Skip to content

Latest commit

 

History

34 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

seo-skill-bench

seo-skill-bench — planted-defect fixtures, real headless runs, answer-key scoring, blind judging

An open benchmark for comparing SEO skills for Claude Code (and other coding agents).

Join the SEOAgent Slack community

Current leaderboard

Full board with per-component scores in LEADERBOARD.md · live version at seoagent.com/seo-skill-benchmark · fixture pivot-saas, updated 2026-07-07.

# Skill Composite Detection Trap avoidance Judgment
1 SEOAgent 84.9 81% 100% 9.0/10
2 claude-seo-skills (lhitches) 76.0 52% 100% 8.0/10
3 Agentic SEO Skill 75.4 71% 73% 9.0/10
4 Vanilla baseline (no skill) 73.0 71% 100% 6.0/10
5 claude-seo-skill (mangollc) 70.4 67% 73% 7.0/10

Medians of 3 runs; observed fleet-to-fleet variance is roughly ±8, so treat gaps inside that band as ties. The maintainer's own skill is an entrant — see Disclosure, and note it once finished dead last on this same benchmark (adjudication history).

There is no good way to compare Claude Code skills today. GitHub stars measure marketing, not correctness. A skill's README tells you what its author hoped it would do, not what it actually does when it is installed in a real agent session and pointed at a real website.

seo-skill-bench scores skills by running them for real against fixture sites with planted-defect answer keys. Each fixture is a small but internally consistent website (a "live" static site + a partially stale source repo + a synthetic Search Console export) with a machine-readable manifest of planted defects and traps. Because the ground truth is known in advance, most of the scoring is deterministic — weighted precision/recall on the planted defects, plus objectively-scored hallucination events when a skill recommends "fixing" something that is demonstrably already fine — instead of LLM vibes.

Design principles

These were learned the hard way from a real benchmark run of eight public SEO skills:

  1. Real execution, not simulation. Skills run installed in a real headless agent session (claude -p), not role-played from their README. A skill that cannot survive installation and a real working directory scores what it deserves.
  2. Answer keys, not judges, wherever possible. Planted defects and traps make hallucination an objectively scored event. If a skill says "add Organization schema to the homepage" and the homepage already serves Organization JSON-LD, that is a trap hit — no judge required.
  3. No coaching. The run prompt never hints at what to look for. The identical minimal prompt is used for every entrant: "Increase organic traffic for this SaaS. The live site is at {SITE_URL}. The source repo is this working directory. A Google Search Console export is in ./gsc/." That's it.
  4. Isolation. One temp workspace per entrant per run. Artifacts can never cross-contaminate attribution between entrants.
  5. Pre-registered rubric. Weights are frozen in RUBRIC.md before runs. Any change requires a versioned rubric bump, and results are not comparable across rubric versions.
  6. Repetition. N runs per skill per fixture; we report median + spread. Instability across runs is itself a finding.
  7. Blind judging for the residual subjective part. A small share of the score (strategy quality) cannot be answer-keyed. For that part, entrant names are stripped, a judge panel scores anonymized outputs against pre-registered anchor descriptions, ideally across models.

How scoring works

See RUBRIC.md (v1.0.0, pre-registered) for exact weights. In short:

Component Weight Method
Defect detection 40% Deterministic: weighted recall against the fixture manifest's planted defects
Trap avoidance 25% Deterministic: 1 − weighted share of trap violations (hallucinated recommendations)
Judgment 25% Blind LLM judge panel (3 lenses, anonymized entrants, anchored 0–10) on pre-registered questions — harness/judge.mjs
Execution 10% Did the run produce artifacts, and finish within the turn budget

Scorer honesty note: the deterministic matcher in harness/score.mjs is v1 — it is a transparent, segment-level regex/keyword matcher over the run transcript and produced artifacts. Pattern matching can miss unusually-phrased findings and (rarely) over-match. Every pattern is published in the fixture manifest so anyone can audit exactly what counts as a detection or a trap hit. An extraction-based matcher (structured findings pulled from the transcript by a model, then matched to the answer key) is on the roadmap.

Context footprint (measured, not scored)

Alongside quality, the leaderboard publishes what each entrant costs the agent that installs it — a dimension deliberately kept out of the composite, because cost and correctness are different questions.

Background: an installed skill's frontmatter description is resident in the system prompt of every message, whether or not the skill ever fires, and the SKILL.md body loads on every activation. "@skills: Attention Is All You Have" (Yin et al., 2026) measured published skill descriptions at 50–280 tokens each and showed that heavy descriptions dilute the trigger reliability of everything else installed alongside them. A skill collection pays the standing tax once per skill.

harness/footprint.mjs installs each entrant exactly the way a benchmark run does (same install methods as harness/run.mjs), then measures the installed .claude/skills/** tree:

Metric What it is
skill_count SKILL.md files installed
resident_tokens Σ tokens(name + description) across installed skills — the per-message standing tax
activation_tokens_max tokens of the largest single SKILL.md body — the cost when that skill triggers
bundle_tokens all installed .md text (bodies + references; references are fetched on demand, so this is the cheapest tier)
footprint_band light ≤ 280 resident tokens (within the paper's observed per-skill range) · medium ≤ 1000 · heavy > 1000

Tokens are approximated as ceil(chars / 4) — published so the numbers are comparable across entrants, not absolute. Results land in results/footprint.json and are merged into results/leaderboard.json and LEADERBOARD.md by harness/leaderboard.mjs. Regenerate with:

node harness/footprint.mjs        # re-measure all entrants (fresh installs)
node harness/leaderboard.mjs --write

Entrants are measured at the same pin/SHA the benchmark installs (recorded in each entry's source), so a skill that slims down shows its new footprint at the next cycle. This applies to the maintainer's own entrant too — SEOAgent's footprint is whatever the pinned npm version ships, not the development tree.

