Skip to content

geo-score 1.8.4

Latest

Choose a tag to compare

@alexxu0909 alexxu0909 released this 13 Nov 16:02
· 1 commit to main since this release

Added

  • Website: geo-score.dev, with docs and a free online check that runs the published base score (no --ask, no API keys). Both READMEs link to it; the package Homepage URL points to it from the next release, and the source link moves to Source. Documentation only.
  • Sponsors section in both READMEs (CCSub), with the rule that sponsorship never affects the rubric, scores, leaderboard or technical recommendations, and a note that sponsors are not tested for the citation levels. Documentation only.
  • A 60-second intro video at the top of both READMEs (English in README.md, Chinese in README.zh-CN.md): one command, the 0–100 score, the three gates, the gaps worth fixing first, and why every point can be re-checked. The demo run is jianruntech.com scored with 1.8.3; the soundtrack is synthesized for the video, no third-party music. Documentation only.
  • Rubric governance (rubric/governance.md + zh, linked from the READMEs, rubric/README.md and CONTRIBUTING):
    who can propose a change and what evidence it needs; the five steps from proposal to release, with at least 30 days
    of public comment for a minor version and 60 for a major one, and an independent reviewer; the four ways a
    published score can change; what keeps one version giving the same score whoever runs it; our conflict of interest
    (Jianrun Tech sells GEO services; the rubric, tool and leaderboard are free, with no certification or paid
    placement); and how the rubric relates to standards bodies (an open specification, no paid seat, drafting only on a
    formal invitation and after legal review). Documentation only: rubric v1.1, every check and every score are
    unchanged. validate.py rule 14 now holds the English page to English.

Changed

  • Both READMEs restructured, with screenshots. README.md and README.zh-CN.md now follow one outline, section for
    section: what problem it solves (with a diagram of the three levels), screenshots, core capabilities, what it checks,
    the leaderboard, how accurate it is, quick start, levels 2 and 3, who it is for, boundaries and governance, a
    developer section, and a documentation index in both languages. The six PNGs in assets/screens/ (four per README)
    are real outputs captured on 2026-10-17: a 1.8.3 run on jianruntech.com in a terminal, the GitHub Action's job
    summary for that report, and the public leaderboard's English and Chinese pages. Long reference material (the MCP
    tool table, what level 3 reports, the engines, caps and cost) is folded into collapsible blocks; the generated MCP
    blocks are untouched, and the old anchors that moved (#ways-to-run-it, #stability-privacy-and-install-options
    and their Chinese twins) still resolve. Documentation only: every figure, the rubric, every check and every score
    are unchanged.
  • reference/ai-crawlers.md: Anthropic's training-preference signal. Since 2026-10-30 Anthropic recognises a
    machine-readable "no AI training" signal (a draft from the IETF's aipref working group), and a robots.txt rule
    naming ClaudeBot is followed over it; the ClaudeBot row says so, with the article's own caveats (no syntax shown,
    rolling out over time, reading may change). The page also notes that Anthropic-hosted browsers (Claude-Agent) are
    outside the article it cites, and that g.robots reads a whole-site Disallow under section 2.2.2 (see Fixed).
    Documentation only: the CLI does not read the signal.
  • The watch example asks about portable power stations. The illustrative run in both READMEs, the two saved
    runs in cli/fixtures/watch/, the watch ask example in geo_watch.py and the texts in the watch tests now use
    made-up questions about portable power stations, with the same made-up brands (Lumo and Pixa; the third is now
    Trailcell). Every count, rate and interval in the example is unchanged. Documentation and tests only.

Fixed

  • g.robots reads a whole-site Disallow the way RFC 9309 does. A Disallow: / (or /*) used to close the site
    to a crawler whatever stood beside it, while the page-by-page reading in the same tool (outside the score) already
    let the longest matching rule decide, with an Allow winning a tie (RFC 9309, section 2.2.2). The two now agree: the
    Disallow counts only when it decides the root path. So Allow: / in the same group, a site's own Allow: / merged
    with a CDN-managed block that names the same crawler (section 2.2.1), or a whitelist that opens the homepage with
    Allow: /$ before ending in Disallow: /, no longer scores 0 and no longer caps the score at 40. The evidence says
    when a Disallow: / is outweighed this way, and a training crawler is no longer reported as disallowed when its own
    Allow outweighs the Disallow. Direction: up, for those files only. guide/chatgpt-access.md says the same. Reported from a reading of third-party robots.txt guidance and
    reproduced with three minimal files, now in the tests together with the whitelist case.
    On the 2026-10-02 leaderboard it moves one site: figma.com, whose robots.txt is such a whitelist, scores 51
    (Growing) instead of 40 on its stored report, so 72 sites have a gate check at zero, not 73, and 19 of them through
    robots.txt, not 20. The rows stay as measured; benchmark/README.md and both READMEs carry the correction, which
    benchmark/run.py writes from a list of confirmed tool errors (build_page.ERRATA) that the tests re-score from
    raw/.

Verify downloads: shasum -a 256 -c SHA256SUMS.
中文更新日志:https://github.com/jianruntech/geo-score/blob/v1.8.4/CHANGELOG.zh-CN.md