geo-score

# geo-score

> An open, versioned rubric for Generative Engine Optimization — score any site 0–100
> on whether AI answer engines can find, parse, trust and cite it. MIT licensed,
> machine-readable, with a zero-dependency Python CLI and a public benchmark of 317
> well-known sites.

The rubric it publishes is the AIV score: 21 tiered checks totalling 100 points, plus
4 bonus checks worth up to +6 outside the denominator. Every tier states a count out of
the 8 sampled pages, so two people scoring the same site agree on the arithmetic.
Three checks are gates — score zero on crawler access, live reachability or
server-rendered content and the result caps at 40.

## The benchmark

- [The state of AI visibility](https://jianruntech.github.io/geo-score/): 317 well-known sites scored. Median 56, range 11–99. 26% of them are unreadable to AI crawlers — 23 block AI crawlers by name, 47 serve a page whose body only exists after JavaScript runs, 13 hand a crawler an outright error.
- [简体中文版](https://jianruntech.github.io/geo-score/zh.html): the same data and findings in Chinese.
- [Raw data](https://github.com/jianruntech/geo-score/blob/main/benchmark/results.json): every site, every score, machine-readable.

## The rubric

- [AIV rubric v1.1](https://github.com/jianruntech/geo-score/blob/main/rubric/v1.1.md): the full specification, tier by tier.
- [简体中文](https://github.com/jianruntech/geo-score/blob/main/rubric/v1.1.zh-CN.md): the same specification in Chinese.
- [Machine-readable rubric](https://github.com/jianruntech/geo-score/blob/main/rubric/v1.1.json): stable check ids, tier conditions, bands.
- [Report schema](https://github.com/jianruntech/geo-score/blob/main/schema/report.v2.json): emit this and results from different implementations are comparable.
- [Calibration record](https://github.com/jianruntech/geo-score/blob/main/rubric/calibration-v1.1.md): how the thresholds were set, against four public benchmarks.

## The tool

- [CLI](https://github.com/jianruntech/geo-score/tree/main/cli): level 1 is one file, standard library only, Python 3.8+. `curl -sL .../cli/geo_score.py | python3 - yoursite.com`. Levels 2 and 3 (`--ask`, `watch`, and an MCP server) check whether AI engines actually cite a site, through their APIs with your own keys, and never add that to the score.
- [MCP server](https://github.com/jianruntech/geo-score/blob/main/guide/mcp.md): `geo-score-mcp` (stdio) gives an agent the same three levels. `score_site` is free and needs no key; `ask` and `run` spend the user's own API credits, and `--read-only` removes them. Setup for Claude Code, Claude Desktop, Cursor, VS Code, Codex, Gemini CLI and Zed is in the guide.
- [GitHub Action](https://github.com/jianruntech/geo-score/blob/main/action.yml): score on every push, fail the build on regression.
- [Five hand-scored audits](https://github.com/jianruntech/geo-score/tree/main/examples/audits/v1.1): all 21 checks scored by hand with reproducible evidence.

## Scope

This measures readiness — whether an engine *can* cite a site. Whether one *does*
depends on competition and query intent, which no site-side audit can observe.
Remediation is deliberately out of scope: the rubric names the gap and what the next
tier requires; it does not ship fix templates.

Maintained by [Jianrun Tech](https://www.jianruntech.com) (见润科技), Shenzhen. MIT.