How Claude Code Runs a Full SEO Team: The v2 Architecture Explained

BLUF: SEOKit v2 ships 19 commands, 4 auto-loading skills, and 2 read-only specialist agents — all measured at 16,004 tokens. There is no orchestrator agent, no blocking reviewer gate, and no Python tooling. Every command ends with a falsifiable deliverable: a diff, a ranked backlog, or a verified file. If you have been waiting for Claude Code to do defensible SEO work, this is the architecture that makes it possible.

AI Overviews now appear on 48% of Google queries (March 2026, up from 34.5% in December 2025). Ranking and AI citation correlate at only r=0.18, and 47% of AIO citations come from below position 5. That means the old playbook — rank high, get traffic — is no longer sufficient. You need a system that targets both traditional ranking signals and AI extractability simultaneously. SEOKit is that system.

What does SEOKit v2 actually contain?

SEOKit is part of the ClaudeKit product family, which now spans 5 kits, 101 commands, 19 skills, and 13 read-only agents totaling 82,197 measured tokens across the full suite.

SEOKit's own numbers: 19 commands, 4 skills, 2 agents, 16,004 tokens.

The flagship commands are /seo quick-wins and /seo citations. Quick-wins targets positions 8-20 plus low-CTR pages — the two buckets where incremental work has the highest return. Citations runs N-iteration AI-citation measurement with confidence intervals, so you have actual data on whether your pages are being cited in AI Overviews and Perplexity, not guesses.

The full command roster includes: seo-audit, seo-write, seo-check, seo-pseo, seo-extractable, plus supporting commands for keyword research, schema, internal linking, and competitor gap analysis. See the complete list on the SEO kit page.

Why did we remove the orchestrator and reviewer gate?

The original v1 SEOKit post described a 15-agent system with an seo-orchestrator, an audit-scorer, a revenue-prioritizer, and a content-reviewer blocking gate. That architecture sounded impressive. It was also wrong for how Claude Code actually works.

Here is what happened in practice:

1. The orchestrator accumulated context across the entire session. By the time specialist agents returned findings, the main context had bloated — expensive and fragile.
2. The blocking reviewer gate created false precision. Scoring a draft 86 vs. 92 on "E-E-A-T signals" was a number we invented. It added friction without adding accuracy.
3. Orchestrators re-introduced the single-agent failure mode they were supposed to solve. One corrupted state meant everything downstream was wrong.

The v2 answer is structural: commands end with EVIDENCE, not a reviewer gate. When you run /seo quick-wins, you get a ranked markdown backlog with position data, CTR delta, and a modeled revenue estimate attached to each item. That backlog is the evidence. You review it. You decide what ships. No agent blocks you; no gate invents a score.

The two read-only agents we kept are genuinely useful:

• seo-researcher — pulls live SERP data, entity context, and AI-citation status for a target keyword cluster. Read-only. Returns a structured findings file.
• seo-auditor — runs technical and indexation checks (crawl health, status codes, canonical logic, robots, GSC coverage gaps) and returns a findings diff. Never touches your files.

Neither agent orchestrates anything. Neither blocks the workflow. They produce findings that a command then processes into a deliverable.

For a deeper look at why we moved away from the reviewer-gate pattern, see our post on agents vs. skills vs. slash commands.

How does /seo quick-wins actually work?

Quick-wins is the command most teams run first. Here is the step sequence:

1. Load seo-context skill (auto-loaded, covers your site's current keyword map, revenue model, and GSC integration).
2. Pull GSC data for positions 8-20. Filter to pages where impressions exceed 200/week — you want real traffic intent, not tail noise.
3. Cross-reference against low-CTR pages (CTR below expected for position). These are pages where rank is fine but the title/meta is leaking clicks.
4. Run the seo-auditor agent on the shortlisted URLs for crawl and indexation issues.
5. Score each item: impact x confidence x revenue proxy.
6. Output a ranked markdown backlog with a leading indicator per item (what metric, what threshold, what timeline).

The output is a file you can drop directly into a sprint. We measured the full quick-wins flow at roughly 3,200 tokens on a 200-URL site. That is not a rough estimate — ck tokens seo will recount any time you want.

Perplexity data shows 90% of top-cited sources answer in the first 100 words of a page. Quick-wins includes an seo-extractable check for each targeted page that flags answer-first structure issues. You fix the ranking and the AI-extractability in the same pass.

How does /seo citations measure AI citation rates?

Citation measurement is the part of SEOKit that does not exist anywhere else. The problem: AI Overviews and Perplexity cite unpredictably. A page can rank 3rd and never get cited. A page at position 11 can appear in 60% of AI Overview runs for that query.

/seo citations solves this with N-run measurement:

1. You specify a keyword cluster and a sample size (default N=20 runs per query).
2. The command queries each target keyword N times across different phrasings — leveraging the fact that one prompt fans out to 8-16 sub-queries.
3. It records which URLs appear in AI Overview citations, Perplexity top-3, and ChatGPT search results per run.
4. It calculates citation rate with confidence intervals. A page cited in 12/20 runs has a 60% rate with a 95% CI of [38%, 78%] — that is actionable. "We appeared in AI Overviews sometimes" is not.

We cross-reference against schema coverage (schema'd pages get cited at 47% vs. 28% for non-schema'd pages, per 2026 data) and first-100-words answer structure. The output is a per-page citation scorecard with specific fix recommendations.

