SEO · 9 min read
Keyword research, done as a sequence of decisions
Most keyword research produces a spreadsheet nobody opens twice. This is the method that produces decisions instead: where demand evidence genuinely comes from, how to read a results page to find out what a query is worth, and what to leave out on purpose.
Written by Zubair Afzal, FounderUpdated
The point of the exercise
This is a filtering job, not a collection job
Most advice on how to do keyword research stops after the collection half. You open a tool, type a seed phrase, export what comes back, and end up with a file containing more rows than your business will publish pages in five years. The file feels like progress because it is large. It is not progress, because nothing in it says what to do on Monday.
The half that matters is the filtering. Out of everything people type, a small number of queries describe a person your business can help, at a moment when they would accept help, in a format you are able to produce well. Research is the process of finding that small number and being able to say, in one line each, why everything else was left out.
That reframing changes the order of the work. You do not start with the tool. You start with what you sell and who buys it, then use the tools to find the language those people use and to check whether the results for that language are open to a business like yours.
Start where the evidence is real
The queries you already have evidence for
Search Console reports what people actually typed before your pages appeared. It is the only demand data about your business that is measured rather than modelled, and it is usually the last place people look.
Before any tool is opened, sort your existing Search Console queries into these three groups. Each implies a different action, and the first two are the cheapest work available to most sites because eligibility has already been demonstrated.
Impressions, almost no clicks
Demand exists and you are already eligible. The gap is usually format, specificity or the title, not authority.
- phrases where you appear low on the first or second page
- question-shaped queries your page answers in its fourth section
- comparison phrases where a competitor holds the top result
- queries whose landing page is a category listing rather than an answer
Clicks arriving at the wrong page
A mapping problem. Somebody has to decide which single page is the answer, then make that decision visible in the linking.
- a buying query landing on a blog post
- a support query landing on a sales page
- a city query landing on a national page
- two of your own pages alternating for one phrase
Queries you would rather not have
Worth naming explicitly. Traffic from these is not a win, and letting it inflate a report is how a programme loses credibility with finance.
- searches for a free version of what you sell
- searches for jobs at companies like yours
- searches from countries you do not serve
- searches by students and researchers rather than buyers
These are examples of how customers in this market search, drawn from keyword research and from the questions that come up on sales calls. They are illustrative, not a volume claim — the actual demand in your area is something we size before recommending anything.
The method
Six steps, in this order
The order matters more than the tooling. Each step narrows the field using a different kind of evidence, and skipping one produces a list nobody can act on.
Write down what you sell and who buys it
One line per offering: what it is, who it is for, what it replaces, and what a good customer for it looks like. This is the filter every later step runs against. It sounds obvious and it is skipped constantly, which is why so many content plans are full of subjects the business cannot convert.
You get: A one-page description of offerings and buyers
Collect the language, not the keywords
Take phrasing from places where real people wrote it: sales call notes, the subject lines of enquiry emails, support tickets, your own site search log, review text, and the questions salespeople answer weekly. Then add autocomplete, related searches and the questions shown on the results page for your obvious seed terms. You are building a vocabulary, not a shortlist.
You get: A vocabulary list grouped by where it came from
Expand, and get an order of magnitude
Now open the estimate tools. Keyword Planner reports average monthly searches for advertisers; Google Trends shows relative movement and seasonality rather than counts. Use them to sort the vocabulary into rough bands — meaningful demand, thin demand, effectively none — and do not treat any single figure as precise enough to decide between two similar phrases.
You get: The vocabulary sorted into three demand bands
Read the results page for every candidate
This is the step that changes plans. For each phrase under serious consideration, look at what currently ranks. Note the format, the type of organisation, whether the pages are national or local, and whether the top of the page is occupied by ads, a map, a shopping row or a generated answer. You are finding out what the engine has decided the query means and who it believes should answer it.
You get: A short note per candidate: format, incumbent type, page furniture
Group by what the searcher wants
Put phrases together when a person typing any of them would be satisfied by the same page, and split them when they would not, even if the words are nearly identical. Grouping by string similarity is fast and wrong. Comparing the actual results pages is slower and correct, because two queries returning substantially the same results are being treated as one question by the engine.
You get: Groups, each with one question stated in plain language
Assign, reject, and record the reasons
Every surviving group gets one page — existing or new — and one owner. Everything else goes in a rejected column with a one-line reason: no commercial fit, wrong searcher, results page closed to us, format we cannot produce well, or demand too thin to justify the effort. The rejected column turns out to be the most reused part of the document.
