You exported 3,000 keywords. Now you need to know which ones belong on the same page, which ones need their own page, and which ones will make two of your posts fight each other in Google.
Quick answer: install AgriciDaniel/claude-seo and use its free seo-cluster skill, which groups keywords by SERP overlap and can pull SERPs from DataForSEO. Use ClawFu’s keyword-clusterer (archived, so no updates) when you only need a fast semantic pass. Either way, cluster on real SERP data, not keyword text alone. Still building the keyword list? Start with skills for keyword research.
That is the clustering problem. Claude Code can do most of the grouping work, but only if you give it the right skill and the right data. A language model guessing from keyword text alone will happily put “best crm” and “what is a crm” in one cluster, and Google disagrees.
This guide covers the skills worth installing (free and paid, with verified links, star counts and prices as of 1 November 2026), the two clustering methods and when each one is right, the data you need, and the workflow from cluster to published page.
What a Claude Code keyword clustering skill does
Short answer: a keyword clustering skill is a SKILL.md file (sometimes with scripts) that teaches Claude Code a repeatable method for grouping a keyword list into page-sized clusters. The good ones also label search intent, pick a primary keyword per cluster, and flag cannibalization with pages you already have.
If skills are new to you, read what Claude Code skills are first. The short version: a skill is a folder with a SKILL.md file that lives in ~/.claude/skills/ (available everywhere) or .claude/skills/ (one project). Claude loads it when your request matches its description. Skills also ship inside plugins, which you add with /plugin marketplace add owner/repo and /plugin install. Many repos also support npx skills add owner/repo. The full walkthrough is in how to install Claude Code skills.
A clustering skill usually does five jobs:
- Ingest a CSV export or a seed keyword and expand it.
- Group keywords, either by SERP overlap or by semantic similarity.
- Label each cluster with intent (informational, commercial, transactional, navigational) and funnel stage.
- Choose a primary keyword and a page type per cluster.
- Map clusters to existing URLs or to new pages, with internal link suggestions.
What a skill cannot do on its own is know search volume, difficulty or what currently ranks. Claude has no live keyword database. That data comes from your exports or from an MCP server such as DataForSEO, which we cover below.
Takeaway: the skill supplies the method; your data supplies the truth. You need both.
Semantic clustering versus SERP-overlap clustering
Short answer: semantic clustering groups keywords whose meaning is similar (using embeddings or the model’s judgment). SERP-overlap clustering groups keywords that Google ranks with the same URLs. SERP overlap is slower and costs money per query, but it reflects how Google actually treats the terms, so it is the safer basis for deciding what goes on one page.
How SERP-overlap clustering works
You pull the top 10 organic results for each keyword, then compare pairs. If two keywords share enough ranking URLs, Google sees them as one topic, and one page can rank for both.
The thresholds differ by skill. AgriciDaniel’s seo-cluster skill (inside the claude-seo plugin) uses this table in its SKILL.md:
| Shared top-10 URLs | Relationship | Action |
|---|---|---|
| 7 to 10 | Same post | Merge into one target page |
| 4 to 6 | Same cluster | Group under the same spoke cluster |
| 2 to 3 | Interlink | Adjacent clusters, add cross-links |
| 0 to 1 | Separate | Different clusters, or exclude |
Rampstack’s seo-keyword skill uses a looser rule: if two keywords share at least 3 of the top 10 results, they target the same page. Neither threshold is “correct”. A stricter threshold gives you more, narrower pages. A looser one gives you fewer, broader pages. Pick one and apply it consistently.
How semantic clustering works
Semantic clustering turns each keyword into a vector (an embedding) and groups vectors that sit close together. The ClawFu keyword-clusterer skill does this with sentence-transformers and scikit-learn, and falls back to a simpler mode if you only install pandas and click. You can also ask Claude to cluster by judgment alone, which is the cheapest and least reliable version.
When to use which
| Situation | Better method | Why |
|---|---|---|
| Deciding what goes on one URL | SERP overlap | Google’s results are the ground truth for “same intent” |
| First pass on 10,000+ raw keywords | Semantic | Free and fast; cuts the list before you pay for SERP pulls |
| Building a topical map or hub structure | Semantic, then SERP check on borderline pairs | Hubs are about topic; pages are about ranking |
| PPC ad groups | Semantic | Ad relevance follows wording more than organic overlap |
| Two existing pages may be competing | SERP overlap | You need to see whether Google treats them as one query |
The practical answer for most sites is a hybrid. Run semantic clustering to shrink the list into candidate groups, then spend SERP queries only on keywords where the grouping is unclear. The claude-seo skill takes a similar line: it groups by an initial intent guess first to cut down the number of pairwise comparisons.
