Audit a site for search and AI answers, and know which points are real

Four Claude Code skills measure a site, fix the source, and prove that nothing else changed. Every finding carries an evidence tier, and the score is reported with and without the points that no engine has been shown to use.

Install Source on GitHub

audit.sh on a live site, 2026-10-07 (lines shortened)

$ bash audit.sh https://one-scroll-bible.com/
== https://one-scroll-bible.com/   (geo-optimizer-skill 4.18.3)
   score 90/100 (excellent)   excluding Tier C: 68/76 = 89%
   robots=18/18, llms=16/18, schema=16/16, meta=14/14, content=12/12,
   signals=3/6, ai_discovery=6/6, brand_entity=5/10, penalty=0
   [B] Add sameAs links in Organization schema to Wikipedia, Wikidata, ...
   [C] Add RSS/Atom feed and link it in <head> ...
   [C] Add potentialAction (SearchAction) to WebSite schema ...
   [SEC] This page exposes user-generated content ('respond' region) ...

Evidence tiers

AI search checkers mix three kinds of advice in one list: documented crawler behavior, plausible practice, and conventions such as llms.txt that no major engine has been shown to read. Google's AI optimization guide (updated 2026-07-10) says Google Search ignores AI text files and special markup.

x-eo labels every finding with a tier. A score that rose only through Tier C points is reported as such.

Tier	Meaning	Examples	Treatment
A	Documented by the engine or directly observable	HTTP 200, content in raw HTML, canonical, hreflang, sitemap, JSON-LD that parses and matches the page	Fix first
B	Supported only by third-party or correlational studies	Answer-shaped passages, freshness, consistent entity name and `sameAs`	Do it when it also helps readers
C	Convention without demonstrated use	`llms.txt`, `ai.txt`, `/ai/*.json`, WebMCP	Only if free; never counted as a result
SEC	Security finding	Prompt-injection surface, exposed secrets	Confirm a real write path before reporting

Four skills, one loop

x-eo-auditMeasure the live site
x-eo-fixChange the source, not the output
x-eo-verifyBuild diff, JSON-LD, JavaScript on and off
deployThen audit the live URL again

After deploy, the loop starts again from x-eo-audit with the same tool and version.

x-eo holds what the other three share: the tiers, honesty rules, report template, 13 pitfalls, a Korea guide (Naver, Daum, KakaoTalk), and 12 topic guides from claude-seo (MIT).

Skill	Scripts
`x-eo-audit`	`audit.sh` (score with and without Tier C, `--threshold` for CI), `host-root-check.sh`, `site-sample.py`, `ai-recall.mjs` (citations of your URLs by search-enabled AI APIs)
`x-eo-fix`	`gen-robots.py`, JSON-LD templates, root files for sub-path sites, `html-to-llms-full.mjs`, `gen-ai-files.mjs`, `indexnow.mjs`
`x-eo-verify`	`diff-builds.sh`, `jsonld-lint.py`, `render-diff.mjs` (Playwright), `local-audit.py`

Install

Copy all four skills; they reference each other. There is no installer, hook or background agent.

Clone and copy the skills

git clone --depth 1 https://github.com/jungrok5/x-eo.git
mkdir -p ~/.claude/skills && cp -r x-eo/skills/* ~/.claude/skills/

Ask Claude Code
Audit https://example.com with x-eo, fix the Tier A findings, and verify.

Requirements: bash, git, python3 3.10 or later. render-diff.mjs also needs playwright-core and a Chromium path in PLAYWRIGHT_CHROMIUM. ai-recall.mjs needs an Anthropic or OpenAI API key and costs one search-enabled request per question and engine.

Run the scripts without an agent

S=x-eo/skills
bash    $S/x-eo-audit/scripts/audit.sh --threshold 80 https://example.com/
bash    $S/x-eo-audit/scripts/host-root-check.sh https://example.com/blog/
python3 $S/x-eo-audit/scripts/site-sample.py https://example.com/ --max 8
node    $S/x-eo-verify/scripts/render-diff.mjs https://example.com/
python3 $S/x-eo-verify/scripts/jsonld-lint.py ./dist --require WebSite --strict
python3 $S/x-eo-fix/scripts/gen-robots.py --sitemap https://example.com/sitemap.xml --out robots.txt

Example: a 214-language static site, 62 to 90

12 of the 28 points were Tier C. Without them the score moved from 52/76 to 68/76.

The change with measurable value was prerendering the Korean root page: text visible without JavaScript rose from 11% to 53% of the rendered page, and the JavaScript-on DOM stayed byte-identical.

Category	Before	After	Tier
`robots`	15	18	A
`llms`	10	16	C
`ai_discovery`	0	6	C
`schema`	7	16	B
`content`	11	12	A
`penalty`	-3	0	A/B
Total	62	90

Full write-up

Limits

audit.sh assigns tiers with keyword rules. A label can be wrong; read the finding.
Scores come from one tool and its author's weights. Compare the same tool and version before and after, not different tools.
ai-recall.mjs takes a sample, not a rate. Answers vary between runs; Perplexity is not covered.
local-audit.py patches geo-optimizer internals and can break on upgrade; the version is pinned.
Tested on Linux. The evidence ledger and the Korea guide are dated; check them again before relying on them.