jianruntech/geo-score

English · 简体中文

Can AI engines cite your site? A free 0–100 score from one command. Do they? Check through their APIs with your own keys.

Score (free, no key) → Ask one question (your API keys) → Watch a question set every week (your API keys). Three levels, one tool · the citation half is never added to the score.

curl -sL https://raw.githubusercontent.com/jianruntech/geo-score/v1.8.2/cli/geo_score.py \
  | python3 - stripe.com --brief

_{Recorded with geo-score 1.1.0 on 2026-09-09. A run with 1.2.0 on 2026-09-27 read 72.}

_{One command. Every check, and what the next tier needs. The report says how long its run took.}

  AIV READINESS  https://stripe.com
──────────────────────────────────────────────────────────────────────────
  72 / 100   Solid
  11 points to Leading

  Reachable 13/15
   ◐ Crawlers allowed in robots.txt   ███████████░░░░░░░  3/5
   ✓ Reachable to retrieval agents    ██████████████████  5/5
   ✓ Main content server-rendered     ██████████████████  5/5
  Understandable 17/22
   ◐ Sitemap discoverable and fresh   █████████░░░░░░░░░  2/4
   ✓ llms.txt present and structured  ██████████████████  5/5
   ✓ Organization + WebSite schema    ██████████████████  6/6
   ✗ BreadcrumbList on nested pages   ░░░░░░░░░░░░░░░░░░  0/3
   ✓ Page-type schema (Product, FAQ…) ██████████████████  4/4
  Content Citability 21/35
   ✓ Self-contained answer passages   ██████████████████  9/9
   ◐ Headings match how people ask    ████████░░░░░░░░░░  3/7
   ◐ Freshness signal present         █████████░░░░░░░░░  3/6
   ◐ Statistics carry a source        ████████░░░░░░░░░░  3/7
   ◐ Named, verifiable authorship     █████████░░░░░░░░░  3/6
  Brand Credibility 9/10
   ⊘ Third-party listings             ··················   —
   ⊘ Independent mentions             ··················   —
   ✓ Knowledge-graph entity           ██████████████████  4/4
   ✓ sameAs links resolve             ██████████████████  3/3
   ◐ Video and multimodal presence    ████████████░░░░░░  2/3
  Answer Fit 2/4
   ◐ Content shaped for extraction    █████████░░░░░░░░░  2/4
   ⊘ Covers the questions people ask  ··················   —
   ⊘ Chinese engine readiness         ··················   —
  ✓ full · ◐ partial · ✗ zero · ⊘ not measured (left the denominator)

  Biggest gaps
   +3 next tier (+3 to full)  Freshness signal present          needs: most pages do, and dateModified agrees with the visible date
   +3 next tier (+3 to full)  Named, verifiable authorship      needs: and the name links to a verifiable identity page
   +2 next tier (+2 to full)  Crawlers allowed in robots.txt    needs: mainstream retrieval user-agents explicitly allowed

  Scored 62 / 86 observable · 4 checks left the denominator · rubric v1.1 · took 91.5 s
  Needs off-site search or judgement: p3.listings, p3.mentions, p4.question-coverage
  Not observable this run: p4.cn-engines (not a Chinese-language site)

  Full rubric and what each tier means:
  https://github.com/jianruntech/geo-score/blob/v1.5.0/rubric/v1.1.md
  Score moved, or a check shows —? https://github.com/jianruntech/geo-score/blob/v1.5.0/guide/troubleshooting.md

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 324 well-known sites score → · a quarter of them are unreadable to AI crawlers.

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

Contents: MCP server · Three levels · What it checks · Leaderboard · Ways to run it · Does AI cite you? · Scope · How accurate is it? · Install and stability · Repository layout · Contributing

geo-score-mcp is a local MCP server (stdio, Python standard library only). It gives an agent the same three levels as the CLI: score a site for free, ask the AI engines one question, and track a question set week over week.

The buttons and the lines below run geo-score 1.8.2 from PyPI with uvx, which needs only uv: no git and no build step. They add no keys of their own, but a server started from a shell that exports a provider key (Claude Code passes its environment on) can use it. For a server that cannot spend, add --read-only after mcp.

Claude Code

claude mcp add --scope user geo-score -- uvx geo-score@1.8.2 mcp

Standard config (no keys in the file) for any client that reads an mcpServers file:

{
  "mcpServers": {
    "geo-score": {
      "command": "uvx",
      "args": ["geo-score@1.8.2", "mcp"]
    }
  }
}

Claude Desktop does not read your shell's PATH: put the full path that which uvx prints in command.