Here’s the catch with semantic clustering: words that look alike can carry different intent. “Claude code skills” and “claude code skills tutorial” sound like one topic, but one SERP may be full of directories and the other full of how-to posts. Only the SERP tells you that.
Takeaway: use meaning to sort the pile, and use SERPs to decide the page.
The best free Claude Code skills for keyword clustering
Short answer: for most people the best free starting point is the seo-cluster skill in AgriciDaniel/claude-seo, because it does SERP-overlap clustering and can pull SERPs from DataForSEO. Rampstack’s seo-keyword is a good lighter option, and ClawFu’s keyword-clusterer covers embedding-based clustering.
Star counts and last-push dates below come from the GitHub API on 1 November 2026. Stars measure attention, not quality, so read each SKILL.md before you install.
1. seo-cluster (AgriciDaniel/claude-seo)
- Source: github.com/AgriciDaniel/claude-seo (about 25,400 stars, pushed 1 Nov 2026)
- Price: free, open source
- What it does: SERP-overlap clustering, hub-and-spoke architecture, internal link matrices and an interactive cluster map.
/seo cluster plan <seed>runs the planning workflow;/seo cluster executewrites posts with the claude-blog skill if installed, or outputs briefs if not. If the DataForSEO MCP is connected, it uses live SERP data and runs a cost check first; otherwise it falls back to web search. - Install:
/plugin marketplace add AgriciDaniel/claude-seo
/plugin install claude-seo@agricidaniel-claude-seoThe plugin brings a full SEO suite, so it is overkill if you only want clustering. It pairs well with the audit tools in Claude Code skills for SEO audits.
2. seo-keyword (rampstackco/claude-skills)
- Source: github.com/rampstackco/claude-skills (about 1,000 stars, pushed 27 Oct 2026)
- Price: free
- What it does: keyword discovery, intent classification, clustering and prioritization in one skill. It describes two approaches, with SERP overlap (3 of top 10 shared) as the mechanical option. It is tool-agnostic, so it works with any CSV.
- Install:
/plugin marketplace add rampstackco/claude-skills, then/plugin install rampstack-seo@rampstack.
3. keyword-clusterer (guia-matthieu/clawfu-skills)
- Source: github.com/guia-matthieu/clawfu-skills, path
skills/seo-tools/keyword-clusterer(about 150 stars, pushed 1 Oct 2026; archived and read-only as of 2 November 2026, so expect no updates) - Price: free, MIT
- What it does: semantic clustering with embeddings via a Python script. Good for the first pass on a large list and for PPC ad groups.
- Install: copy the folder into
~/.claude/skills/keyword-clusterer, then runpip install scikit-learn sentence-transformers pandas click.
4. seo-keyword-cluster-builder (secondsky/claude-skills)
- Source: github.com/secondsky/claude-skills, under
plugins/seo-keyword-cluster-builder(about 240 stars, pushed 31 Oct 2026). Also listed on skills.sh. - Price: free
- What it does: expands seeds with modifiers and groups them into topic and intent clusters. Judgment-based, so check borderline groups against real SERPs.
- Install: clone the repo and copy the skill folder into
~/.claude/skills/, or use the install command shown on its skills.sh listing.
5. keyword-cluster-builder (MadAppGang/claude-code)
- Source: github.com/MadAppGang/claude-code, under
plugins/seo/skills/keyword-cluster-builder(about 290 stars, last pushed 15 Mar 2026, so slower moving) - Price: free
- What it does: seed expansion with question, comparison, intent and audience modifiers, then clusters labelled with intent and funnel stage.
- Install: clone the repo and copy the skill folder into
~/.claude/skills/.
Where else to find clustering skills
Directories help you discover skills but they are catalogs, not reviewers. claudskills.com lists seo-cluster and a “keyword clustering and mapping” skill that maps clusters to existing pages. skills.sh, skillsmp, claudemarketplaces.com and cultofclaude also index SEO skills. Always click through to the source repo and read the SKILL.md. The broader roundups in Claude Code skills for keyword research and Claude Code skills for SEO cover adjacent tools.
coreyhaines31/marketingskills (about 57,800 stars) is the most-starred marketing skills repo, but it focuses on strategy, copy and programmatic SEO rather than a dedicated clustering method. Use it alongside a clustering skill, not instead of one.
