llm-seo-lab
Path-breaking AEO / LLM-SEO platform. Closed-loop autonomous citation engineering powered by the Claude Code CLI subscription (no per-token API), TRIZ-driven design, and attractor-flow orchestration.
What this is
The AEO / GEO / LLM-SEO category in 2026 is dominated by monitoring dashboards (AthenaHQ, Profound, Otterly.AI, Peec.ai, Geol.ai, Goodie, Search Party, SerpAPI). They surface visibility gaps but they do not close them. llm-seo-lab exists to close that loop: detect the gap, generate the fix, publish the artifact, verify the citation lift.
North-star contradiction (TRIZ frame)
Measure vs. Act. Every incumbent improves measurement of AI citation share at the cost of human effort to act. The Ideal Final Result is a system in which measurement and intervention are the same loop — the act of measuring closes the gap.
Components
| Surface | Purpose | Status |
|---|
Dual-target plugin (plugin/) | Same source tree, two manifests (.cursor-plugin/plugin.json for Cursor, .claude-plugin/plugin.json for Claude Code CLI). Commands /aeo:bootstrap, /aeo:loop, /aeo:audit, /aeo:track, /aeo:fix, /aeo:results, /aeo:compete. Hooks: SessionStart (Sākṣī witness), PreToolUse (claude-CLI guard), Stop (compaction nudge). | v0.2.0 |
MCP server (mcp/) | 16 JSON-RPC tools at POST http://127.0.0.1:7301/rpc — audit_page, track_citations, generate_brief, emit_schema, compare_competitors, oracle_query, read_config, open_pr (with live clone+commit+push+create mode), list_sites, read_latest_audit, list_prs, read_citation_trend, plus 3 widgets. Token-bucket rate limiting, fail-open Claude fallbacks, full HTTP integration test in CI. | v0.2.0 |
Pratyakṣa epistemology gate (vendored at tools/pratyaksha/) | Witness invariants (Sākṣī), conflict detection, sublation-with-evidence. Wired into the AEO loop runner as the Manas/Buddhi pair: Manas drafts the brief, Buddhi consults Pratyakṣa before the PR opens. Adoption rationale in docs/decisions/2026-04-26-pratyaksha-integration.md. | v0.2.0 |
Claude Code skills (skills/) | aeo-audit, citation-oracle-loop, content-brief-from-gap, schema-generator, freshness-radar, competitive-citation-intel | v0.2.0 |
Next.js web dashboard (apps/web) | Consumer-facing surface; orchestrates the Claude Code CLI worker via the MCP HTTP bridge (POST /rpc on :7301) and a WebSocket | v0.2.0 |
CLI worker daemon (packages/cli-worker) | Drives the Claude Code CLI subprocess; persistent JSONL job queue; rate-limits to subscription quotas; streams events to the web app over WebSocket; honours payload.dry_run for safe end-to-end smokes | v0.2.0 |
Repository layout
llm-seo-lab/
├── docs/
│ ├── research/ # Research expansions, competitor matrix, citation mechanism studies
│ ├── triz/ # Contradiction cards, ARIZ session logs, principle applications
│ ├── spec/ # Design spec for plugin/MCP/skills/app
│ ├── prd/ # Product requirements + pricing
│ ├── plans/ # Implementation plan (TDD, bite-sized tasks)
│ ├── benchmarks/ # Methodology + statistical analysis vs SOTA
│ ├── use-cases/ # Real-site validation reports
│ └── whitepaper/ # Substack-publishable long-form
├── plugin/ # Cursor plugin (commands + agents)
├── mcp/ # MCP server (Python/FastMCP)
├── skills/ # Claude Code skills bundle
├── apps/web/ # Next.js dashboard
├── packages/
│ ├── cli-worker/ # Claude Code CLI orchestration daemon
│ └── shared/ # Shared types, utilities
├── benchmarks/ # Benchmark harness scripts
├── tools/ # Vendored tooling (e.g. attractor-flow submodule)
└── .github/workflows/ # CI
Methodology
The product is not designed by intuition. The design pipeline is:
- Research expansion — extend the existing competitive landscape research with deep-dives into citation mechanisms and the GEO evidence base.
- TRIZ contradiction mapping — formalize the five candidate contradictions, run the matrix, formulate Ideal Final Results, deepen with ARIZ-85C on the top-1.
- Solution generation — TRIZ-led divergence (6–8 sketches scored on the 4×25 rubric) followed by attractor-flow convergence (basin depth + Lyapunov + perturbation) to two finalists.
- Spec → PRD → Plan — gated documentation written before any product code.
- Build — TDD, subagent-driven implementation across worktrees.
- Statistical benchmarking — power analysis, two-proportion z-tests, Bonferroni correction, bootstrap CIs, against the SOTA tools.
- Real-site proof — pre/post measurement on
technektar.dev, technektar.substack.com, a SharathSPhD GitHub Pages site, and two additional indie sites.
SharathSPhD
