Keyword clustering methods fall into two camps: SERP overlap, which groups keywords Google answers with the same pages, and semantic, which groups keywords that mean similar things. Use SERP overlap to decide pages and semantic grouping to sort big lists cheaply.
By Kristijan GjorgjioskiIndependent SEO consultantPublished 18 October 2026 · 8 min read
You have a keyword list with thousands of rows and one question: which of these belong on the same page?
That is the job of keyword clustering. And the method you pick decides whether your answer is right.
I build clustering into my own keyword pipelines, and I have watched both main methods fail in different ways. Here is how they compare.
If you only read one section, make it this one.
SERP overlap clustering is the more accurate method for deciding which keywords share a page. It groups keywords by the URLs Google ranks for them, so it reflects how Google actually splits intent.
Semantic clustering is the faster, cheaper method for sorting a large list into topics. It groups keywords by meaning, which helps you organise a list but tells you nothing about how the results behave.
So the honest answer is not "pick one".
Use semantic clustering to get your messy list into rough topic families.
Then use SERP overlap to decide which keywords in each family share a page and which need their own.
The rest of this guide explains why, with the trade-offs you will hit when you run this on a real site.
There are four methods in common use. Two of them do most of the work:
| Method | Groups keywords by | Data you need | Relative cost | Best for | Main blind spot |
|---|---|---|---|---|---|
| SERP overlap | Shared ranking URLs in the top results | A live results pull per keyword | Highest | Deciding pages | Unstable or local results |
| Semantic (embeddings) | Similarity of meaning | Keyword text plus an embedding model | Low | Sorting big lists into topics | How Google splits intent |
| Lexical | Shared words or word stems | Keyword text only | Lowest | A quick first pass | Synonyms and word order |
| Manual | Your judgement | Time and market knowledge | Time-heavy | Short, high-value lists | Consistency at scale |
The first two rows are the ones that matter.
Lexical grouping is a weaker version of semantic grouping.
Manual grouping is what you do when your list is short enough to read. It is also how I start research for a small business (see my guide to keyword research for a service business).
SERP overlap clustering compares the top results for each keyword. If two keywords share enough ranking URLs, they go in the same cluster.
The logic fits in one line:
If Google ranks the same pages for two searches, one page can probably rank for both.
Here is the process:
That primary term is the one your page's title tag and H1 should carry.
You can apply the threshold in two ways.
Hard clustering requires every keyword in a cluster to share enough URLs with every other keyword. Clusters come out small and tight.
Soft clustering only requires each keyword to share enough URLs with the cluster's main keyword. Clusters come out bigger and looser.
Hard clustering is safer for service pages, where a wrong merge can cost you a page that converts.
Soft clustering suits blog topics, where one broad article can answer a wider set of questions.
The threshold is how many shared URLs you treat as "the same intent".
Set it too low and unrelated keywords merge.
Set it too high and near-identical keywords end up on separate pages, competing with each other.
I start in the middle, around three or four shared URLs out of ten, then adjust after reading the output. No threshold suits every market. Reading a sample of clusters by eye is your real test.
Semantic clustering ignores search results entirely. It groups keywords by what the words mean.
The modern version uses text embeddings. An embedding model turns each keyword into a list of numbers that represents its meaning.
Keywords with similar meanings sit close together, and a clustering algorithm groups the close ones.
So "boiler repair cost" and "how much to fix a boiler" can land in the same cluster, even though they share almost no words.
Before embeddings were cheap, most tools grouped keywords by shared words or stems. "Roof repair", "roof repairs" and "repair my roof" share a stem, so they cluster.
Lexical grouping still works as a first pass.
But it misses synonyms entirely, and it happily merges keywords that share a word but not a meaning.
Three reasons:
If you are staring at a big raw list, those advantages are real.
Here is the problem with clustering by meaning:
Google does not split pages by meaning. It splits them by intent.
Two keywords can mean nearly the same thing and still return different results. A pattern I see often: one phrasing returns providers, while a close variant returns guides and explainers.
Semantic clustering puts them together. The results say you need two different pages.
The reverse happens too. Two keywords that look unrelated can return the same set of pages, because searchers want the same answer.
SERP overlap catches both cases. It measures the thing that decides rankings, not a proxy for it.
