Topic clusters that actually work (and the mistake almost everyone makes)
If your clustering tool groups keywords by how similar the words look, it is telling you about English, not about Google.
Most keyword clustering tools group terms by string or embedding similarity. That reliably puts "best running shoes" next to "best running shorts" — two words apart, two completely different SERPs, two completely different pages.
SERP-overlap clustering compares the actual ranking URLs for each query and groups terms only where Google demonstrably treats them as the same topic. It is more expensive to compute and it is the only method that answers the question you actually have: do I need one page here, or five?
How SERP overlap works
For each keyword pair, fetch the top ten results and count shared URLs. Three or more shared URLs is a reasonable threshold for "same page can serve both". Below that, they need separate pages. Run this pairwise across your keyword set and the clusters emerge from the data rather than from a similarity score.
The results are frequently counter-intuitive. Terms that look identical often have zero overlap because one is informational and the other transactional. Terms that share no words often overlap almost completely because Google has decided they are the same question.
Structuring the cluster
Once clusters are real, the architecture follows:
- Pillar page targeting the highest-volume, broadest term in the cluster. Comprehensive, but not a dumping ground.
- Spoke pages for each sub-topic with a distinct SERP, each going deeper than the pillar does.
- Bidirectional links. Pillar links to every spoke; every spoke links back to the pillar. This is the part teams skip and it is most of the value.
- Lateral links between spokes only where genuinely relevant. Linking everything to everything dilutes the signal.
Label intent before briefing anyone
Each cluster gets one label, and the label decides the page type:
| Intent | Page type | Primary goal |
|---|---|---|
| Informational | Guide or explainer | Coverage and citations |
| Commercial | Comparison or listicle | Shortlist inclusion |
| Transactional | Product or category | Conversion |
| Navigational | Brand or docs | Retention |
Writing a guide for a transactional cluster is the most common and most expensive briefing error in content marketing.
Three failure modes
The pillar that is a table of contents. A pillar page needs to be genuinely useful on its own. If it is a list of links to spokes, it will not rank and it will not be cited.
Spokes that cannibalise. Two spokes competing on the same query split the signal and neither wins. If SERP overlap says they are the same topic, they are one page.
Clusters with no commercial edge. Beautiful informational clusters that never touch a buying decision generate traffic and no pipeline. Every cluster should have a path to something you sell, even if it is three clicks away.
Sequencing the build
Publish the two or three highest-intent spokes first, then the pillar, then the remaining spokes. Spokes ranking before the pillar gives the pillar internal support from day one — the reverse order leaves the pillar sitting alone for months.