Without access to PyPI, the release wheel: uvx --from https://github.com/jianruntech/geo-score/releases/download/v1.8.2/geo_score-1.8.2-py3-none-any.whl geo-score-mcp

From source (needs git): uvx --from git+https://github.com/jianruntech/geo-score@v1.8.2 geo-score-mcp

Resources: geo-score://rubric/v1.1.

Prompts: audit_site (url), check_citations (question, brand), weekly_watch, compare_runs (from, to), explain_check (check_id).

Bold arguments are required. Hints are the tools' MCP annotations: clients may use them to decide what to confirm, and they are hints, not guarantees. Only ask and run can spend money, and only with the provider keys you give the server.

Things to ask your agent:

Keys and config. score_site needs no key. ask and run read the provider keys (OPENAI_API_KEY, PERPLEXITY_API_KEY, GEMINI_API_KEY, ANTHROPIC_API_KEY, OPENROUTER_API_KEY) from the server's environment, which is not your terminal's: a desktop app does not see what you exported in a shell. The level 3 tools also need a geo-score-watch.json. Pass it as -c /absolute/path/geo-score-watch.json, because a GUI client does not start the server in your project. From 1.4.0, --read-only gives a server that cannot spend anything.

Setup for each client (Claude Code, Claude Desktop, Codex, Cursor, VS Code, Gemini CLI, Devin Desktop, Zed), the Claude Code plugin, timeouts, security and troubleshooting: guide/mcp.md.

Level 1 is the score. Levels 2 and 3 measure the outcome the rubric keeps out of the score on purpose (two scores, never one): they are reported next to the 100 and never summed into it. A site can score 90 and still lose every answer to a competitor with more third-party coverage, and the reverse also happens, which is why you want both.

Levels 2 and 3 need cli/geo_watch.py next to cli/geo_score.py. The package on PyPI has both: uvx geo-score@1.8.2 example.com runs it without installing, and pipx install geo-score==1.8.2 gives you a geo-score command. A clone, or both files downloaded, works too. The one-line curl | python3 above runs level 1.

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. Because engines retrieve passages, this rubric gives Content Citability the most weight. That is a design choice, not a measured effect: the evidence for how much passage shape matters is thin, and the evidence table says where.

Most of what the rubric checks 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.

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 by design: answer engines retrieve passages, not domains. Whether passage shape outweighs domain authority is not established: the heaviest single check, p2.answer-passages, rests on thin support (where the evidence is thinnest).

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.

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. In this repository's own benchmark, 85% of 324 sites reach Early and 6% reach Leading (where the cuts fall).

A quarter of them are unreadable to AI crawlers. 73 sites have a gate check at zero — an AI retrieval crawler cannot get the content, so it has nothing of theirs to quote. 20 block AI crawlers by name in robots.txt, which is an editorial choice and reported as such — amazon.com lands at 17 for exactly this reason. 41 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. At a further 12 the server refuses crawler user-agents (unverified probes from the auditor's address).

Median 55 (95% CI 51–58, stats.py). Range 11 to 96. 66 sites were scored on fewer than four pages; without them the median is 59.

The full table, by sector → · markdown · raw data · CSV · every site's full report · re-run it

Two more findings worth the click. Sites in the five China sectors score 22 points lower than everyone else (95% CI 16–30; median 36 against 58; earlier samples measured with 1.1.0 put the gap between 16 and 23 points). Grouped by the language a site serves instead, the 54 Chinese-language sites score 26 points lower than the 270 others (95% CI 22–32), and within matched sector families the gap runs from -8 to 27 points. Counted check by check in raw points, 69% of the gap sits in checks the CLI reads with a heuristic, so at most that share could be the tool misreading Chinese pages (the breakdown). The two largest per-check differences, question-shaped headings and self-contained answer passages, are heuristics the CLI also read lower than a human on the one Chinese site in the hand-audit comparison, so part of the gap may be the tool reading Chinese pages conservatively. From 1.5.0 the CLI reads each page in its own language and counts Chinese task, explanation and comparison headings; the leaderboard above was measured with 1.5.0, so its figures include that change. And statistics that carry a source and a named, verifiable byline are among the three largest gaps on more than half the sites.

