Repository navigation
Releases
Releases · Olexan-art/ai-search-audit
ai-search-audit 2.113 — drops with a reason, --explain
- The agent can drop a finding only by citing a
dropWhenclause from the check's rubric (--drops drops.json). A drop without a valid clause is refused, and every dropped finding stays in the report in its own group, "Dropped by the agent", with the clause and the reason. Nothing disappears silently. --explain <code>prints one check's title, severity, sources and rubric, and--codeslists every check, so the agent no longer reads the whole reference to explain one finding.- A link to GLLmolfar, the GetLoopLoop desktop crawler, in the report footer, for when one page is not enough.
- Vectorscope 2.116 inside.
ai-search-audit 2.112 — modern-web-guidance checks, plugin install, update check
- Headings folded with
hidden="until-found"are no longer reported as hidden headings (a closed<details>never was, and a test now pins it). The fix for hidden headings names<details>,hidden="until-found"andbeforematch. - Three new low-severity checks: render-blocking scripts in
<head>(consent and A/B-test snippets are named as often intentional), a preloaded image the same page loads lazily, andfetchpriority="high"on more than two images. A missingfetchpriorityis never reported. - modern-web-guidance is cited as a further source for the image-priority, blocking-script and hidden-content rules, pinned to one commit. Google's own pages stay the source for every search claim.
- The six checks most prone to false positives carry
keepWhen/dropWhenclauses, in the references and in the JSON export. - Install as a plugin in Google Antigravity (
agy plugin install https://github.com/Olexan-art/ai-search-audit) or Claude Code (/plugin marketplace add Olexan-art/ai-search-audit, then turn on auto-update in/plugin). - A daily update check with a notice in the report, which also counts installs and runs anonymously (skill, version, OS, a random install id, runs; nothing about the audited pages). Off with
GETLOOPLOOP_NO_UPDATE_CHECK=1or--no-update-check. - Vectorscope 2.114 inside.
ai-search-audit 2.110 — fewer tokens, fewer errors
- Agent instructions rewritten in controlled technical English (ASD-STE100): rules, procedure, verification checklist.
- SKILL.md: −69% tokens (≈8,800 → ≈2,800). Details moved to
references/, read only when needed. - Shorter description in the agent's context: 968 → 618 characters.
- Measured on a full agent run, page to report: 119 → 86 s, script errors 1 → 0.
- Every allowed value and limit is written down; JSON files are written with the file tool, not a shell heredoc.
- Fix: Bulgarian and Belarusian pages pass the questions step.
- The review control platform is no longer named in the instructions:
read.mjswritesreviews.template.json. - Vectorscope 2.112 inside.
ai-search-audit 2.109
- Checks one page against published rules: access, robots.txt for search and AI bots, the Google and Bing checklist, structured data.
- Adds AI readings: the questions people ask, the brand, the reviews.
- Every problem is a label, a severity and a place on the page map.
- How generated the text sounds.
- Report in English, Ukrainian and Polish.
- The scripts call no model.