Semantic vector clustering reveals the architecture algorithms see
Use embedding clusters to map a site’s real semantic architecture, find editorial gaps and improve content decisions.
highlights
- • Editorial categories and algorithmic clusters do not always match.
- • Vector clustering groups pages by meaning rather than folders.
- • The map exposes misplaced content, weak hubs and fragmented topics.
- • A Site Focus Score measures distance from the domain’s thematic center.
- • The semantic map should be recalculated as the site grows.
what vector clustering actually does
Embeddings place pages in a shared semantic space. Clustering algorithms then group nearby pages without using the site’s folders or editorial labels, producing an independent view of the information architecture.
K-Means or HDBSCAN can allocate each page to the nearest semantic group. The resulting clusters represent the territories the content actually covers, while the page closest to each cluster centroid acts as its most representative document.
what the map reveals
A cluster may contain pages split across several categories, or one category may contain unrelated groups. Outliers can signal misplaced content, while dense groups without a clear hub can reveal a missing pillar page.
Very similar pages inside one group can also expose fragmentation that should be consolidated. Small isolated clusters deserve review because they may represent a deliberate new subject—or content that falls outside the domain’s focus.
site focus score
Distance from the site centroid provides a useful directional metric for thematic coherence. It is not a ranking factor; it helps identify pages whose subject matter sits far from the rest of the domain.
A specialized publication will normally have lower average drift than a broad portal. High-drift pages are not automatically wrong, but they prompt a useful editorial question: do they strengthen the domain’s expertise or dilute it?
editorial decisions from clusters
Use clusters to review navigation, hub pages, internal links and the content roadmap. A visualization is only useful when every pattern can lead to a testable editorial or architectural decision.
Large clusters with no authoritative center may need a pillar page. Closely overlapping low-traffic pages may need consolidation, while valuable small clusters can justify deeper coverage and stronger internal links.
how often to rerun the analysis
Recalculate after meaningful publishing or migration cycles. Comparing snapshots reveals whether the site is becoming more coherent, expanding into a new territory or creating isolated pockets of content.
Quarterly analysis is useful for active publishers; a semiannual cycle may be enough for stable sites. The important point is to compare snapshots, because every new group of pages changes both local clusters and the overall centroid.
final thoughts
The goal is not to force the site to match an algorithmic map. The value lies in comparing the intended architecture with the structure that emerges from meaning.
Clustering makes that gap visible, but editorial judgment still decides what to consolidate, strengthen, reposition or leave untouched. The map is a diagnostic instrument, not the final architecture.
sources and further reading
Lucas Cassapula
Partner & Head of SEO at Wesearch and Co-founder of Mentionflow
I am a partner at Wesearch and co-founder of Mentionflow. I have worked with SEO for almost 10 years. I am a data-driven geek who is always testing hypotheses, looking for patterns and turning ideas into products. I share studies, experiments and automations focused on SEO and GEO.