Keyword clustering: the three methods, how to set the threshold, and how to check the result

What is keyword clustering?

Keyword clustering is grouping a list of keywords so that each group can be targeted by one page. Keywords belong in the same cluster when one page can satisfy the people searching all of them, which usually means they share a search intent. The output is a set of clusters, each with a primary keyword and the variants that belong on the same page.

Clustering is the step between keyword research and a content plan. The structure it feeds, pillars and the cluster pages around them, is covered in the guide to pillar pages and topic clusters. A note on terms: a keyword cluster is a group of keywords for one page, while a topic cluster is a group of pages around a pillar. Clustering keywords comes first.

Why keyword clustering matters

Clustering prevents two expensive mistakes that keyword lists invite. The first is writing several pages for keywords that are really one intent, which splits the links and rankings that one page would have earned, and is the most common cause of keyword cannibalization. The second is the reverse: forcing keywords with different intents onto one page, which then satisfies none of them well.

It also changes how you size an opportunity. A keyword list sorted by volume overstates the number of pages you need and understates what each page is worth. Grouped into clusters, the same list shows how many pages to build and the combined demand behind each.

Keyword	Volume	Cluster	Page
content brief template	1,300	Content brief template	One page
content brief templates	260	Content brief template	Same page
seo content brief template	390	Content brief template	Same page
what is a content brief	880	Content brief definition	A different page, or a section of the guide
content brief example	320	Content brief template	Same page as the template
content brief vs creative brief	140	Content brief vs creative brief	A section, unless the results page says otherwise

Hypothetical keywords and volumes to show the method

The three methods of keyword clustering

Keyword clustering tools use one of three methods, or a mix. Knowing which one a tool uses tells you what it will get wrong.

By shared words

Strengths

• Fast, free and runs on any list, with no data beyond the keywords
• Reliably groups plurals, word-order changes and close variants
• Transparent: you can see why two keywords were grouped

Limits

• Cannot see synonyms: ‘CRM for startups’ and ‘customer database for new companies’ stay apart
• Groups keywords that share words but not intent, such as ’email template’ and ’email template builder’

A first pass on any list, and the only option without a results-page source.

By results-page overlap

Strengths

• Groups by what Google actually ranks, so synonyms with the same intent end up together
• Separates keywords that look alike but have different intents

Limits

• Needs a results page for every keyword, which costs money and time on large lists
• Results change, so clusters drift; mixed-intent results pages produce unstable groups

The deciding method for keywords you will build pages for.

By meaning

Strengths

• Uses language models or embeddings to group keywords with similar meaning, including synonyms
• No results-page cost

Limits

• Similar meaning is not the same as the same intent: ‘buy running shoes’ and ‘best running shoes’ are close in meaning but may rank different pages
• Harder to explain why two keywords were grouped

A supplement for finding synonyms, checked against the results page.

Clustering by shared words

Word-based clustering, also called lexical clustering, reduces each keyword to its main words, often with stemming or lemmatisation so that ‘template’ and ‘templates’ match, removes filler words, and groups keywords whose word sets overlap enough. Wikipedia’s article on keyword clustering describes lemmatisation-based grouping as the method that came before results-based clustering. The Topical Map Generator uses this method: it takes keywords highest volume first, adds each to the cluster it shares the most words with if the overlap clears a threshold you choose, and starts a new cluster otherwise, so clusters form around real demand rather than the order of your list.

Clustering by results-page overlap

Results-based clustering, usually called SERP clustering, searches each keyword, records the top results, and groups keywords whose results share enough URLs. The reasoning, as Ahrefs’ guide to keyword clustering puts it, is that keywords sharing many of the same top-ranking pages likely share intent and can be targeted on one page. According to Wikipedia, the approach and its first principles were set out by Alexey Chekushin in Russia in 2015, the same year the first tool built on it appeared there.

Clustering by meaning

Meaning-based clustering, often called semantic clustering, converts each keyword into a numerical representation of its meaning, an embedding, and groups keywords whose representations are close. It finds synonyms that word overlap misses without paying for results pages. Its weakness is that it measures meaning, and two keywords can mean nearly the same thing while the people searching them want different pages. Use it to suggest candidates, then decide with the results page.

How results-page overlap works, and how to set the threshold

Results-page clustering needs two decisions: how many results to compare, and how many shared URLs make two keywords one cluster. Both change the answer.

