SEO cluster generator

You build a defensible topic cluster from one seed keyword in a working day. Pillar page brief + 20–30 spoke briefs + internal link map + intent matrix. You do not draft the pages (that’s the eval-gated-drafting skill). You produce the planning artefact every page in the cluster references.

Inputs to gather first

1. Seed keyword the cluster should be built around
2. Target locale and language (e.g. en-GB, en-US)
3. Search Console data for the brand (last 90 days) — or Ahrefs / Semrush data if GSC isn’t connected
4. Three named competitors for SERP analysis
5. Brand differentiation — what original data, customer stories, expert access, or tools does the brand have that competitors don’t? Read .lens/positioning-brief.md and .lens/message-house.md if present.
6. Min volume threshold — default 50 monthly searches for spokes, 500+ for pillar candidates

The pipeline (six phases)

Phase 1 — Seed expansion

Pull 200+ related queries from the supplied data. Filter:

• Target locale + language
• Min volume threshold
• Strip brand-name queries

Phase 2 — Intent classification

Classify each query into the four canonical intents (informational, commercial, transactional, navigational) and the five buyer stages (problem-aware, solution-aware, product-aware, evaluating, decided).

SYSTEM: You classify search queries by intent and buyer stage. Four
canonical intents, five buyer stages. One intent + one stage per query.
You also return confidence (high / medium / low). You do not invent
middle-ground categories.

Heuristics:
- "best X" without brand qualifier → commercial / evaluating
- "X vs Y" → commercial / evaluating
- "how to X" → informational / problem-aware or solution-aware
- "X near me" → transactional / decided
- "[brand] X" → navigational / decided
- "what is X" → informational / problem-aware

USER:
Brand context: {BRAND_NAME}, category {CATEGORY}, audience {AUDIENCE}.
Queries: {QUERIES_JSON}

Return JSON array (same order as input):
[
  {
    "query": "...",
    "intent": "<informational|commercial|transactional|navigational>",
    "stage": "<problem-aware|solution-aware|product-aware|evaluating|decided>",
    "confidence": "<high|medium|low>",
    "fits_brand_audience": <true|false>
  }
]

Rules:
- One intent, one stage. Do not pick two.
- "fits_brand_audience": err on the side of false if uncertain.
- "low" confidence acceptable. Better than wrong-with-confidence.

Phase 3 — Cluster shape

Group queries by topic. Identify:

• Pillar candidate — highest volume + informational/commercial intent that subsumes the spokes
• Spokes — one page per topic, multiple queries per page
• Orphans — queries that don’t fit; signal for other clusters, not failures

Phase 4 — SERP analysis

For the pillar candidate and the top 5 spokes by volume, pull top 10 SERP results. Extract:

• Format (article / guide / tool / video)
• Length range
• Who’s ranking (publishers / brands / aggregators)
• Evidence types dominating (data, expert quote, original research, examples)

The cluster has to look credible alongside what’s ranking.

Phase 5 — Differentiation hooks

For each page in the cluster, surface 2–3 hooks where this brand can plausibly beat the existing SERP. Sources:

• Original data the brand has (customer survey, product analytics, transaction data)
• Customer stories the brand can tell
• Expert access the brand controls
• Tools or interactives the brand can ship

If a spoke has no defensible hook, flag as “competitive — uncertain win” and deprioritise.

Phase 6 — Internal link graph + briefs

Generate:

• Link graph — pillar links to all spokes; each spoke links contextually to 2–4 sibling spokes; no orphans
• Pillar brief — 3,000–5,000 words target, sections, must-include queries, differentiation hooks, evidence requirements
• Spoke briefs — 800–1,500 words each, structured the same

Output

Single document:

1. Cluster map — pillar + spokes diagram with internal links
2. Intent matrix — every page mapped to intent + stage
3. Pillar brief — full
4. Spoke briefs — all 20–30
5. Build order — recommended sequence (high-volume + clear differentiation first)

Save to .lens/seo-clusters/{seed-keyword}/.

Evals

Self-check before delivery:

• Intent classification accuracy — hand-classify 50 queries from the brand’s data; pipeline accuracy ≥85%
• Cluster coherence — pairwise embedding similarity within each spoke ≥0.6
• SERP differentiation reality — for top 5 spokes, manually check that proposed hooks aren’t already table-stakes in the current top 10
• Link graph density — no spoke has fewer than 2 incoming links

Failure modes to watch

• Seed too narrow — fewer than 12 spokes is a thin cluster. Honest read: either the seed needs broadening or the cluster isn’t defensible.
• SERP analysis sees yesterday’s SERP — differentiation hooks should be sturdy enough to survive minor SERP movement.
• Model invents search volumes — always feed actual data. If volume is missing, pause and request — never let the model estimate.
• Cluster cannibalisation — sibling spokes with overlapping primary queries dilute each other. Flag for merge if 70%+ overlap in intent + audience.
• Helpful Content Update vulnerability — spokes that can’t articulate a hook beyond synthesis shouldn’t ship.
• Pillar-spoke length mismatch — pillars 3–5x longer than spokes. If your pillar is 1,200 words and spokes are 1,000, it’s not a pillar.

Hand-off

Spoke briefs feed:

• eval-gated-drafting — drafts each spoke with eval gates against the brief
• social-content-factory — repurposes anchor spokes into channel-native posts
• earned-media-pitch — surfaces pillar stories worth pitching to journalists