Paid options: clustering APIs, MCP servers and skill packs
Short answer: paid options either sell you clustering as a service (Keyword Insights), sell you the SERP and volume data the free skills need (DataForSEO, Ahrefs, Semrush), or sell bundles of skill files (Gumroad packs). Pay for data before you pay for prompts.
Keyword Insights Claude skill
- Source: docs.keywordinsights.ai/integrations/claude-skill
- Price: the docs say API access needs the Premium plan. The pricing page’s structured data lists Premium at $299 per month; the visible plan cards show Basic at $58 and Professional at $99 per month. Confirm on keywordinsights.ai/pricing before you buy. Clustering costs about 1 credit per keyword; a content brief is about 100 credits.
- What it does: sends your keyword list to Keyword Insights’ clustering and intent API, parses CSV exports, summarises hub structure and can generate briefs. Claude checks your credit balance and warns before expensive orders. Minimum 5 keywords per order. The docs themselves recommend the API or web UI, not Claude, for orders over 500 keywords.
- Install: download
keyword-insights-SKILL.mdfrom their GitHub, upload it under Settings, Capabilities, Skills in Claude, addapi.keywordinsights.aito your domain allowlist, and setKWI_API_KEY. In Claude Code you can drop the file into a skills folder instead.
DataForSEO MCP server
- Source: github.com/dataforseo/mcp-server-typescript (about 260 stars, pushed 1 Nov 2026)
- Price: pay as you go. DataForSEO pricing shows a $50 minimum top-up, no subscription, a $1 trial credit, SERP data from $0.0006 per SERP and keyword data from $0.0012 per task. Live mode costs more than standard mode.
- What it does: gives Claude live SERPs, volumes and keyword data. This is what turns the free SERP-overlap skills from “web search approximation” into real data.
- Install:
npx dataforseo-mcp-server@latestwith your API login and password as environment variables, then register it withclaude mcp add.
Ahrefs, Semrush and Google Search Console MCP servers
Ahrefs and Semrush data can also reach Claude through MCP servers (check each vendor for an official one), and there are community GSC MCP servers. They feed clustering the same way DataForSEO does. You need a paid Ahrefs or Semrush plan with API access; check each vendor’s current pricing page, because API tiers change often. GSC is free and is the best source for cannibalization checks on your own site.
Gumroad skill bundles
The Master Marketing Skills Bundle on Gumroad is listed at $167 (the product title also says “167 skills”; the price in the page data is separately $167). It is a general marketing pack and the listing does not name a dedicated clustering skill, so it is not a reason to buy if clustering is your only goal. We found no clustering-specific paid skill on Contra with a public price.
Free vs paid comparison
| Option | Free or paid | Price (verified 1 Nov 2026) | Method | Needs external data? |
|---|---|---|---|---|
| seo-cluster (claude-seo) | Free | $0 | SERP overlap | Optional DataForSEO; falls back to web search |
| seo-keyword (Rampstack) | Free | $0 | SERP overlap or judgment | Your CSV |
| keyword-clusterer (ClawFu) | Free | $0 | Embeddings | Your CSV |
| seo-keyword-cluster-builder (secondsky) | Free | $0 | Judgment | Your CSV |
| keyword-cluster-builder (MadAppGang) | Free | $0 | Judgment | Your CSV |
| Keyword Insights skill | Paid | Premium plan, listed at $299/mo; plus credits | Their SERP-based API | No, included |
| DataForSEO MCP | Paid (usage) | $50 minimum top-up; from $0.0006 per SERP | Data source | It is the data |
| Ahrefs / Semrush MCP | Paid | Pricing varies by plan; check vendor | Data source | It is the data |
| Gumroad marketing bundle | Paid | $167 one-time | General marketing skills | Not clustering-specific |
The data you need before you cluster
Short answer: you need a keyword list with volume, a SERP snapshot for the keywords you will compare, and a list of your existing URLs with the queries they already rank for. A CSV export covers the first; DataForSEO or another SERP source covers the second; Google Search Console covers the third.
Minimum viable dataset
| Data | Where it comes from | Why clustering needs it |
|---|---|---|
| Keyword, volume, difficulty | Ahrefs, Semrush, Keyword Planner CSV, or DataForSEO | Picks the primary keyword per cluster |
| Top-10 URLs per keyword | DataForSEO SERP API, or web search fallback | Required for SERP-overlap grouping |
| Your URLs and their queries | GSC performance export (query + page) | Maps clusters to pages, finds cannibalization |
| Current page list | Sitemap or crawl | Stops you planning a page you already have |
Preparing a CSV export
Clean before you cluster. Give Claude one file with these columns: keyword, volume, kd, intent (optional), current_url (optional). Then ask the skill to:
- Lowercase and trim whitespace.