You get: A question-to-page map, plus a rejected list with reasons
Evidence
Where demand evidence comes from, and what each source is
These are not interchangeable. Two of them measure something real, two are models, and one is not about volume at all.
| Dimension | What it actually reports | What it cannot tell you | Best used to decide |
|---|---|---|---|
| Search Console | Your own impressions, clicks, average position and the queries that triggered them, measured rather than estimated. | Anything about queries you have never appeared for, and the complete query list — rare queries are withheld. | Which existing pages to improve first, and which questions you are already close to owning. |
| Keyword Planner | Average monthly searches, built for advertisers estimating reach, with close variants grouped together. | How commercially valuable a query is to you, or how that average is distributed across the year. | Rough order of magnitude, and whether a phrase carries any advertiser demand at all. |
| Google Trends | Relative interest over time and by region, indexed rather than counted. | Absolute volume, and anything about queries below its reporting threshold. | Seasonality, whether a subject is growing or fading, and regional differences. |
| The results page itself | Who currently ranks, in what format, and what the engine has decided the query means. | How many people search it, and whether those people buy anything. | Whether the query is realistically open to a business like yours, and what shape the page has to be. |
| Your own business records | The words customers use, the objections they raise, and which enquiries turned into revenue. | How many strangers use those same words in a search box. | Which questions are worth answering even when the estimated volume looks unexciting. |
What to skip
Things that look like keyword research and are not worth the hours
Each of these is common, defensible-sounding, and produces no decision at the end. Cutting them is usually what makes the research finishable.
- Chasing a target keyword density. There is no ratio to hit, and writing towards one reliably makes the page worse to read, which is the thing being measured all along.
- Treating a third-party difficulty score as a verdict. It models link counts; it does not judge whether you can answer the question better than what currently ranks. Read the results page instead.
- Building an exhaustive long-tail export. Thousands of near-identical phrasings rarely represent thousands of opportunities. They usually represent one page with a broad enough answer.
- Making a separate page per phrasing. That is how a site ends up with several pages competing for one question and none of them clearly being the answer.
- Planning for queries you cannot serve. A national page written for a query where every result is a nearby business is effort spent on a competition you were never entered into.
- Re-running the export because six months have passed. Demand language moves slowly. Your own Search Console data moves weekly, and it is the file that deserves the recurring review.
- Arguing about the volume figure. If a decision turns on whether a phrase gets four hundred or seven hundred searches a month, the decision does not really turn on the volume figure.
Where this leads
What sits either side of the research
Questions
Questions people ask while doing this
Do I need a paid keyword tool?
Not to start. Search Console gives you your own measured query data, Keyword Planner gives you advertiser-facing estimates, autocomplete and related searches give you phrasing, and the results page itself gives you the competitive picture. Those four cost nothing and cover most of a first pass.
Paid tools earn their money on breadth and speed: seeing what competitors rank for, exporting large sets, and tracking movement over time. They do not make the underlying estimates more true, and they cannot tell you which queries your sales team values.
How accurate are search volume numbers?
They are modelled estimates rather than counts, and they are best read as an order of magnitude. Different tools disagree because they are built on different panels and extrapolations.
The practical response is to stop using the number as a tiebreaker. Use it to separate hundreds a month from a handful a year, then decide on evidence that is actually reliable: who is searching, what the results page rewards, and whether the query maps to something you sell.
Should one page target one keyword?
One page should answer one question. That question is usually expressed through a family of phrasings, and all of them belong on the same page.
The test is not whether the words differ. It is whether a person typing each phrase wants the same thing. "Cost of a new boiler" and "how much is a boiler installation" want the same page. "Boiler installation cost" and "boiler installers near me" do not, because the second wants a list of suppliers and the first wants a number.
How many keywords should be in the plan?
Fewer than you expect, each attached to a named page and a named owner. Thirty questions that will genuinely be answered well beats eight hundred rows that will not.
The useful discipline is keeping the rejected column: every phrase you decided against with one line saying why, so the decision does not have to be re-argued when somebody new joins and reruns the export.
What if everything relevant is dominated by big publishers?
Then that layer of the market is not currently open to you, and the honest move is to go narrower rather than to write a better version of the same article. Narrower usually means more specific: a segment, a constraint, a geography, a use case, a price point.
The other option is to compete on a different axis altogether. Original data, a working tool, or genuine first-hand experience are three things a large publisher covering a subject at arm’s length cannot easily reproduce.
Bring your query data and we will read it with you
The fastest version of this conversation is opening your own Search Console export together and separating the questions you are already close to owning from the ones that were never available. Forty minutes, and you keep the notes.
Related
Where to go next
Last updated