seo-keyword-research— agent skill

Overview

Most keyword research dies at the head term. A DR-25 site chasing "project management software" (KD 72) loses to Asana and Monday.com forever, but the same audience also searches "free task tracker for two-person agencies" (KD 8, weak SERPs, transactional intent), the keyword class that converts on month-three traffic instead of year-three. This skill is the funnel that finds those reliably: seed, modifier expansion, SERP-mined patterns, and competitor gap, then a SERP-authority rubric that beats raw KD, then a prioritized brief mapped to content types and intent.

Mental model

Long-tail wins on four compounding axes that head-term chasing loses:

1. SERP authority profile beats KD. KD is a dumb average; what matters is whether the actual top 10 are beatable. A "KD 35" keyword topped by three generic 1,500-word AI posts is more winnable than a "KD 18" one held by Wikipedia, a 5-year-old definitive guide, and the official product page.
2. Intent specificity is the conversion lever. "[head term] for [audience] under [constraint]" attracts only buyers. A 100-volume long-tail at 8% conversion beats a 10,000-volume head term at 0.2%.
3. Long-tail clusters earn topical authority. Five 200-volume long-tails on one sub-topic, internally linked to a pillar, signal more expertise than one head-term page. Google's core ranking (which absorbed the Helpful Content System in March 2024) rewards original information, depth, and entity-coverage breadth, not surface listicles.
4. Query fan-out rewards coverage, not page count. Google's AI features issue concurrent related sub-queries (Google AI optimization guide, May 2026). A cluster that genuinely answers them gets surfaced; a page per variation built to manipulate AI answers is scaled content abuse, and Google says AI search needs no extra long-tail keywords. Map fan-out questions to sections first, new pages only for distinct intents.

2026 algorithmic context (kill-rules upstream of the funnel)

Instructions

Core rules

• Long-tail = 4+ words OR a clear intent-specific modifier on a head term. "Best CRM" is short-tail; "best CRM for solo financial advisors" is long-tail. Specificity matters, not word count.
• SERP review is non-negotiable. Inspect the actual top 10 for every keyword that survives expansion. KD from any tool (Ahrefs, SEMrush, SE Ranking, Serpstat) is a starting estimate, not a verdict.
• Realism rule (heuristic): target keywords whose top-10 average DR is within roughly +15 of your DR. A DR-25 site can credibly rank against DR 10–40; a top 10 averaging DR 60+ is a multi-year project. Widen the band when the top 10 are weak or old, narrow it when they are definitive.
• Information gain is the on-page kill rule (Google's self-assessment asks for "original information, reporting, research, or analysis"; "information gain" itself is patent language, not a documented ranking system). If you can't name what the new piece adds over the top 10 (data, perspective, structure, freshness), the keyword isn't yours yet; picking it without an angle is the same mistake as picking a head term without DR.
• Intent must match the site's commercial position. Informational keywords on a transactional product page, or commercial-investigation keywords on a thin glossary page, neither rank nor convert. Classify intent before assigning content type.
• Five+ candidates per sub-topic, not one. Single-keyword targeting anchors to whatever came first; even if four are dead ends, the comparison clarifies why the survivor wins.
• Zero-volume keywords are a real 2026 category. AI search shifts demand to long-tails that tools mark "0"; one ranking for 12 related semantic queries can earn 200+ visits/month. Check related-keyword overlap instead of auto-striking.
• GEO is the second axis of every keyword, not a separate discipline. Google's AI-feature visibility comes from the same ranking systems: no special phrasing, chunking, schema, or llms.txt (Google AI optimization guide, May 2026). ChatGPT, Perplexity, and Claude cite pages that answer definition, question, and comparison queries clearly. Note GEO potential per keyword; on-page execution belongs to another domain.
• Output a brief, not a verdict (see Stage 6); writing, schema, internal-link wiring, and outreach are downstream.
• Compound over spike. Trend keywords decay within 12 months; "how to [permanent task]" compounds for years.

The funnel

#### Stage 1 — Profile the site (kill or proceed)

Get concrete answers before generating any list; if unclear, stop and ask.

1. Niche: one sentence, plus sub-niche.
2. Audience: the specific segment that buys, not "small businesses."
3. Current DR: rough estimate (Ahrefs free tool, Moz Link Explorer, or "we're new"). Sets the realism cap.
4. Existing content: count and coverage. For an existing site, check what already ranks (Search Console queries) before generating overlap.
5. Commercial position: informational, affiliate, SaaS, e-commerce, local services. Decides which intent classes to prioritize.
6. Geo + language: global English, US-only, multi-region, non-English. Affects SERP review and modifier set.