ChatGPT search rides the Bing index; pages updated within 3 months average 6 citations vs. 3.6 for stale pages. The citations command tracks your update recency as a leading indicator. Claude search uses the Brave index (86.7% citation overlap with Brave top results). Knowing which index you are targeting matters — the command makes that explicit.

How does SEOKit compare to running raw Claude Code prompts?

The honest answer: for single-page tasks, you may not need SEOKit. One good prompt can rewrite a meta title. Where the kit earns its cost is on the systematic work.

The gap widens as scope grows. For programmatic SEO work — template-plus-dataset systems at hundreds of pages — the structured output and thin-content guards in seo-pseo are genuinely hard to replicate with a raw prompt.

What does installation look like?

SEOKit installs through the claudekits CLI (v0.1.3 on npm) or through the Claude plugin marketplace.

CLI path:

npm install -g claudekits
ck auth <your-key>
ck install seo

The token ledger prints on install. Global install lands in ~/.claude; pass --local for project-scoped install. ck tokens seo recounts at any time. ck doctor diagnoses path and permission issues. ck list shows your entitlements.

Plugin marketplace path:

/plugin marketplace add Madni-Aghadi/claudekit-seo

Pricing: $14.99/month for SEOKit alone. $29.99/month for Pro (any 3 kits, swap 1 per cycle). $49.99/month for All-Access. Annual pricing at $119/$239/$399. Lifetime access is $99 one-time per kit as shipped (no future updates). 14-day refunds, 3 devices per license. Full details on the pricing page.

If you are already using MarketingKit, the two work well together. SEOKit handles the keyword and technical layer; MarketingKit's /mkt repurpose turns your optimized posts into 5 distribution formats. They share the same install pattern and token accounting.

What is the real token cost of running SEOKit?

SEOKit measures at 16,004 tokens if all 4 skills load simultaneously. In practice, skills load on demand — a /seo quick-wins run does not load the AEO-optimization skill unless you ask for it.

We measured a typical audit session across 3 representative commands:

• /seo quick-wins on a 200-URL site: ~3,200 tokens
• /seo citations for a 10-keyword cluster at N=20: ~4,800 tokens
• /seo audit full crawl + indexation: ~5,100 tokens

That is roughly 13,100 tokens for what used to take an SEO contractor two days and cost $800-1,200. At Sonnet 4.5 rates, that session is under $0.20 in model costs.

For the full methodology behind token measurement, see our token-cost dataset post.

FAQ

Does SEOKit have an orchestrator agent that coordinates the other agents?

No. We removed the orchestrator architecture in v2. The original v1 system had a top-level seo-orchestrator that planned, fanned out, and merged — but it bloated context and reintroduced the single-agent failure mode it was supposed to solve. In v2, commands are the coordination layer. Each /seo command runs a defined sequence, calls a read-only agent for data if needed, and ends with a deliverable. No agent plans for another agent.

What are the two read-only agents in SEOKit, and what do they do?

seo-researcher pulls live SERP data, entity context, and AI-citation status for a target keyword cluster. seo-auditor runs technical and indexation checks — crawl health, status codes, canonical logic, GSC coverage gaps — and returns a findings diff. Both are read-only: they cannot write files, cannot modify your project, and do not orchestrate other agents. They produce structured findings that commands then process into a deliverable you review.

Can Claude Code replace an SEO agency?

For systematic, data-backed SEO work — audits, quick-win identification, AI citation measurement, programmatic content at scale — SEOKit covers most of what a junior-to-mid SEO contractor does, faster and at a fraction of the cost. It does not replace judgment calls about brand positioning, editorial strategy, or relationship-driven link building. The right frame: SEOKit handles the repeatable 80%, freeing you to focus on the 20% that requires human judgment and relationships.

How does SEOKit handle AEO (Answer Engine Optimization)?

The seo-extractable command audits pages for AI-extractability: answer-first structure, first-100-words coverage of the target query, schema markup, entity clarity, and llms.txt presence. The /seo citations command then measures actual citation rates with confidence intervals. The pairing gives you a fix list (extractable audit) and a measurement loop (citations). Given that 44.2% of AI Overview citations come from the first 30% of a page, the structural fixes are often the highest-leverage interventions. See our AEO vs. SEO post for the full framework.

Is there a reviewer gate that blocks content from shipping?

No. We removed the blocking reviewer gate from v2 because it created false precision — scoring content 86 vs. 92 on a made-up scale — without improving quality. In v2, every command ends with evidence: a diff, a ranked backlog, or a verified file. You are the reviewer. The seo-auditor agent flags issues; you decide what to act on. This is explained in depth in our post on agents vs. skills vs. slash commands.

How is the quick-wins backlog prioritized?

By impact x confidence x revenue proxy, in that order. Each item in the /seo quick-wins output includes: the affected URL, the current position and CTR, the modeled improvement, a specific fix recommendation, a leading indicator (what metric to watch), and a timeline. The seo-researcher agent models incremental traffic from position and CTR delta; you supply the conversion rate and AOV from your analytics, and the command attaches a dollar estimate. The top items on a real mid-size ecommerce site typically include discovered-not-indexed faceted URLs and title-tag CTR leaks — both high-confidence, high-impact, measurable within 21 days in GSC.

SEOKit is the right starting point if you are doing serious SEO work with Claude Code. The 19-command architecture covers auditing, keyword research, content writing, AI citation measurement, programmatic SEO, and schema — all with measured token costs and falsifiable deliverables. Start with the SEO kit page to see the full command roster, or go straight to pricing if you are ready to install.