Keyword Research That Starts With Your Own Data

Most keyword research starts in the wrong place: a tool, showing estimated volumes for a market you have no position in.

Start instead with Search Console, which shows queries your site already appears for, how often, and roughly where. That is measured data about you rather than an estimate about a market — and no competitor can see it.

The highest-value finding in keyword research is usually already in your account: the queries where you rank eighth to twentieth. You are relevant enough to be shown and not visible enough to be clicked. A page that already half-works is a far better investment than a page that does not exist.

---

What volume figures actually are

Worth being precise about, because the entire industry treats these numbers as measurements.

Google's Keyword Planner figures are rounded averages, not counts. Google's own documentation says so: "Your search volume statistics are rounded. This means that when you get keyword ideas for multiple locations, the search volumes might not add up as you'd expect."

They are averages over a date range, defined as "The average number of times people have searched for a keyword and its close variants based on the month range as well as the location and Search Network settings you selected" — and Google notes that "web traffic is influenced by seasonality, current events, and a number of other factors."

And third-party tools do not have Google's data at all. Their volumes are models built from clickstream panels and other signals, calibrated against whatever they can. Difficulty scores are entirely their own invention — useful for ranking one keyword against another inside the same tool, meaningless as absolutes.

None of this makes them useless. It makes them a compass rather than a map. Relative comparison is what they are good for; precise planning is not.

---

What Search Console gives you, and its limits

Real data, with documented gaps you should know about.

The gaps, from Google:

• Rare queries are removed. "Tables in the performance reports omit rare queries to protect user privacy." So the long tail you most want to see is partly invisible
• The table caps out. "Our tables can show a maximum of 1,000 rows, so some rows might be omitted." On a site of any size, filter rather than scroll
• Totals will not add up, for both reasons above. That is expected behaviour, not a bug

And "position" is an average of a topmost result across many searches, personalised and localised differently for every person. Treat it as a band, not a rank.

Even with those limits it beats every alternative, because it is your actual performance rather than a model of a market. Reading it properly is covered in Google Search Console reports, read properly.

---

The four questions research has to answer

Not "what is the volume". These:

Question	Why it decides things
What does the searcher want?	Determines the page format entirely — and whether you can win at all
Does a page for this already exist on my site?	Two pages competing is worse than one page ranking
Can I say something the current results do not?	If not, you are producing a fourth identical page
Does anything happen if they arrive?	Traffic that cannot convert or inform is a cost

The first is the one most research skips, and it is the one that decides the format, the depth and the winnability of the page. Search intent is covered in full in search intent: the four types and what they change.

The second prevents the most common self-inflicted damage in SEO, which is two of your own pages competing for the same query — see keyword cannibalisation.

---

The method

Seven steps.

Step 1 — Export what you already rank for

Search Console → Performance → Queries, last three months, and export.

Then sort by impressions and filter to positions 8 to 20. That band is the opportunity list: Google already considers you relevant, and you are below where clicks happen.

Read the queries themselves, not just the numbers. You will find questions you did not know you were being shown for, and phrasing you would never have guessed.

Step 2 — Separate "improve" from "create"

For each query in that band, ask which page is ranking and whether it is the right page.

• Right page, ranking low → improve it. Faster and more reliable than anything else on this list
• Wrong page ranking → you have a relevance or an internal linking problem, not a content gap
• No page at all → a genuine gap, and a candidate

Most sites find more improve-work than create-work, and improve-work pays off sooner. The method is in refreshing old blog posts.

Step 3 — Classify the intent before writing anything

Search the query and look at what Google is actually returning.

The results tell you the intent more reliably than any tool's classification. If every result is a product page, an article will not rank. If every result is a tutorial, a product page will not.

And check whether the answer is being given on the results page itself. If it is, the click may not happen regardless of your position — which changes whether the keyword is worth targeting at all.

Step 4 — Group into topics, not lists

Twenty variations of the same question are one page, not twenty.

The grouping test: would the same page satisfy both searchers? If yes, one page. If no, two — and they need to be genuinely different, not two angles on the same thing.

This step is what prevents cannibalisation, and it is why a cluster is planned as a cluster rather than assembled from a keyword export. See internal linking and topic clusters.

Step 5 — Check you have something to say

The honest gate.

Read the top five results properly. If you cannot name what your page would add — original data, direct experience, a correction, a clearer explanation — the page will not earn its place, and Google's own guidance names this pattern directly. Its self-assessment questions include: "Are you producing lots of content on many different topics in hopes that some of it might perform well in search results?" and "Is the content primarily made to attract visits from search engines?"

