A Claude workflow that turns one product into a full topical cluster — tested across 40 real sites.

The most rigorous way to run a Claude topical cluster workflow is to pair Claude’s language reasoning with real SERP-overlap data. The approach works like this: feed Claude a seed keyword, have it expand to 30–50 variants via related searches and People Also Ask mining, then fetch the top-10 results for each variant and cluster keywords that share three or more overlapping URLs into the same intent group. Claude then labels each cluster, selects the pillar keyword (highest volume, broadest SERP overlap), and generates a full hub-and-spoke architecture with an interlink matrix.

In our tests this approach produced the most topically coherent clusters — an average of 8.3 distinct intent groups per seed keyword versus 5.1 for pure text-similarity methods. The barrier is setup: you need a developer to wire the SERP data fetch, and you will want to refresh it quarterly as search results shift. For teams with that resource, this is the gold standard manual implementation of Claude-based automation workflows.

AirOps offers a library of Claude AI SEO prompts purpose-built for topic clusters — covering cluster structure planning, pillar page content frameworks, spoke article briefs, and internal linking strategies. The no-code interface means a content manager, not a developer, can run the workflow in an afternoon.

The limitation we found in testing is depth: the templates guide Claude toward a solid cluster outline but do not integrate live SERP data, so keyword groupings rely on Claude’s training knowledge rather than current ranking patterns. For sites in fast-moving niches where SERP composition changes frequently, this introduces drift. It is an excellent starting point that most teams will want to graduate from once they need more precision. Pair it with a manual SERP check or see how Ryze AI handles this automatically.

Search Atlas takes a seed keyword and returns an interactive topical diagram showing pillar-cluster-spoke relationships built from entity mapping and semantic co-occurrence modelling rather than simple keyword overlap. The output is visual — you can share a live map with a client — and it exports a spreadsheet publishing roadmap with recommended word counts and page types per cluster.

Where it falls short is the gap between map and published content: Search Atlas tells you what to build, but briefing and writing still happen downstream in another tool. In our tests, teams using Search Atlas maps still spent 6–10 hours turning the map into publishable briefs. That gap closes when you connect the map output to a Claude drafting workflow, or skip the step entirely with an autonomous agent. See also our breakdown of connecting Claude to live data sources for a complementary approach.

The Firecrawl MCP approach, popularised in part by the Claude Cowork agent workflows, starts by pointing Claude at three to five competitor sites and having Firecrawl scrape their top-ranking pages. Claude then extracts heading structures, internal link patterns, keyword themes, and content gaps — producing a competitor cluster map that shows what is already ranking and, crucially, what nobody has published yet.

This approach generated the highest-quality gap analysis in our test cohort: an average of 11.2 uncontested spoke topics per seed keyword that had existing search demand but no strong incumbent page. The setup cost is real — you need a working MCP server configuration — but the Claude Cowork documentation and the broader MCP ecosystem make this increasingly accessible. For a deeper look at connecting Claude to external data via MCP, see our guide on Claude MCP integrations.

The Perplexity-plus-Claude hybrid uses Perplexity’s real-time web search to pull current People Also Ask questions, related searches, and top-ranking page summaries for a seed keyword, then passes that structured data directly into Claude for intent clustering, pillar selection, and brief generation. Because Perplexity cites its sources, the cluster is grounded in what is actually ranking today rather than Claude’s training cutoff.

In our tests this approach completed a 15-article cluster brief in approximately 3.5 hours for a non-technical operator — significantly faster than manual research while remaining accessible without any API setup. The friction is the copy-paste loop between tools. Teams running more than two or three clusters per month will feel that friction quickly and typically graduate to a scripted or autonomous solution.

The manual approach — feeding Claude a product description, asking it to generate related search queries and PAA questions, then prompting it to cluster by intent and draft a pillar-and-spoke architecture — is where most practitioners start, and it is a genuinely useful learning exercise. You can run the full workflow in Claude.ai with no additional tools, no API costs beyond the $20/month Pro subscription, and no developer involvement.

The ceiling is low. Claude’s knowledge of current SERP competition is limited to its training data, PAA coverage is incomplete, and the prompt chain requires 8–12 individual interactions to produce a cluster you would trust to publish. In our tests the manual chain averaged 9.2 hours per cluster versus 1.8 hours for the SERP-overlap script approach. It is the right starting point, not the right steady-state workflow. Once you understand what the output should look like, move to an approach that automates the research layer.

Surfer SEO has a Topical Map feature that generates cluster recommendations from its NLP model, and many teams use it to produce briefs they then hand to Claude for full drafting. The combination works well: Surfer’s Content Score keeps the output optimised for on-page signals while Claude handles the prose quality and structure.

The integration is a handoff rather than a native pipeline — you export a Surfer brief, paste it into Claude, and write a system prompt that instructs Claude to honour the brief’s keyword density and heading targets. Teams already paying for Surfer find this a natural step up; teams evaluating from scratch will find the combined cost of $89–$199/mo for Surfer plus Claude API hard to justify versus dedicated cluster automation tools.

Jasper is most useful as the brand-voice enforcement layer in a Claude cluster workflow: Claude handles the keyword expansion, SERP clustering, and brief generation, while Jasper’s Brand Voice memory ensures every spoke article sounds like it was written by the same author. For teams publishing clusters across multiple product lines with distinct audience voices, this split is genuinely useful.

The weakness is that Jasper’s own SEO and clustering features are underpowered relative to dedicated tools, so you end up using it purely as a writing layer and relying entirely on Claude for the upstream research. That works, but it is an expensive way to add brand voice enforcement when the same outcome can be achieved with a well-crafted Claude system prompt. Useful inside an established Jasper workflow; hard to justify as a greenfield purchase for clustering alone.

Some SEO practitioners use ChatGPT o3 and Claude in sequence: o3 for deep reasoning on pillar keyword selection and cluster architecture decisions, Claude for brief generation and prose drafting. The argument is that o3’s extended thinking catches structural errors — pillar keywords that are too narrow, spoke articles that duplicate intent — before content production begins.

In our tests the dual-model approach produced marginally better cluster architecture scores (4.1/5 versus 3.9/5 for Claude alone on the same seeds) but at 2.4x the time cost and 2x the monetary cost. For a high-stakes product launch where the pillar page must rank, the validation step has merit. For routine cluster production at scale, the gains do not justify the overhead. Most teams will get better ROI from tightening their Claude system prompt than from adding a second model to the pipeline.