That is why live results are the final check in my own keyword pipeline. A term only gets routed to a page after its results confirm the intent.
If the results are full of guides, the term goes to the blog plan, not a service page. My guide to search intent shows you how to read those signals.
Pro tip: When a semantic cluster mixes "cost" keywords with "near me" keywords, check the results before you merge them. Price research and provider searches often want different pages.
SERP overlap has one big drawback. Every keyword needs a fresh results pull.
That costs you money through a SERP API or a keyword tool's credits. It takes time. And results change, so a clustering run from last year may no longer match what ranks today.
Semantic clustering costs a fraction of that. The keyword text is all it needs.
On a few hundred keywords, the cost difference rarely matters. On tens of thousands, it decides whether you can cluster at all.
There is a middle path that keeps your costs down:
The first two points come straight from how I run automated keyword research: a relevance filter before any paid data, and cached calls throughout.
If your list came from competitor research, expect plenty of off-topic terms. Filter them before you pay to cluster them.
Neither method is safe on autopilot. These are the failure patterns I see most.
For local services, results change by location. A pull for the wrong city describes the wrong market.
Set the location on every request, and expect map results and directories to muddy your overlap count.
Low-volume keywords often have weak results pages: a few directories, a forum thread, a stray PDF.
Overlap scores built on those pages are noisy, so treat the clusters as suggestions.
Keywords containing a brand name return that brand's own pages. They cluster with each other and tell you little about intent. Filter them out first.
Embedding models are good at meaning and weak on commercial nuance.
They will merge "best X" with "X near me" without a second thought. Check the results before those clusters become pages.
Choose by the decision you are making, not by the tool you already pay for.
Choose SERP overlap if:
Choose semantic clustering if:
Use both if:
In that last case, semantic grouping gets your list into topic families. Then SERP overlap splits each family into pages.
That is also how keywords become a working topic cluster structure, with a pillar page and supporting articles around it.
Here is the order I recommend:
Step five is where clustering pays off.
A cluster with no page is a note. A cluster mapped to a page is the start of a content brief.
Want to see how that mapping fits a full research process? Read my walkthrough on how to automate keyword research.
Once clusters turn into topics, check them against what your site already covers. My guide to content gap analysis covers that next step.
The best clustering method is the one that produces pages Google wants to rank.
SERP overlap is the most reliable signal for that, because it measures what Google already shows. Semantic clustering is the quickest way to make a big list manageable.
Together, they give you accurate page decisions without paying for a results pull on every row.
Your first action: take your most important page and check which keywords share its results. You may find two of your pages chasing the same cluster, which is exactly what keyword cannibalisation looks like.
If you would rather have the clustering, routing and checks built and run for you, that is what my keyword pipeline work covers. For the research itself without the automation, see my keyword research service.
For deciding which keywords share a page, SERP overlap is the most reliable keyword clustering method because it uses the URLs Google actually ranks. For sorting a very large raw list into topics, semantic clustering is faster and far cheaper. Most serious keyword plans use both: meaning first to group the list, then live results to decide pages.
There is no universal number. A middling threshold, such as three or four shared URLs in the top ten, is a sensible place to start. Then read a sample of the clusters by eye. If unrelated keywords are merging, raise the threshold. If near-identical keywords are splitting apart, lower it.
You can ask a language model to group keywords, and it will group them by meaning. That makes it a form of semantic clustering, with the same blind spot: it cannot see how Google splits intent in the results. It can also produce confident groupings that do not hold up, so check its clusters against live results before building pages.
No. Keyword clustering groups individual search terms so that each group can be targeted by one page. A topic cluster is a site structure: a pillar page and the supporting articles that link to it. Keyword clusters are the raw material you use to plan the pages inside a topic cluster.
Refresh keyword clusters when you plan a new batch of pages or when rankings shift noticeably. Search results change over time, so overlap-based clusters drift. Rerunning the overlap check before a major content round is usually enough, and caching earlier results keeps the cost of reruns down.
About the author
Kristijan Gjorgjioski
Independent SEO consultant · 6+ years in SEO · 60+ client websites worked on
I'm an independent SEO consultant with 6+ years in SEO. I run audits, migrations and content programmes for businesses in Europe, the UK and the USA, and I build the automation behind my own research.
Book a free 30-minute call. I look at your site first and tell you plainly what I would fix, and in what order.