These are not rhetorical questions. Producing pages that fail them at scale is what Google's spam policy calls scaled content abuse: "when many pages are generated for the primary purpose of manipulating search rankings and not helping users."

Step 6 — Sequence by winnability, not volume

A page that ranks third for a small query beats a page that ranks fortieth for a large one.

Order candidates by:

• Existing position, if any — the 8-to-20 band first
• Specificity — narrower questions have less established competition, covered in long-tail keywords
• Whether you have unusual authority on that particular question
• Whether the traffic does anything for you once it arrives

Volume comes fourth, deliberately.

Step 7 — Write for the person, then check the phrasing

In that order.

Use the searcher's own words — the exact phrasing from your Search Console export, which is how real people ask, not how marketers write. Then stop.

Google on word count: "Are you writing to a particular word count because you've heard or read that Google has a preferred word count? (No, we don't.)"

And on repetition, the spam policy defines keyword stuffing as "filling a web page with keywords or numbers in an attempt to manipulate rankings", with examples including "blocks of text listing cities and regions a page is trying to rank for" and repeating words unnaturally.

The practical rule: if you would not say it aloud to a customer, take it out.

---

What has changed, and what has not

What has not: people still type questions, and pages that answer them well still get found. The fundamentals of this discipline have been stable for a decade.

What has: more answers are now given on the results page or by an AI system, which means a ranking position no longer guarantees a visit. That does not make keyword research obsolete — it changes which keywords are worth having.

Two adjustments follow.

Value queries by what happens when someone arrives, not by volume. A query with modest volume where the searcher needs something you sell or know is worth more than a large informational query answered in the results.

And accept that being cited without being clicked has value now. Appearing as a source in a generated answer builds recognition even without a session. The mechanics are in AI search optimisation and tracking AI search traffic.

---

The mistakes that cost most

Starting with a tool. Your own data is better, free, and invisible to competitors.

Treating volume as truth. It is a rounded average of a model, and Google says as much about its own figures.

One page per keyword variation. This is how sites build cannibalisation deliberately and then wonder why nothing ranks.

Ignoring intent because volume looked good. You cannot rank an article where Google is returning products.

Writing to a word count. Google has stated it has no preferred length.

Publishing without an answer to "what does this add". At scale this is the pattern Google's spam policy names.

Never checking what actually happened. Research is a hypothesis; Search Console three months later is the result.

---

Frequently asked questions

How do I do keyword research for free?
Start with Search Console's Performance report, which shows queries your site already appears for. Filter to positions 8 to 20 — those are queries where Google considers you relevant but you are not visible enough to be clicked, and they are the best available opportunities.

Are keyword search volumes accurate?
They are estimates. Google's own Keyword Planner figures are rounded averages over a date range, and Google notes they may not add up as expected. Third-party volumes are models built from panel data. Use them to compare keywords, not to forecast traffic.

What is the best starting point for keyword research?
Your own Search Console data, because it is measured rather than estimated and no competitor can see it. Tools are better for exploring markets you have no presence in yet.

How do I know what a keyword's search intent is?
Search it and look at what Google returns. If every result is a product page, that is the intent, and an article will not rank there. The results page is a more reliable classifier than any tool's intent label.

Should I create a separate page for every keyword?
No. Variations of the same question belong on one page. The test is whether the same page would satisfy both searchers — if yes, it is one page, and splitting them causes your own pages to compete.

Does keyword research still matter with AI search?
Yes, but what makes a keyword valuable has changed. A ranking no longer guarantees a click, so queries are better valued by what happens when someone arrives than by volume alone.

How long should an article be for SEO?
There is no target. Google's own guidance asks whether you are writing to a particular word count because you heard Google has a preferred one, and answers plainly that it does not.

Why doesn't Search Console show all my queries?
Google omits rare queries from performance tables to protect user privacy, and tables show a maximum of 1,000 rows. Both are documented, which is why chart totals will not match the sum of the rows.

---

What to do next

Open Search Console, go to Performance → Queries, set the range to the last three months, and export it. Sort by impressions, then look only at rows where average position is between 8 and 20.

Read those queries. That list is your keyword research, it is specific to you, and it took four minutes.

---

Related guides

• Search intent: the four types and what they change — the question research must answer first
• Long-tail keywords — where a small site can actually win
• Keyword cannibalisation — when your own pages compete
• Keyword research without paid tools — the full free method
• Google Search Console reports, read properly — the source data
• Internal linking and topic clusters — turning groups into structure
• Refreshing old blog posts — the improve-work that pays first
• On-page SEO — what happens after the research