ranking.

• Head terms help name a broad topic and often fit pillars. They're competitive, so they need comprehensive material.
• Long-tail phrases describe a specific case. They often fit well as sections in a guide, comparison, or satellite article — unless they deserve a standalone page.
• Branded phrases show how people search for your brand and products. They won't replace visibility for new needs, but help control whether your offer and naming are understood.
• Natural-language questions are especially valuable in the era of AI-generated answers. This isn't a reason to write fake FAQs; it's material for real, standalone sections with clear answers.

Where AI helps and where it gets in the way

An assistant can quickly group hundreds of variants, catch synonyms, or suggest cluster gaps. It shouldn't decide on its own that a phrase has purchase intent or invent volume it doesn't see in data. The best division of labor is simple: the tool speeds up organization, the human verifies results, picks the page owner, and owns the text's thesis.

AI also doesn't replace checking current results. The model may remember an outdated market layout, mix ad language with organic, or propose questions nobody asks. Treat its list as verification material: confront every important phrase with your own data, the search result, and business sense.

We develop this process in five principles of content creation. If you already have lots of materials and don't know which URLs to strengthen, start from the basics in What is SEO or from organizing within SEO and organic marketing.