- Drop exact duplicates and near-duplicates (plural, word order).
- Remove branded competitor terms unless you want comparison pages.
- Drop zero-volume rows unless they come from GSC with real impressions.
Controlling DataForSEO cost
SERP overlap needs a SERP per keyword. The comparison is pairwise, but the pulls are per keyword, so cost scales with list size. Use the per-SERP price on DataForSEO’s pricing page to estimate before you run, and pre-filter with semantic clustering so you only pull SERPs for keywords that matter. The claude-seo skill runs a cost check script before live SERP calls, which is a good habit to copy in your own skills.
Takeaway: a cheap, clean list beats an expensive, messy one.
How to turn keyword clusters into pages
Short answer: each cluster becomes one URL with one primary keyword. Pick the cluster’s page type from what ranks in its SERP, decide whether it is a hub or a spoke, and write a brief that names every secondary keyword the page should cover.
Step-by-step workflow in Claude Code
- Cluster. Run
/seo cluster plan "<seed>"or your chosen skill on the cleaned CSV. - Assign one primary keyword per cluster. Usually the highest-volume term whose SERP best represents the group.
- Match the page type to the SERP. If the top results are lists, write a list. If they are tools, build a tool. Fighting the format rarely works.
- Decide hub or spoke. Broad clusters with overlap to many others become hubs; narrow ones become spokes. This is the step that becomes your topical map, covered in Claude Code skills for topical maps.
- Write the brief. Primary keyword, secondary keywords, questions to answer, competing URLs. See Claude Code skills for content briefs.
- Plan links. Spokes link up to the hub; sibling spokes with 2 to 3 shared URLs cross-link. Claude Code skills for internal linking covers automating this.
- Write and publish. Use a writing skill from Claude Code skills for content writers.
Cluster to page decision table
| Cluster signal | Page decision |
|---|---|
| High SERP overlap, one intent | One page, all keywords as sections |
| Same topic, mixed intent (guide vs product) | Separate pages, cross-link |
| Template-like pattern (“X for Y” across 50 Ys) | Consider programmatic pages: programmatic SEO skills |
| Location modifiers | Local pages: local SEO skills |
| Cluster matches an existing URL | Refresh, do not create: content refresh skills |
A sample prompt that works with any skill
Use the clustering skill on keywords.csv.
Method: semantic first pass, then SERP overlap for any pair in the same
group with ambiguous intent. Merge at 7+ shared top-10 URLs, cluster at 4-6.
Output a table: cluster, primary keyword, secondary keywords, intent,
page type, hub or spoke, existing URL match from gsc.csv.
Flag any cluster where two of my URLs both get impressions.How to avoid keyword cannibalization
Short answer: cannibalization happens when two of your pages target the same cluster, so Google alternates between them or ranks neither well. Prevent it by mapping every cluster to exactly one URL before writing, and detect it by checking GSC for queries where two of your pages both get impressions.
Prevention: one cluster, one URL
Keep a single keyword map (a CSV or sheet) with one row per cluster and one owning URL. Before any new brief, have Claude check the new primary keyword against that map using SERP overlap, not just string matching. If the new keyword shares 7 or more top-10 URLs with an existing cluster, it belongs on the existing page.
Detection: read your GSC export
Export GSC performance data with both query and page dimensions. Ask Claude to list every query where two or more of your URLs get impressions, sorted by impressions. Not every overlap is a problem. A hub and a spoke can both appear for a broad query. It becomes a problem when the ranking URL flips between pages or neither holds a stable position.
Fixes, in order of preference
| Situation | Fix |
|---|---|
| Two thin pages on one cluster | Merge into the stronger URL, 301 the other |
| Two pages, different intent, same keyword in titles | Retitle one to its true intent, re-anchor internal links |
| Hub and spoke both target the spoke term | Remove the spoke term from the hub’s title and H1, link down |
| Old post outranking the new one you prefer | Merge content into the old URL instead |
Changes in the technical layer (redirects, canonicals) are covered in Claude Code skills for technical SEO.
Takeaway: cannibalization is a mapping failure. Fix the map, then fix the pages.
Common mistakes with AI keyword clustering
Short answer: the biggest mistakes are trusting judgment-only clusters for page decisions, clustering without your existing URLs, and skipping a human review of borderline groups.
- Clustering on text alone. Judgment-based clusters look tidy and often hide mixed intent. Check borderline groups against SERPs.
- No existing-URL map. You end up planning pages you already have, which is cannibalization by design.