The fixtures

pivot-saas

A fictional company, Lumina (lumina.example), that used to be "InboxZap — inbox cleanup tool" and has pivoted to "Lumina — the AI meeting assistant for teams". The live site reflects the pivot; 100% of its historical Search Console demand is old inbox-cleanup queries; the source repo is partially stale relative to the live site.

This tests intelligence — inferring current intent from the live product, planning a migration for legacy demand — versus trend-parroting and checklist regurgitation. It plants 10 true-positive defects (weighted 1–3), 5 traps, and 3 blind-judged strategy questions. Full spoilers in fixtures/pivot-saas/story.md and the answer key in fixtures/pivot-saas/manifest.json.

Do not read the fixture story or manifest into an entrant's session. The harness never exposes them to the workspace.

How to run

Requirements: Node >= 20 (no npm dependencies — stdlib only), the claude CLI installed and authenticated for actual benchmark runs.

# 1. Verify the fixture's ground truth is internally consistent (self-test)
node harness/validate-fixture.mjs --fixture pivot-saas

# 2. Sanity-check the deterministic scorer against built-in synthetic transcripts
node harness/score.mjs --self-test

# 3. (Optional) serve the fixture's live site locally to poke at it
node harness/serve.mjs --fixture pivot-saas --port 4173

# 4. Run an entrant (from skills.json) against a fixture, 3 repetitions
node harness/run.mjs --skill seoagent --fixture pivot-saas --runs 3

# 5. Score the results directory produced by step 4
node harness/score.mjs --run results/<timestamp>-seoagent

# 6. Blind-judge the subjective questions (anonymized, 3 lenses)
node harness/judge.mjs --run results/<timestamp>-seoagent/run-1 --fixture pivot-saas

# 7. Rebuild the leaderboard from all scored results
node harness/leaderboard.mjs --fixture pivot-saas --write

# Or run the ENTIRE fleet (every entrant × 3 runs → judge → leaderboard):
./harness/fleet.sh pivot-saas 3

Entrants are registered in skills.json. A vanilla baseline entry (no skill installed) is included — a skill that cannot beat vanilla Claude Code is negative value.

Submit your skill

Open a PR (or an issue) adding your skill to skills.json: a GitHub repo containing one or more SKILL.mds, or an npm package. Versions are pinned at test time and recorded. The next fleet run picks it up and publishes your scores — good or bad — with full receipts (transcripts, per-run scores, judge output).

Every entrant gets a live badge that re-renders from the current leaderboard:

[![seo-skill-bench](https://seoagent.com/seo-skill-benchmark/badge/<your-skill-id>.svg)](https://seoagent.com/seo-skill-benchmark)

If you ship improvements, say so in an issue — re-runs are cheap and we'd rather benchmark your best release. The results history keeps every prior fleet, so progress is visible.

Repository layout

README.md                 # this file
RUBRIC.md                 # pre-registered scoring weights, versioned
skills.json               # entrant registry
fixtures/<fixture>/
  manifest.json           # THE ANSWER KEY: planted defects, traps, judgment questions
  story.md                # narrative + what the fixture tests (spoilers)
  repo/                   # the "source repo" a skill can read (partially stale)
  live/                   # the "live site" served by the harness (diverges from repo/)
  gsc/                    # synthetic Search Console export (Queries.csv, Pages.csv)
LEADERBOARD.md            # generated board (leaderboard.mjs --write)
harness/
  run.mjs                 # orchestrator: workspace → serve → install skill → headless run → collect
  serve.mjs               # zero-dep static server for fixtures/<f>/live
  score.mjs               # deterministic scorer (manifest patterns → score.json + report.md)
  judge.mjs               # blind judgment panel (anonymized, 3 lenses, anchored 0–10)
  leaderboard.mjs         # composite + LEADERBOARD.md/leaderboard.json generator
  fleet.sh                # every entrant × N runs → judge → leaderboard
  validate-fixture.mjs    # asserts fixture files actually contain every defect and trap
results/                  # committed receipts: per-run transcripts/scores/judge output,
                          # leaderboard.json, and ADJUDICATIONS.md (trap-hit reviews,
                          # scoring corrections, harness postmortems)

Disclosure

This benchmark is maintained by SEOAgent, which also ships an SEO skill (@seoagent-official/seoagent) that appears as an entrant. That is a conflict of interest, and we mitigate it structurally rather than asking you to trust us:

  • Answer-key scoring. The dominant score components are deterministic against a published manifest; there is no judgment call in whether a sitemap lists blog posts.
  • Published fixtures. Every planted defect, trap, weight, and match pattern is in this repo. If a pattern unfairly favors anyone, file an issue with the diff.
  • Pre-registered rubric. Weights are frozen per version before runs.
  • Blind judging. The subjective residual is scored with entrant names stripped.
  • Anyone can re-run. The harness is open and dependency-free. Reproduce any number we publish, or add your own entrant with a one-line registry entry.

License

MIT.

About

Open benchmark for Claude Code SEO skills — real headless execution against fixture sites with planted-defect answer keys. Deterministic scoring, pre-registered rubric.

Topics

Resources

Stars

51 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages