English · 简体中文

geo-score

Will ChatGPT cite your site? Score it in 20 seconds.

curl -sL https://raw.githubusercontent.com/jianruntech/geo-score/main/cli/geo_score.py \
  | python3 - stripe.com --brief

One command. About twenty seconds. Every check, and what the next tier needs.

Python 3.8+, standard library only, nothing to install. It reads public URLs and prints a score against a published, versioned rubric — not a black box.

See how 314 well-known sites score → · a quarter of them cannot be cited at all.

GEO means Generative Engine Optimization — getting cited by ChatGPT, Perplexity, Google AI Overviews, Gemini and Copilot. Nothing to do with geography or maps.

---

Why this is a different question from SEO

Classic SEO asks where do I rank. Answer engines don’t rank — they retrieve passages, decide whether a source is worth quoting, and cite it. Different question, different failure modes: a site can sit at position 3 on Google and never be quoted, while a page nobody links to gets cited daily because its passages are clean.

Most of what determines this is mechanical and cheap to fix — a robots.txt line, a JSON-LD block, a date in a template, a paragraph rewritten so it stands on its own. The hard part is knowing which of them you are missing, and what each one is worth.

What it checks

21 tiered checks totalling 100 points, plus 4 bonus checks worth up to +6 outside the denominator. Full specification: rubric/v1.1.md · 简体中文

Content Citability carries the most weight on purpose: answer engines retrieve passages, not domains. Passage shape beats domain authority more often than classic SEO intuition expects.

Every scored check is tiered — 2 to 4 tiers, each naming 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, because until a crawler can reach the content nothing else you change has any effect.

Bands

Band names describe a stage, not a verdict. External benchmarks put most business sites in the 30–55 range, so a score in the forties is ordinary, not alarming.

314 sites, scored in public

A quarter of them cannot be cited at all. 80 sites have a gate check at zero — a retrieval crawler simply cannot get the content. 19 block AI crawlers by name in robots.txt, which is an editorial choice and reported as such — amazon.com lands at 12 for exactly this reason. 45 serve a page whose body only exists after JavaScript runs. Their content is there, a browser sees it, and a crawler gets an empty shell. That group almost certainly did not choose it. A further 16 hand a crawler an outright error.

Median 56. Range 12 to 98.

The full table, by sector → · markdown · raw data · re-run it

Two more findings worth the click: sites built for the Chinese market score 19 points lower than everyone else (median 40 against 59 — a gap that has held between 16 and 23 points across five separate samples, from 38 sites up to 314), and the same three cheap things — a date in the page template, an opening paragraph that stands on its own, one JSON-LD block — are missing from more than half the field.

Every number here is reproducible with the command at the top of this page — and we measured how reproducible. Running the whole benchmark twice and comparing every site: 96% land within ±5, 45% land identically. Read one site’s score as ±5 rather than as exact; medians are stable. The unstable part is the gate checks, where five sites flipped between runs because their bot protection answered a crawler differently. The band, the control experiment and the per-site pairs are in benchmark/REPRODUCIBILITY.md.

For five reference sites we also publish hand-scored audits covering all 21 checks, with the evidence behind each one: examples/audits/v1.1/.

Three ways to run it

CLI — no install, no dependencies, 20 seconds.

python3 cli/geo_score.py example.com            # human-readable
python3 cli/geo_score.py example.com --explain  # with the evidence behind every check
python3 cli/geo_score.py example.com --json     # conforms to schema/report.v2.json
python3 cli/geo_score.py example.com --compare competitor.com   # side by side
python3 cli/geo_score.py example.com --badge aiv-badge.svg      # embeddable SVG
python3 cli/geo_score.py example.com --share                    # one line to paste somewhere

GitHub Action — score on every push, fail the build when it regresses.

- uses: jianruntech/geo-score@v1
  with:
    url: https://example.com
    fail-under: 40

Claude Code skill — the CLI measures what a static fetch can see. Four checks need off-site search or human judgement, and the skill does those too.

git clone https://github.com/jianruntech/geo-score ~/.claude/skills/geo-score
# then: /geo-score audit https://example.com

The CLI leaves those four checks out of the denominator rather than guessing, so it reads a little lower than a full audit — typically by 5 to 15 points on an established brand, which has listings and mentions the CLI cannot see.

Why a rubric, not just a tool

A score you cannot audit is a number someone made up. So the specification is the product, and the tools are implementations of it:

• Versioned. Every score reports the rubric version. 71 (v1.1) is a claim; 71 is not.
• Tiered, with counts. Each tier names a page count out of 8, not “most”.
• Evidence-bound. Every check requires an observation someone else can reproduce.
• Calibrated against public benchmarks, with the record published — including the four external sources the thresholds were checked against, and the eight specification ambiguities that real audits surfaced and v1.1 settled.
• Machine-readable. rubric/v1.1.json with stable check ids, and schema/report.v2.json so results from different implementations are comparable.

Implement it in your own stack, disagree with a weight, open a rubric proposal. That is the main thing we want contributions on.

Scope — what this does not do

This is the part most tools leave out, so it’s stated plainly.

AIV Score measures. It does not fix.

Other honest limits:

• It measures input-side readiness, not outcomes. A high readiness score means engines can cite you. Whether they do depends on competition, query intent and factors no external audit can observe. Citation performance is reported as a separate, unscored block and never folded into the 100 — see Two scores.
• Brand Credibility and the named-author check need human judgement. “Is this a real identifiable person” and “is this mention independent” are not fully automatable. Treat those ~24 points as assisted, not automatic.
• Tiers reduce disagreement, they do not remove it. Every tier names a count out of the 8 sampled pages, so two auditors agree on the arithmetic. They can still disagree on whether a given paragraph is a self-contained answer. The settled ambiguities are the ones we found; there will be more.
• Heavily client-rendered sites score low, sometimes unfairly. If your content only appears after hydration, most checks will read the pre-hydration HTML — which is also roughly what a crawler sees, so the low score is usually right, but verify by hand.
• Engine behaviour moves. The rubric is versioned for exactly this reason. A score from an older rubric version is not comparable to a current one.

Research behind the weights

The weights are opinionated but not invented. The two findings that most shaped them:

• Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024 — citing sources, adding statistics and quoting experts raise visibility by up to 40% (measured as Position-Adjusted Word Count, not citation count). Notably, the paper found an authoritative tone produced no significant improvement — which is why this rubric scores structure and attribution, not voice.
• llms.txt proposal, Answer.AI — the convention this rubric checks for in the Understandable pillar (p1.llms-txt).

Where a check rests on our own field observation rather than published research, the rubric says so. If you have evidence that a weight is wrong, open a rubric proposal — that is the main thing we want contributions on.

Related tools

Deliberately naming what this is not, so you can pick correctly:

If you need remediation and not just a score, those projects overlap with the part this repo deliberately excludes.

Who maintains this

Built and maintained by Jianrun Tech (见润科技), Shenzhen — we run GEO and AI-adoption programs for cross-border commerce companies. The rubric came out of client work and out of optimizing our own products; publishing it is how we’d like AI visibility to be measured consistently, including by people who never become our clients.

Commercial use of this repository is unrestricted under MIT — including inside paid consulting work. You do not need our permission, and there is no separate commercial licence.

Contributing

The most valuable contribution is evidence about the weights. See CONTRIBUTING.md.

Citation

If you reference the rubric in research or a report, see CITATION.cff.

License

MIT