- Feeding unclean exports. Duplicates and plurals inflate cluster counts and waste SERP credits.
- Changing thresholds mid-project. Pick 3-of-10 or 7-of-10 and stick with it for one site.
- Asking Claude for volumes. It does not have them. Volumes must come from a tool or MCP server.
- Running huge lists through chat. Keyword Insights’ own docs suggest their API or UI past 500 keywords. A script-based skill handles big lists better than a conversation.
- Ignoring AI search. Clusters built for Google still need answer-first sections to be cited by AI engines. See Claude Code skills for GEO.
If no skill fits your method, write your own. How to build a Claude Code skill shows the structure, and clustering is a good first skill because the method is easy to write down. Marketers who want a wider kit can start at Claude Code skills for marketers.
Frequently asked questions
Can Claude do keyword clustering?
Yes. Claude can group keywords by meaning and intent from a CSV. For page-level decisions it needs SERP data, either from web search or an MCP server such as DataForSEO.
What is the best Claude Code skill for keyword clustering?
For SERP-based clustering, seo-cluster in AgriciDaniel/claude-seo is the most complete free option. For embeddings, ClawFu’s keyword-clusterer. For a paid managed service, the Keyword Insights skill.
What is the difference between semantic and SERP clustering?
Semantic clustering groups keywords with similar meaning. SERP clustering groups keywords that share ranking URLs on Google. SERP clustering is more reliable for deciding what goes on one page.
How many shared URLs mean two keywords belong on one page?
It depends on the method. Rampstack’s skill uses 3 of the top 10; claude-seo merges at 7 to 10 and clusters at 4 to 6. Choose one threshold per project.
Does Claude have search volume data?
No. Claude has no live keyword database. Use CSV exports from Ahrefs, Semrush or Keyword Planner, or connect a data MCP server.
Do I need DataForSEO to cluster keywords with Claude Code?
No, but it helps. Without it, SERP-overlap skills fall back to web search, which is slower and less complete. DataForSEO has a $50 minimum top-up and charges per request.
How do I install a clustering skill?
Add the plugin marketplace with /plugin marketplace add owner/repo and install the plugin, run npx skills add owner/repo where supported, or copy the skill folder into ~/.claude/skills/.
Can Claude Code cluster 10,000 keywords?
It can process large files with a script-based skill, but SERP pulls for 10,000 keywords cost money and time. Pre-filter semantically, then pull SERPs only for ambiguous groups.
Is keyword clustering still worth it in 2026?
Yes. Clustering decides how many pages you need and which keywords each page owns, which matters for both Google rankings and AI citations.
How do I find keyword cannibalization with Claude?
Export GSC data with query and page dimensions, then ask Claude to list queries where two or more of your URLs get impressions.
Should I merge cannibalizing pages?
Merge when both pages serve the same intent and one is clearly weaker. Keep them separate and retitle when the intents differ.
What is a keyword cluster versus a topic cluster?
A keyword cluster is a set of keywords one page targets. A topic cluster is a group of pages (a hub and spokes) covering a broader subject.
Are free clustering skills good enough?
For most sites, yes, when paired with real data. Paid tools mainly save time on large lists and provide the SERP data for you.
Can I use ChatGPT instead of Claude Code for clustering?
You can paste lists into any chat model, but Claude Code skills give you a repeatable method, file access to your CSVs and MCP data connections in one place.
Are paid Claude skill packs worth buying for clustering?
Only if the pack names a clustering method you can read before buying. General marketing bundles rarely add much over the free repos for this task.
Sources and methodology
Star counts and last-push dates came from the GitHub API on 1 November 2026. Thresholds were read from each skill’s SKILL.md on GitHub. Prices were read from live vendor pages on the same date and can change.
- AgriciDaniel/claude-seo,
skills/seo-cluster/SKILL.md: github.com/AgriciDaniel/claude-seo - Rampstack
seo-keywordSKILL.md: github.com/rampstackco/claude-skills - ClawFu
keyword-clusterer: github.com/guia-matthieu/clawfu-skills - Keyword Insights Claude skill docs: docs.keywordinsights.ai
- DataForSEO pricing: dataforseo.com/pricing
- DataForSEO MCP server: github.com/dataforseo/mcp-server-typescript
Your next step
Pick one method and one data source today. If you want free and SERP-based, install claude-seo, connect DataForSEO with a small top-up, and cluster one seed topic you already write about. Map the clusters against your GSC export before you write a single new brief. That one map will prevent most of the cannibalization problems you would otherwise fix later.