Most tools compare the top ten organic results, though some use fewer. Keyword Insights’ documentation says it now examines the top seven, because Google no longer consistently shows ten traditional links in the top ten, and groups keywords when they share 40% or more of their URLs, a threshold you can adjust. Wikipedia calls the minimum number of shared results the clustering level, and notes the trade-off: a higher level produces more groups with fewer keywords each, and a low level of one or two produces fewer, larger groups.

Shared URLs in the top ten	Effect	When to use it
2 or 3	Large clusters that can mix related intents	Grouping a list into topics for a hub or pillar, not into pages
4 or 5	Clusters that usually match one page	Deciding which keywords one page should target
6 or more	Small, tight clusters, more pages than you need	Checking whether two near-identical pages should be merged

A working guide, not a published standard. The four-URL rule is the one used on this site

The working rule used across this site, set out in the keyword cannibalization guide, is four or more shared URLs in the top ten meaning one page. The exact number matters less than applying the same number every time, and checking by hand the pairs that fall one URL either side of it.

How a cluster is linked together

There is a second choice that tools make quietly: whether every keyword in a cluster must overlap with every other, or only with one keyword in the group. Wikipedia describes three versions: each keyword linked to the highest-volume keyword in the group, each linked to at least one other keyword, or all keywords sharing the same matching URLs. The first two produce larger clusters that can drift: keyword A overlaps with B, B with C, but A and C have little in common. The third is strict and safe. If your tool lets you choose, use the strict version for page planning.

How to cluster keywords, step by step

The method below works with any tool, or a spreadsheet for a small list. It uses word clustering for speed and results-page overlap for the decisions that matter.

1. 1CollectExport the keyword list for the topic with volume and difficulty. Remove irrelevant and branded terms.
2. 2LabelMark each keyword’s intent, at least informational, commercial, transactional or navigational.
3. 3GroupRun a word-based clustering pass to group variants and plurals.
4. 4CheckFor keywords you will build pages for, compare the top ten results and apply your threshold.
5. 5ResolveDecide the borderline and mixed-intent pairs by hand, using the decision tree below.
6. 6NamePick each cluster’s primary keyword: usually the one with the most volume that describes the whole group.
7. 7MapAssign each cluster one URL, existing or planned, in a keyword map.

The Intent Classifier handles the labelling step for a whole list, with the reason for each label, and the Topical Map Generator handles the grouping, naming and mapping steps from an exported list, including matching clusters to pages you already have and flagging likely cannibalisation.

Choose the primary keyword carefully. It becomes the page’s target query and usually shapes its title, so it should describe the whole cluster, not just carry the most volume. ‘Content brief template’ is a better primary keyword for a cluster about templates than ‘content brief’, even if the latter has more volume, because the shorter term has a different intent.

When to put two keywords on one page: a decision tree

Most clustering decisions are easy. The hard ones are the pairs near the threshold and the keywords whose results page shows mixed intent. This tree settles them.

Do the two keywords share at least your threshold of URLs in the top ten results?

Yes

Do the shared pages answer both keywords fully, rather than ranking for one as a side effect?

Yes

One page. Make the higher-volume keyword that describes both the primary, and cover the other as a variant.

No, they rank for one only in passing

Probably two pages, or one page with a section for the smaller keyword. Look at what the top results do for each.

No

Does the results page show mixed intent, such as guides and product pages together?

Yes, mixed

Decide by your business: build the page type you can make best and that serves your goal, and expect a less stable ranking. Recheck the results page in a few months.

No, the results are consistent but different

Two pages. If one keyword is small, it may still be better as a section of a related page than a page of its own.

Use the same threshold for every pair. Pairs one URL either side of it are the ones worth checking by eye

Mixed-intent results pages deserve a note. When Google ranks guides and product pages together, it is hedging because searchers want different things. Neither page type is wrong, but the ranking of either is less stable, and the results page can shift towards one type over time. Note mixed-intent clusters in the keyword map, so they are rechecked rather than forgotten.

What keyword clustering costs

Word and meaning-based clustering are nearly free to run. Results-page clustering costs a results-page lookup for every keyword, every time you refresh it, and checking pairs by hand grows much faster than the list. The calculator puts those costs side by side.

Hypothetical defaults. Results-page data prices vary widely by provider and volume.

Keywords in the listCost per 1,000 results-page lookups$Times you re-cluster a yearShare of pairs near the threshold that need a manual check%Minutes to check one pair by handminutes

Keyword pairs to compare

1,999,000

Every keyword against every other

Lookup cost per run

$6.00

Lookup cost a year

$24.00

Pairs near the threshold, per run

19,990

Hours to check those pairs by hand, per run

666.3

This is why you check only the pairs you will build pages for

Tools avoid comparing every pair by grouping candidates first, usually by shared words, which is why a word-based pass before results-page clustering saves time and money.

