Why cluster
Google's own guidance has long said that it understands synonyms and variants, and the results page (SERPThe page of results a search engine shows.) confirms it: "how to season a cast iron pan", "seasoning cast iron" and "cast iron pan seasoning" return the same pages. Building three pages for them splits the site's effort and creates the cannibalisation the engine then has to resolve. Clustering is the step that prevents that at planning time, and it also reveals the opposite error: two queries that look alike but whose results differ — "cast iron pan" (products) and "cast iron pan care" (guides) — need different pages.
Clustering is distinct from the topic cluster model in the content section. Keyword clusters decide the pages; topic clusters decide how those pages are grouped and linked around a subject. The keyword cluster for "seasoning cast iron" is one page; the cast-iron topic cluster is that page plus the buying guide, the cleaning guide, the recipes and the hub that links them.
Two ways to cluster
By results-page overlap. For each pair of queries, compare the top results. If a sufficient share of the ranking URLs is shared, the queries cluster. This is the method most clustering tools use, and it is the most faithful to how the engine actually behaves, because it asks the engine. Its costs: it needs a results snapshot per query, which is a tool or a lot of manual checking; the results change over time; and the threshold for "sufficient overlap" is a judgement (tools let it be tuned).
By meaning. Group queries that ask for the same thing in different words, by reading them, by stemming and modifier stripping, or by embeddingA list of numbers that represents a piece of text so that texts with similar meaning sit close together, used for search, retrieval and clustering. similarity. Fast, cheap, and prone to grouping queries the engine treats differently (same words, different intent) or separating ones it treats the same. The mechanism behind embedding similarity, and why it can disagree with the results page, is explained in how embeddings group similar text.
The practical method uses both: a semantic pass to propose groups, a results-page check to confirm or split them, and intent labels from the classification guide as a hard boundary — queries with different intents never share a page, whatever the words say.
The rules
- Different intent, different cluster. Always.
- Same results, same cluster. When the top results are largely shared, one page.
- Different results, different cluster, even when the words are close.
- Modifiers that change the results split; modifiers that do not, merge. "For beginners" may change the results (a different, simpler page ranks); "2026" usually does not.
- Locations split by place where the results page is local, and merge where it is not.
- Singular and plural, word order, stop words, common misspellings merge unless the results say otherwise.
- The head term is the query with the most demand whose results page best represents the cluster; the rest are variants to be covered naturally on the page, not listed.
Clustering at scale
- Start from the labelled list with intent, volume estimates and the site's current ranking URL from Search Console (GSCGoogle's search performance and indexing tool.).
- Partition by intent and by topic area so the comparison space is manageable.
- Run a semantic grouping — a tool, a script that strips modifiers and stems, or an embedding similarity pass — to propose clusters. A language model can also propose groups by meaning, with the limits the AI for keyword research guide describes.
- Confirm with results-page overlap for every proposed cluster above a value threshold and for a sample of the rest, using a SERP tool's snapshots. Split where overlap is low; merge where two proposed clusters share results.
- Choose the head term and list the variants per cluster.
- Record the cluster, its head term, intent, total estimated demand, the site's current best URL and position for it, and the results page's dominant format.
- Hand over to keyword mapping, where each cluster gets one URL. The keyword map template gives the sheet each cluster becomes a row of.
Clustering a payroll software site's queries
A list of informational queries around payroll proposes, by meaning, a cluster of phrasings for "how to run payroll": "how to do payroll", "running payroll for the first time", "payroll process steps", "how to run payroll for a small business", "how to run payroll manually". Results-page overlap confirms the first four share most ranking URLs; the fifth returns pages specifically about doing it without software, with little overlap, and is split into its own cluster. A separate proposed cluster, "payroll compliance checklist", turns out to share most of its results with "payroll process steps" and is merged in. The head term becomes "how to run payroll", with the small-business phrasing covered as a section; "how to run payroll manually" becomes its own page. The site's current ranking for the head cluster is a help-centre article that ranks weakly; the map will decide whether it is rebuilt or replaced.