Every number here is reproducible with the command at the top of this page (the benchmark was run with 1.5.0 on 2026-10-02) — and we measured how reproducible. On 2026-10-12, 1.8.0 scored the whole list three times from one place the same day (the third run re-reading the first run's pages): of the 317 sites scored every time, 99% landed within ±5 between the first two runs and 81% scored identically (ICC 0.99, SEM 1.9; REPRODUCIBILITY.md, derived by benchmark/retest.py). An older 1.1.x build, measured on 2026-09-09, put 96% within ±5 and 45% identical. Read one site's score as ±5 rather than as exact; medians are stable. Most of what still moves is the knowledge-graph lookup, which went unanswered on most sites from that vantage, and the gate checks of sites whose bot protection answers a crawler differently from one run to the next. 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/.

CLI — level 1 is one file with no dependencies, and every report records how long its run took (elapsed_s). Levels 2 and 3 add cli/geo_watch.py from the same release. With the package from PyPI, uvx geo-score@1.8.2 (or geo-score, after pipx install geo-score==1.8.2) takes the place of python3 cli/geo_score.py below, all three levels included.

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 --badge-json aiv-badge.json  # the same badge as shields.io endpoint JSON
python3 cli/geo_score.py example.com --share                    # one line to paste somewhere
python3 cli/geo_score.py diff before.json after.json             # what changed between two reports, check by check
python3 cli/geo_score.py example.com --baseline before.json --fail-on-drop   # CI: re-score the same pages, fail only on a drop

GitHub Action (level 1) — score on every push, fail the build when it regresses. For level 3 on a schedule, see examples/ci/watch-weekly.yml.

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

A gate at zero caps the score at 40: a site that would otherwise score 40 or more reads exactly 40 and passes fail-under: 40. fail-on-gate: true fails any capped site.

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

npx skills add jianruntech/geo-score     # add -g for a global install
# then: /geo-score audit https://example.com

npx skills add runs the open skills CLI. It finds one skill in this repository, geo-score, and copies the whole repository (about 9 MB) as it is on the default branch, main, into .claude/skills/geo-score for Claude Code. Without Node.js, or for the release rather than main, clone the release tag into your own skills folder instead: git clone --branch v1.8.2 --depth 1 https://github.com/jianruntech/geo-score ~/.claude/skills/geo-score.

The CLI leaves those three checks out of the denominator rather than guessing (and a fourth, p4.cn-engines, unless the site serves Chinese). Re-scoring the five published hand audits on the same pages, it matched the auditor's tier on 80% of 84 check pairs (92% where it reads a rule, 69% where it approximates a judgement). On the checks both scored it read 2.4 points lower on average (per site from 12 lower to 10 higher; mean absolute difference 8.8). Its full normalised score read 5.2 points lower on average, from 16 lower to 7 higher, because the hand audits also score p3.listings, p3.mentions and p4.question-coverage, which the CLI leaves out. n=5, all of them calibration sites: VALIDITY.md lists every disagreement.

MCP server (all three levels) — uvx geo-score@1.8.2 mcp from PyPI (geo-score-mcp after a pipx install), or python3 cli/geo_score.py mcp from a clone, for Claude Code, Cursor and other agents: see MCP server.

--ask and watch put the questions your buyers ask to ChatGPT, Perplexity, Gemini and Claude, through each provider's search-enabled API with your own keys, and record who the answers cite: you, your competitors, or the third-party pages (forums, review sites) the engines lean on instead. Keys are read from the environment; geo-score never writes them anywhere.

git clone --branch v1.8.2 --depth 1 https://github.com/jianruntech/geo-score && cd geo-score/cli
export OPENAI_API_KEY=… PERPLEXITY_API_KEY=…         # any subset of engines works
python3 geo_score.py acme.com --ask "best invoicing app for freelancers"   # level 2
python3 geo_score.py watch init --brand Acme --domain acme.com --competitor "Rival=rival.com"
# put the questions your buyers ask an AI assistant into queries.csv, then:
python3 geo_score.py watch run --dry-run            # the plan, the caps, what a diff could detect and how precise one question's rate is; no calls, no cost
python3 geo_score.py watch run                      # level 3: every question on every engine, saved
python3 geo_score.py watch diff                     # this run against the last, with a significance test
python3 geo_score.py watch report --since 30        # the last 30 days pooled, one interval over questions

geo-score watch · Lumo · run 20260926T083000Z · api channel

6 questions × 2 engines × 2 = 24 planned · 23 answered · 1 failed · 0 skipped

Cited in 38% of answers to questions that do not name Lumo (6 of 16, 4 questions, 95% CI 12–62%) · mentioned in 38%
Questions that name Lumo (2, kept out of the headline): cited in 100% (7 of 7, 95% CI 65–100%) · mentioned in 100%
Your site was cited somewhere for 5 of 6 questions.