If the user can't name DR or commercial position, kill here; qualification would be blind: "I'd be guessing whether the top 10 are beatable. Get a DR estimate (Ahrefs free Backlink Checker takes 30 seconds) and tell me what conversion looks like, then I'll run the funnel."

#### Stage 2 — Generate seed keywords (parallel)

Fire all seed sources in ONE parallel block:

• Core product/service nouns, in buyer language, not internal jargon.
• Problem phrases: "[problem] when [context]", "[symptom] in [audience]", "stop [bad thing]".
• Solution phrases: "how to [task]", "[task] without [obstacle]", "alternative to [incumbent]".
• Audience qualifiers: "[product] for [segment / team size / use case]".
• Industry / vertical terminology from the buyer's world.
• Existing-content mining: Search Console queries pulling impressions but ranking page 2–5; each seeds related long-tails.
• Competitor SERP gaps: 3 competitors at the same DR tier, their top-200 ranking keywords (Ahrefs Site Explorer / SE Ranking competitor research / SEMrush Domain Overview), filtered to those the user's site doesn't rank for.

Output: 30–80 seed terms, unqualified.

#### Stage 3 — Expand to long-tail (modifier patterns + SERP mining)

Generate 5–15 variants per seed, all seeds in parallel, from the modifier buckets in references/modifier-library.md: question (highest GEO value), intent (commercial/transactional), constraint (highest yield for low-DR sites; every constraint kills a competing page), outcome, comparison (mid-funnel, AI-citation-friendly), and problem-trigger (error messages, support-forum phrasings; highest conversion when the product solves the pain).

Fire the SERP-mined sources listed there (autocomplete a–z, PAA, Related searches, Trends rising, Reddit/forums, Lowfruits, Glimpse / Exploding Topics) in one block. The web search tool isn't Google: autocomplete, PAA, and Related searches come from the user, a connected SERP tool, or public PAA tools; list any source you couldn't reach as unused in Method Notes.

Output: 200–500 candidates, deduped at Stage 4.

#### Stage 4 — Qualify each long-tail (SERP-first rubric)

This separates winnable keywords from KD-low-but-unwinnable ones. Walk top to bottom and STRIKE on the first failure; it's a kill list, not a score.

Survive all 12 or strike. For zero-volume candidates, replace check 2 with "appears in 5+ Reddit threads OR in PAA / autocomplete?": yes means a real query with no tool data, keep; no means strike.

#### Stage 5 — Cluster, map entity coverage, prioritize

1. Pick pillar candidates: broad seeds with the most survivors around them.
2. Group survivors under pillars by sub-topic (one pillar + 4–8 cluster pages).
3. Map entity coverage per cluster: named entities, sub-topics, and concepts the pillar must cover, from the top-10 pages' H2/H3 structure (n-gram-mined), Wikipedia "See also" for the seed, and knowledge-panel related entities. A pillar missing 3+ entities the whole top 10 covers signals shallow coverage.
4. Score each survivor:

• Realism: top-10 authority vs user DR (1 = unwinnable, 5 = clearly beatable).
• Intent value: informational 1, navigational 1, commercial 2, transactional 3, multiplied by relevance to the site's commercial position.
• GEO potential: definition / how-to / comparison / list, 1 each (max 4).
• Cluster fit: pillar 5, primary cluster page 4, secondary 3, orphan 1.
• AIO discount: subtract 2 if an AI Overview is present (or unverified) on an informational keyword AND the page has no information-gain angle.

Composite priority = (realism × 3) + (intent × 2) + GEO + cluster_fit − AIO_discount. Range 9–35.

1. Bucket: P0 quick wins (≥ 28, publish first month), P1 high-value (22–27, next sprint), P2 cluster fillers (16–21, batch with pillar), P3 monitor (< 16, not this quarter). Realism 1 forces P3 regardless of total.

#### Stage 6 — Output the keyword brief

Fill every section of references/brief-template.md (P0–P3 tables with intent, content type, and calendar slot; clusters with entity coverage, GEO notes, risk notes, existing-content audit, downstream open items). The brief is the deliverable; anything outside it is downstream.

Decision guide — common edge cases

Tools

Free-vs-paid tool table: references/modifier-library.md. Use a connected SEO-data MCP (Ahrefs, SE Ranking, Serpstat, Search Console) for volume, KD, and SERP features if present; none is assumed. Paid tools never replace the manual Stage 4 SERP review.