The lookup cost is rarely the obstacle; the manual checking is. With 2,000 keywords there are nearly two million pairs, and even 1% of them is far too many to check by eye. That is why the step-by-step method above checks results-page overlap only for the keywords you will build pages for, and resolves by hand only the pairs that matter.

From clusters to a keyword map

Clusters become useful when each one becomes a row in a keyword map: one cluster, one URL, every variant listed beside it, and the intent and page type noted. The map is what stops the same intent getting a second page months later.

Cluster and primary keyword	Variants	Intent and page type	URL
content brief template	content brief templates, seo content brief template, content brief example	Commercial, template page	/templates/content-brief
what is a content brief	content brief meaning, content brief definition	Informational, guide	/blog/content-brief
content brief tool	content brief generator, ai content brief	Commercial, comparison	Planned

Hypothetical rows to show the format

Every new brief starts by checking its target query against the map. If the query is already in a row, the work is an update to that row’s URL. How to turn a row into a brief a writer can work from is in the content brief guide.

Clusters also group into topics. Clusters that share a head term usually belong under one pillar, and the set of topics you cluster deeply is a choice about where you want to be a recognised source, which is the subject of the guide to topical authority.

How to check a clustering result before you use it

A clustering tool’s output is a draft. Checking it takes an hour for a few hundred keywords and saves building pages that compete or miss.

• Read the largest clusters. Large clusters are where chaining joins keywords that do not belong together. Split any cluster whose keywords would need different page types.
• Read the singletons. Keywords left alone are often synonyms the method missed. Check whether any belong in an existing cluster.
• Check the primary keywords. Each should describe the whole cluster and suit a title.
• Spot-check overlap. For five or ten clusters you will build first, compare the results for the primary and two variants by hand.
• Compare with what ranks today. For clusters where you already have a page, check in Search Console that the page ranks for the variants. If another page of yours does, the cluster may be split already.
• Mark mixed intent. Note clusters whose results page mixes page types, and plan to recheck them.

Keeping clusters current

Results pages change, so clusters built on them drift. A cluster that was one intent last year can split when Google starts ranking different page types for its variants, and new keywords appear in every topic. Re-cluster on a schedule and when something changes.

Trigger	What it can mean	What to do
Every quarter or half-year	Routine: results pages drift slowly	Re-run results-page overlap for the clusters you have built pages for, and compare with the last run
A page loses rankings for some of its variants	The cluster may have split into two intents	Check whether those variants now rank different pages; the cluster may need re-drawing
Two of your pages start trading places for a query	Two clusters may have merged	Check whether their clusters have merged; one page may need to absorb the other
New keywords in a topic	Demand has shifted or grown	Assign them to existing clusters or start new ones, before anyone writes a page for them

When a cluster has split or merged, act on it with the rules in the keyword cannibalization guide. Ahrefs’ guide to cannibalization makes the point that pages ranking for overlapping keywords are not always a problem; act when the overlap is costing performance, not just because it exists.

Common keyword clustering mistakes

Most clustering mistakes come from trusting one method too much. Each item below has the check that catches it.

• Word clusters treated as page plans. Check: have the clusters you will build pages for been checked against the results page?
• No stated threshold. Check: is the overlap threshold written down and applied the same way every time?
• Chained clusters. Check: does every keyword in a large cluster overlap with the primary, not just with a neighbour?
• Primary keyword chosen on volume alone. Check: does it describe the whole cluster, and does it share the cluster’s intent?
• Mixed-intent clusters forgotten. Check: are they marked for rechecking?
• Clusters never become a map. Check: does every cluster have exactly one URL, and do briefs check against it?
• Clustering once. Check: is there a date for the next re-cluster?

Prepare

Keyword list cleaned of irrelevant and branded terms Intent labelled for every keyword Overlap threshold chosen and written down

Cluster

Word-based pass run to group variants Results-page overlap checked for keywords you will build pages for Borderline and mixed-intent pairs decided by hand Largest clusters and singletons read and corrected

Use

Each cluster has a primary keyword that describes the whole group Each cluster has exactly one URL in the keyword map New briefs checked against the map before writing Re-clustering scheduled

Ticks are saved in your browser.

Running clustering, the keyword map and the brief check as one system, so that a team can add pages for years without them competing, is part of an SEO systems engagement. It is also the kind of fix the SEO audit guide argues for: change the process that creates duplicates, rather than cleaning them up after each audit.