Reading the clusters
The clustered list answers questions the raw list could not:
- How many pages the site needs for a topic area, and how many it already has.
- Which existing pages cover several clusters and should be split, and which clusters are covered by several pages and should be consolidated — the cannibalisationTwo or more pages on one site competing for the same query and intent, splitting signals. findings in advance.
- Which clusters are large and winnable, from demand against the strength of what ranks.
- Which format each cluster needs, from the results page's dominant type, which feeds the search intent reading for each page.
- Where click-through (CTRShare of impressions that result in a click.) will be constrained by result features such as snippets and product grids, which affects what demand actually reaches a page.
Tools and methods
Dedicated clustering tools take a list and a results snapshot and return groups at a chosen overlap threshold; rank-tracking tools and keyword tools often include a clustering feature; a spreadsheet with stemmed, modifier-stripped queries sorted alphabetically does a workable semantic pass for small lists; and the keyword research expertise page includes an interactive demonstration of grouping by shared terms. Whatever the tool, the results-page confirmation on the valuable clusters is done by a person.
Setting the overlap threshold
Results-page clustering needs a rule for how much overlap makes two queries one cluster, and the choice shapes the whole plan. A high threshold (most of the top results shared) produces many small clusters and therefore many pages, some of which will compete with each other because the engine treats their queries as near-equivalent on other days. A low threshold (a few shared results) produces few large clusters and pages that try to serve queries the engine actually answers differently. Neither is right in the abstract. The practical approach is to run the clustering at two or three thresholds, look at where the valuable clusters split or merge between them, and decide those cases by reading the results pages directly. The threshold is then recorded with the research so that the next cycle uses the same one, and the exceptions are recorded with their reasons.
Position depth matters too: comparing only the first few results is strict, comparing the first ten is looser, and the engine's own behaviour of showing one page per site means that overlap among the first few positions is the strongest evidence that one page can serve both queries.
Clustering across markets and languages
Clusters do not translate. A cluster in one language or country is rebuilt from that market's queries and that market's results pages, because the phrasings, the intents and the engine's treatment of them all differ. The international section explains why the resulting page sets differ by edition; a site that clones its primary edition's clusters into other markets builds pages for queries nobody there types and misses the ones they do. The tutorial on keyword research for a new market rebuilds the clusters from that market's own queries and results pages.
Question clusters
Question phrasings ("how long does", "can I", "what happens if") often cluster not into their own pages but into sections of a broader page: the results page for the question shows the same guides that rank for the subject, with the answer in a section or a snippet. The rule holds — same results, same cluster — and the questions become the headings and the FAQ entries of the owning page, which the content briefs record as coverage. A question whose results page differs from the subject's, with its own specialised answers, is its own cluster and its own page.
Maintaining clusters
Results pages change, categories mature, and new phrasings appear. The clusters are reviewed quarterly with the rest of the research: new queries from Search Console are assigned to existing clusters or create new ones; head-cluster results are re-checked for merges and splits; and any page whose ranking queries now straddle two clusters is flagged for the map. The same clusters make a good rank tracking set, a head term and a few variants each, and the rank tracking guide shows how to tag tracked queries by cluster so positions are read for the group rather than query by query.
Once clusters are built, the SEO traffic estimator for keyword prioritisation ranks them by estimated conversions at a position the site can realistically reach.
Common mistakes
- Clustering by words alone. The engine disagrees, and the pages compete or miss.
- One page per variant. Cannibalisation built into the plan.
- Ignoring intent as a boundary. A guide and a product page in one cluster.
- Head term by volume only. The variant with the biggest number may have the least representative results.
- Treating clusters as keyword lists for the page. The page covers the subject; the variants follow.
- Never re-checking. Clusters drift with the results page.
The keyword mapping tutorial takes the clusters and assigns URLs; the topic clusters guide organises the resulting pages into a linked structure.