By engine
  engine          model             cited         95% CI   mentioned  named first (tracked)  avg rank
  chatgpt-api     gpt-6-luna        9/12 75%      42–100%  75%        9/12                   1.0
  perplexity-api  sonar             4/11 36%      8–75%    36%        0/11                   2.0       1 failed
  gemini-api      gemini-3.8-flash  not measured                                                       no_key: set GEMINI_API_KEY or GOOGLE_API_KEY

These rows, and the tables below, count every question, the 2 that name Lumo included.

Answer share your pages support: not measured; this run was saved before citation positions were recorded.

Sources listed and used in the answer: not measured; this run was saved before listed sources were recorded.

Repeat consistency of cited domains (per engine; descriptive, not read by diff)
  engine          questions measured  pairs  median Jaccard
  chatgpt-api     6/6                 6      1.00
  perplexity-api  5/6                 5      1.00
  gemini-api      not measured                               no_key

Least consistent questions (up to 3 per engine):
  id   question                         engine          answers that cite  Jaccard
  q02  ai video tool with no watermark  chatgpt-api     2                  0.50
  q01  best free ai video generator     perplexity-api  2                  0.50

- Repeat consistency: for each question an engine answered more than once in this run, the overlap of the domains two of its answers cite, shared domains over all domains of the two (Jaccard), for every two answers, averaged per question; the median over questions per engine. 1: every ask cited the same domains; 0: no domain in common. Domains are read as everywhere in this report, www dropped; the order of citations is not compared.
- Answers that cite nothing are left out: two empty lists are not agreement, and one empty list is not a 0. On an engine, a question with fewer than two answers that cite is not measured. Lowest: up to 3 questions per engine whose answers did not all cite the same domains, lowest first.
- Descriptive only: no interval and no test, never part of the cited rate, and watch diff does not read it. It describes asks minutes apart in one run, which share that day's index; it is not computed over pooled runs.

Share of voice
              cited  95% CI  mentioned  named first (tracked)  sentence share
  Lumo (you)  57%    26–86%  57%        9/23                   100%
  Pixa        52%    33–71%  100%       14/23                  100%

- Named first (tracked): answers that name that brand before any other brand in your config, out of the answers that name at least one of them. Brands your config does not list are not seen, so first among the brands you track may not be first in the answer. Two brands first named at the same place, one name inside the other, are neither first.
- Sentence share: in the answers that name a brand, the mean share of their sentences that name it, every sentence counted the same. A sentence ends at a line break, at 。！？, or at . ! ? before a space.
- Both read the answer as saved, which keeps its first 20,000 characters; a name that first appears after that is not seen.

Sources the engines cite most (not yours)
  domain        answers  share  owner
  pixa.example  12       52%    Pixa
  reddit.com    11       48%

Questions where a competitor is cited and you are not (1)
  id   question                    cited instead
  q03  cheapest text to video app  Pixa

Who is named for each question
  id   question                         answers  Lumo (you)             Pixa               no tracked brand named  searched yes/no/unknown  most cited
  q01  best free ai video generator     4        cited 3/4 · named 3/4  cited 2 · named 4  0/4                     2/0/2                    Lumo (you)
  q02  ai video tool with no watermark  4        cited 1/4 · named 1/4  cited 2 · named 4  0/4                     2/0/2                    Pixa
  q03  cheapest text to video app       4        cited 0/4 · named 0/4  cited 2 · named 4  0/4                     2/0/2                    Pixa
  q04  lumo vs pixa                     4        cited 4/4 · named 4/4  cited 2 · named 4  0/4                     2/0/2                    Lumo (you)
  q05  pixa alternatives                4        cited 2/4 · named 2/4  cited 2 · named 4  0/4                     2/0/2                    —
  q06  鹿末视频免费吗                   3        cited 3/3 · named 3/3  cited 2 · named 3  0/3                     2/0/1                    —

- An answer given without a web search cannot cite anything.
- "No tracked brand named" counts answers that name none of the brands in your config; they may name others.
- Perplexity and OpenRouter search on every call or do not say, so their answers count as unknown under searched.