Examples

Example 1: SaaS, mid-authority, finds quick wins

Open-source self-hosted CRM, DR 28: the rubric keeps "self-hosted CRM for sales teams under 10" (KD 14, top-10 avg DR 22) as P0 and zero-volume "CRM Docker compose" as P1, strikes "best CRM software" (KD 78) and "CRM" (branded), and clusters survivors under a "Self-hosted CRM: complete guide" pillar. Full walkthrough: references/brief-template.md (worked example).

Example 2: New site, no DR — refuse the head term

User: "We just launched. We sell project management software. Find me 50 keywords."

Stop. At a brand-new domain (DR ~0–5), 80% of "project management software" keywords are unwinnable for 12+ months; Asana, Monday, ClickUp, and Notion own them with decade-old backlink graphs. Fifty generic PM keywords would waste your first quarter. Two questions first:

>

1. Which segment converts that the incumbents underserve (e.g. "construction subcontractors with 3–15 person crews", "law firms doing matter management")? Narrower segment, more long-tail surface. 2. What's the 1–2-sentence reason to switch (feature gap, pricing model, integration, vertical workflow)? That becomes the Stage 3 modifier set.

>

If "we don't know yet", positioning is upstream of keyword research: pick the niche, then run the funnel.

Example 3: GSC mining for free quick wins

For any site live ≥ 3 months with Search Console, run this first; it's the highest-ROI single source. Pull queries with impressions > 100 and average position 8–20 (page 2), confirm intent against the existing page, and SERP-review for room to climb. Matching intent: a 200-word section plus an intent-aligned H2 typically lifts it into the top 10 in 4–8 weeks. Mismatched intent or incidental overlap: a new dedicated page that outranks the parent.

Checklist

• [ ] Stage 1 profile captured: niche, audience, DR, existing coverage, commercial position, geo
• [ ] DR and commercial position confirmed before any keyword generation (kill-gate)
• [ ] Seeds from ≥ 4 of the 7 source categories, run in parallel
• [ ] 5–15 variants per seed from the modifier library
• [ ] Manual SERP review for every candidate surviving volume + intent (rubric check 3)
• [ ] Realism rule applied: top-10 average DR within ~+15 of user DR, banded by content-quality gap
• [ ] AIO presence recorded from real evidence or marked Unverified; informational AIO keywords without an information-gain angle demoted
• [ ] Site Reputation Abuse risk flagged on "best [thing] [year]" keywords dominated by Forbes/CNN/Times brand pages
• [ ] Information gain articulated per P0 keyword
• [ ] Zero-volume candidates kept with Reddit/PAA validation or struck, never auto-skipped
• [ ] Survivors clustered into pillar-and-cluster structures, not flat lists
• [ ] Entity coverage mapped per cluster (top-10 H2/H3 mining, Wikipedia "See also", knowledge-panel entities)
• [ ] P0–P3 bucketing via the composite formula incl. AIO discount; realism 1 demotes to P3
• [ ] GEO potential noted per keyword (question / definition / comparison / list)
• [ ] Brief follows references/brief-template.md; downstream concerns listed as Open Items
• [ ] Cannibalization checked: any P0 keyword already targeted by an existing page?
• [ ] AI search visibility tracking listed in Open Items for GEO-flagged P0/P1 keywords

Composition / References

In-skill references (load on demand):

• references/modifier-library.md — full modifier library, SERP-mined sources, and free-vs-paid tool table for Stage 3
• references/brief-template.md — canonical Stage 6 brief template, field definitions, and the full Example 1 walkthrough

Within-domain pairings:

• The sibling SEO skill that prospects backlinks for cluster pillar pages (priority for any P0 cluster lacking link equity).
• The sibling SEO skill that validates free-tool ideas: a cluster surfacing a recurring user problem (calculator, checker, generator) is a tool-build trigger.

Downstream concerns (other domains):

• GEO content writing (clear answers, information gain, crawlable text); the brief only flags GEO-eligible keywords.
• Content briefs, internal-link wiring, on-page schema, outreach copy consume this brief.
• AI search visibility tracking setup (Profound / Otterly / AthenaHQ / Peec); the brief lists which keywords to track.

External authoritative sources:

• Ahrefs — long-tail keyword research
• Backlinko — keyword research definitive guide
• Google Search Central — Helpful Content guidance
• Google Search Central — Site Reputation Abuse policy
• SparkToro — zero-volume keyword research framework
• Search Engine Journal — SERP analysis for keyword qualification
• Search Engine Land — AI Overviews CTR impact studies 2024–2026
• Google Search Central — Optimizing for generative AI features (May 2026) — query fan-out, scaled content abuse, mythbusting
• Princeton/Georgia Tech — GEO research paper (Aggarwal et al) — original framework for AI-citation optimization