Topic-cluster keyword research is the workflow most SEO articles reference and few actually walk through. The reason is that the workflow requires holding the topic graph in mind, and most teams default back to a keyword spreadsheet because the spreadsheet is the shape their tools produce. The steps below are the workflow as it actually runs, with the templates that make each step transferable across topics.
The workflow
Pick a seed topic, not a seed keyword
Step 1 · 15 minutes
Topics are clusters; keywords are nodes. A seed topic is one or two phrases that characterize the cluster you intend to build authority on. Examples: keyword research, topical authority, semantic SEO, knowledge graphs for SEO.
Avoid head terms that are too generic (“SEO,” “marketing”) or too specific (“best keyword research tool for SaaS startups in 2026”). The right seed sits at the centroid of the cluster you want to own.
Build the demand graph
Step 2 · 20 minutes with a graph tool, two hours by hand
Pull the related queries Google surfaces for the seed (autocomplete + People Also Ask + the related-searches block). Pull the question variants from AnswerThePublic or AlsoAsked. Add the entity neighborhood from Google’s Knowledge Graph for the seed term if relevant.
Represent the resulting query list as a co-occurrence network: nodes are entities, edges are co-occurrences of those entities across the query list. The clusters that surface are the actual shape of demand.
Build the supply graph
Step 3 · 20 minutes with a graph tool
Pull the top 30–50 SERP results for the seed query. Extract the entities and topics each page covers (titles, H2s, definitions, first 200 words). Build the parallel co-occurrence graph.
The clusters in the supply graph are what the current SERP rewards. Compare them with the clusters from step 2.
Identify clusters and gaps
Step 4 · 30 minutes
For each cluster in the demand graph, check whether the supply graph has a corresponding cluster of similar size and density. Three outcomes:
- Matched. Demand cluster present in supply. SERP is saturated; publishing here is incremental work.
- Underweight. Demand cluster present in supply but smaller or less coherent. Real opportunity; new content can become the canonical reference.
- Missing. Demand cluster absent from supply. Highest opportunity; first coherent answer becomes the default cited source.
Map the publishing plan
Step 5 · 45 minutes
For each cluster, define one canonical page (the central article on the cluster’s central entity) and three to five perimeter pages (canonical pages on constituent entities). Add bridges to adjacent clusters identified in the demand graph.
Order the publishing plan by gap size, not by volume. Missing clusters first; underweight next; matched clusters last (and possibly not at all if the cluster is fully saturated).
Re-run quarterly
Step 6 · 1–2 hours per topic per quarter
Demand shifts as new platforms and formats emerge. Supply shifts as competitors publish. The gaps move with them. Quarterly refresh keeps the publishing plan aligned with the demand that actually exists rather than the demand at the time of the initial analysis.
Most teams skip this step and discover six months later that two of their target gaps closed and a new one opened. The refresh is much cheaper than the rework.
Templates
Three artifacts that travel across topics. Build them once; reuse across clusters.
- Demand-graph capture sheet. Columns for entity, source (autocomplete / PAA / related searches / external tool), approximate volume, observed co-occurrences. One row per entity. Used in step 2.
- SERP-coverage matrix. Rows are clusters; columns are the top 20 SERP results. Each cell marks whether the page covers that cluster (full / partial / none). Surfaces gaps visually. Used in steps 3 and 4.
- Publishing-plan grid. Rows are planned pages; columns are cluster, page type (central / perimeter / bridge / commercial), priority, target queries, sibling links. Used in step 5 and updated in step 6.
Where this workflow stops being feasible by hand
Two thresholds. Past them, the workflow needs tooling.
Past three clusters in active build. The mental overhead of holding multiple cluster states becomes the bottleneck. A graph tool that persists the demand and supply graphs across clusters lifts the bottleneck.
Past 500 queries per cluster. Manual clustering past this volume produces enough errors that the gap analysis becomes unreliable. AI-assisted clustering (KeywordGraph or similar graph-native tools) handles the volume cleanly.