By question type
  type         cited     95% CI   mentioned
  alternative  2/4 50%   15–85%   50%
  list         4/8 50%   22–78%   50%
  pricing      3/7 43%   0–100%   43%
  vs           4/4 100%  51–100%  100%

What the engines searched for (from 12 answers that show it)
  times  search
  2      best free ai video generator
  2      ai video tool with no watermark
  2      cheapest text to video app
  2      lumo vs pixa
  2      pixa alternatives
  2      鹿末视频免费吗

- Of 12 searches: 0 contain a year, 6 contain a ranking or comparison word.

Ledger
24 questions asked · 23 API requests · 14,620 in / 4,860 out tokens · 23 searches · $0.07 + 12 answers with no price (add prices to geo-score-watch.json)
1 failed after 24 HTTP requests in all; a request that timed out may still be billed.
Caps: at most 100 questions.

Read this before quoting the numbers
- API channel. Answers come from each provider's search-enabled API, which is not the consumer app. Compare runs with runs; never pool them with answers sampled by hand in the apps.
- The same question gets different answers from one ask to the next, and answers to one question move together. Rates carry a 95% interval: Wilson when each question was answered once, a bootstrap over questions when a question has several answers. diff pairs the questions both runs answered and calls a change a change only when an exact paired test (McNemar, or a sign-flip test when a question has several answers), Holm-corrected across engines, says so (p < 0.05).
- A question that names the brand invites an answer that cites it. Those questions are kept out of the headline rate and reported on a line of their own.
- Engines without a key, calls that failed and calls skipped by a cap count as not measured, never as zero.
- This measures citations. It does not predict traffic, rankings or revenue.
- This run was saved before questions were flagged at run time. Questions that name the brand were found when this report was made, from the saved question text and the run's brand name, aliases and domains.

Level 2 asks each engine each question once and prints the result under the readiness report (and into the report's citation object with --json). One ask is an anecdote: use it to see what the engines say today, not to measure a rate.

Level 3 keeps a fixed question list, saves every answer under .geo-score/watch/runs/ (schema), and reports:

What it will not tell you.

• It is not the ChatGPT app. Answers come from each provider's API with web search switched on. The consumer apps use other models, prompts and personalisation. Compare API runs with API runs; never pool them with answers sampled by hand in the apps.
• Some surfaces are out of reach. Google AI Overviews and AI Mode, the AI summaries inside apps such as TikTok, Reddit or Xiaohongshu, an assistant's memory and account settings, location finer than the country hint, and which library a coding agent picks cannot be reached through these APIs.
• A citation is presence, not merit. A citation rate says an engine named and linked you, not that what it said is true or that the page deserved it; engines also cite misleading and self-promotional pages.
• One answer is an anecdote. Every rate carries a 95% interval (a bootstrap over questions when a question has several answers, since those answers move together), and diff calls something a change only when an exact paired test on the questions both runs answered, Holm-corrected across engines, says so (p < 0.05). Too few shared questions (under 6, or too few for the number of engines compared) is reported as too few to tell. The method: guide/watch-methodology.md.
• Questions that name you are kept apart. "acme vs rival" or "is acme worth it" put your name in the engine's search and are cited almost every time. watch flags them when it runs, keeps them out of the headline rate, and reports them on their own line with their own count. An optional branded column in queries.csv corrects the flag by hand.
• Not measured is not zero. An engine without a key, a failed call and a capped call are all reported as not measured, and none of them lowers your rate.
• Citations are not traffic. Nothing here predicts visits, rankings or revenue.

Engines. Pin the model you mean in the config and keep it fixed between runs; diff flags a run where the model changed. OpenRouter covers hundreds of models with one key.

Caps and cost. max_calls is an exact cap on questions asked in a run, and a plan that exceeds it refuses to start. budget_usd is checked before every call against the spend so far plus the most expensive call seen on that engine, so it can be exceeded by at most one call per engine; with a budget set, engines whose cost cannot be known are left out unless you pass --allow-unpriced. Every run keeps a ledger of requests, tokens, searches and cost. Configuration, prices and all commands: cli/README.md.

From an agent. Levels 2 and 3 are also MCP tools, and an agent can only tighten your caps, never loosen them: see MCP server.

Every week. Citation rates move slowly and noisily: run on the same weekday with the same questions and models, and read diff, not single runs. examples/ci/watch-weekly.yml does it on a schedule with keys from repository secrets and commits each run, so the history lives in git. Run files contain your questions and the full answers: use a private repository if they are confidential.

Run it yourself, or have it run for you. Everything here is MIT; the tool has no paid edition. You pay your model providers directly and the ledger shows what each run used. What needs people rather than an API, Jianrun does as a service:

Five questions people ask, and where this repository answers each one:

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.

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. Measure it with levels 2 and 3 above. One reason it stays out: asked again, the same question gets a largely different set of sources. Repeated runs of a prompt within 24 hours, counting runs that cited something, had a mean source overlap (Jaccard) of 0.32 to 0.43 in each of four industries in one market with four engines pooled, and 0.23 to 0.51 engine by engine (Schulte et al. 2026, §5, Tables 4 and 6), and two runs of Google AI Overviews on the same device and location 0.66 (Grossman et al., SIGIR 2026, §4.4, Table 4).
• 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.

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

• Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024 — in a simulated engine, rewriting one source at a time to cite sources, add statistics or quote experts raised that source's share of the answer by up to about 40% (Position-Adjusted Word Count, not citation count). The paper also found an authoritative tone produced no significant improvement, which is why this rubric scores structure and attribution, not voice. The gain does not carry over cleanly: when every source was rewritten at once, the source that had ranked first lost 20 to 30% of its visibility (§5.2, Table 2), and a later benchmark that measures citation rank found most published methods gave no significant gain (C-SEO Bench, arXiv 2506.11097 §6.2). The details are in the evidence table, row p2.sourced-stats (rubric/evidence-v1.1.md).
• llms.txt proposal, Answer.AI — the convention this rubric checks for in the Understandable pillar (p1.llms-txt).

Every check declares what its weight rests on (published research, a vendor's own documentation, our field observation, or a convention no engine has confirmed using), with sources and the date they were last verified: rubric/evidence-v1.1.md. Where the evidence is thin (p2.answer-passages, p2.question-intent, p1.llms-txt), the table says so. If you have evidence that a weight is wrong, open a rubric proposal — that is the main thing we want contributions on.

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.

• Install: nothing to install for level 1 (the curl line above, pinned to a release). For a geo-score command with all three levels, install the package from PyPI: pipx install geo-score==1.8.2. To run it without installing: uvx geo-score@1.8.2 example.com (uvx geo-score@1.8.2 mcp is the MCP server). Neither needs git or a build step. Without access to PyPI, install the same wheel from the release: pipx install https://github.com/jianruntech/geo-score/releases/download/v1.8.2/geo_score-1.8.2-py3-none-any.whl. From source (needs git): pipx install git+https://github.com/jianruntech/geo-score@v1.8.2. The Claude Code skill: npx skills add jianruntech/geo-score (-g for a global install), or a clone into ~/.claude/skills/geo-score (Ways to run it). Standard library only, whichever you pick. Release downloads carry a SHA256SUMS file, and the wheel and sdist on PyPI match it byte for byte (how to check).
• Stability: the rubric and the tool are versioned separately, and the report schema, check ids, CLI flags, exit codes, Action inputs and MCP tools are public contracts. What may change in which release: the stability policy.
• No telemetry. geo-score sends nothing to us or anyone else. Level 1 fetches the site you name (following its redirects), the URLs its markup and robots.txt point to (sameAs profiles, logo, sitemap), and Wikidata and Wikipedia search; levels 2 and 3 call only the AI providers whose keys you set.
• Security: text quoted from the audited site is fenced as data in every report, the MCP server's score_site connects only to public addresses, and keys never reach a file. Threat model and reporting: security policy.

Why did my score move 4 points? Each run samples up to 8 pages, and the sample changes; in a same-day retest with 1.8.0, 99% of sites stayed within ±5 (REPRODUCIBILITY.md). To compare before and after a change, re-score the same pages with --baseline last.json, which also prints what changed, check by check (--urls-from last.json re-scores them without the comparison). Why does a check show —? It was not measured, so it left the denominator instead of scoring 0. More answers: troubleshooting. Every term the README, the reports and the guides use is defined in the glossary; every flag, exit code and watch command is in the CLI reference (简体中文).

Built and maintained by Jianrun Tech (见润科技), Shenzhen. We run GEO and AI-adoption programs for companies in China and abroad: B2B exporters, software companies, professional services, local businesses, education providers, and companies going global, cross-border e-commerce included. 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.

The most valuable contribution is evidence about the weights. See the contributing guide. Which issue form takes a wrong score, a disputed leaderboard row or a tool that does not work: getting help.

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

MIT