GeoLook

GEO = Generative Engine Optimization: getting AI engines (ChatGPT, Perplexity, Gemini, DeepSeek, Doubao…) to proactively mention and cite your brand when answering user questions. Not geographic info, not classic SEO.

More and more users ask AI directly — "best tools for X", "X vs Y, which one". If your brand:

Four stages plus operations, all in one self-hosted dashboard:

Status — Engine performance across 17 engines (10 automated via API + 7 manual, incl. Google AI Overviews and Metaso): mention rate, rank, citation share, what each engine actually cites, sample replay of raw answers, suspected-negative flags; brand mention distribution (you vs. competitors, per engine and aggregated); competitor tables with each rival's strongest engine one click away — plus each rival's citation-source mix and a "channels they have that you don't" hit-list; a 7-category question bank with intent-group cards (buyer / educate / probe — see at a glance which intent class you're absent from) where every question gets a diagnosis type (suspected-negative > competitor-dominated > absent > low-ranked); a sample library where every AI answer's metadata is browsable and human-correctable (regex parsing misreads — name collisions, negation — get fixed here; corrections recompute metrics instantly and survive re-sampling), with per-sample citation-source breakdown (domain × count · share) and sampling-environment provenance (sandbox / incognito / dedicated profile / personal — personal auto-downgrades to "needs review").

Keyword mining — expand the question bank from real search demand: Baidu suggest (CN) + Google autocomplete (Global) terms from brand/competitor/category roots (free public endpoints, no keys). Each round is snapshot-diffed to flag rising demand (affects topic ordering, never metrics); alternative/vs phrasings mined from competitor roots feed the Competitors page. Candidates only — adding to the bank is always a manual check.

Diagnosis — Site audit organized as a four-layer dependency chain (Access → Orientation → Understanding → Quotability: each layer depends on the one above, and a failing Access layer makes everything downstream invisible to engines — the fix order is computed for you). The Access layer goes well beyond robots checks: RFC 9309-compliant robots parsing (catches wildcard-group blocks, shared UA groups and specificity overrides that line-by-line regexes miss), WAF/CDN differential probing with real AI-crawler UAs (robots may allow GPTBot while your CDN 403s it — invisible from a browser), X-Robots-Tag header noindex, llms.txt link validation, hreflang coverage, sitemap index-pollution and duplicate title/content detection. Passage-level quotability: retrieval picks passages, not pages — pages with sections but zero independently quotable passages get flagged with a concrete fix. Plus gap diagnosis (content → channels → facts); a channel map of 19 channels weighted by real citation-corpus data; and a brand facts library as the single source of truth that llms.txt, JSON-LD and content drafts are generated from.

Action — Structured tickets (rationale / owner / effort / window / acceptance criteria) carrying an independent risk grade (low-risk quick wins / observe with 7-14-28-day recheck / high-risk technical changes with backup & rollback discipline — priority says how important, risk says how careful), with "first-measured → current → target" progress bars and automatic reopening on regressions; tickets deep-link to the exact question they most need written; a content workbench (topic pool sorted by "not mentioned + no content", required extraction blocks and brand facts at hand, live citability pre-check, fabrication-risk lint for AI drafts, and a distribution checklist matching each piece to its target channels); deploy assets (llms.txt, JSON-LD, HTML snippets, plus an AI-traffic attribution pack: GA4 "AI engines" channel-group regex, server-log counting script, source-snapshot guidance — closing the loop from "cited" to "converts"); publishing with channels grouped General / China / Global — GitHub, WordPress drafts, WeChat OA drafts, webhook, plus real X (teaser tweet with auto-backlink to your latest long-form publish) and Reddit (full-markdown self-post) integrations, each with an in-dialog step-by-step credential guide; multi-channel checkbox publishing per article, and publish status synced back to the action plan, pending list and question bank. Platforms without a personally usable official API (Weibo, Xiaohongshu, Toutiao, Bilibili, LinkedIn, Facebook, Instagram) are deliberately not integrated — better absent than fake; the page says why and offers the webhook bridge instead. Publishing is always manually confirmed, article by article.

Results — Per-question before/after (all / CN / global tabs), task-level before/after, verification history; boss-ready one-pager, execution plan, and a complete client delivery package (HTML + CSV).

Operations — Scheduled full-cycle re-runs (every 7/14/30 days; register the dashboard as a macOS standing service with scripts/service.sh install — starts at login, restarts on crash, keeps running with every terminal closed, so scheduled re-runs actually fire), multi-brand with one-click switching, and a manual-sampling loop that feeds the same metrics: export a full sheet or a weekly buyer-intent sheet (sample-sheet --intent buyer --limit 20), or use the Chrome sampling assistant below.

Sampling assistant (Chrome extension, extension/) — for the engines without APIs, which is where the real users are. Side panel loads your buyer-intent queue (grouped by intent), fills the question into the chat box (you press Enter in the default manual mode), extracts the finished answer with all citation links in one click, and posts samples back to your local dashboard as A-grade evidence — a weekly check drops from ~30 minutes to ~10. An explicit opt-in auto-run mode exists with hard guardrails (you stay present, rate-limited, capped, halts on any CAPTCHA/anti-bot signal — ToS risks documented honestly in its README). Ships with sandbox.sh: one command launches a disposable clean-browser sandbox (no history, no cookies, extension auto-loaded) or a persistent logged-in profile for engines that require accounts, so sampling hygiene ("what a stranger sees, not what AI thinks of you") is one command instead of a discipline.

Most GEO products are monitoring SaaS: they show mention rates and rankings, charge monthly, and keep your data in their cloud. GeoLook is an implementation platform:

Honest limits: single-machine tool, no accounts or team collaboration; sampling frequency and volume depend on your own API budget; "suspected negative" flags are leads for human review, not verdicts. These are deliberate design choices.

• macOS or Linux (Windows via WSL — the code uses fcntl file locks)
• Python 3.9+
• Exactly three third-party packages: requests, beautifulsoup4, lxml

# 1. Clone and install
git clone https://github.com/aigclink/geolook.git
cd geolook
pip3 install requests beautifulsoup4 lxml

# 2. Start the dashboard (opens your browser)
python3 scripts/geo.py ui        # → http://127.0.0.1:8765
#    macOS: ./scripts/service.sh install registers it as a standing service
#    (starts at login, restarts on crash, survives closing every terminal)

# 3. (Optional) Configure engine API keys
#    A: in the dashboard — Settings → Engines & Keys → "Configure" (writes local .env)
#    B: cp .env.example .env and edit

Zero keys works too: automated sampling is skipped; use the manual sampling sheet loop instead. Crawling, auditing, tickets and assets need no keys. One CN-capable key (e.g. DeepSeek/GLM) unlocks auto-derivation of the question bank / brand facts and AI first drafts.

The server binds to 127.0.0.1 only (a deliberate security boundary — there is no auth layer). For remote access:

ssh -N -L 8765:127.0.0.1:8765 user@your-server   # then open http://127.0.0.1:8765 locally

For multi-user setups put a reverse proxy with auth in front. .env and work/ contain secrets and project data — mind file permissions.

git pull        # your data lives in work/ and .env, both gitignored

python3 scripts/geo.py new --url https://example.com --market both

--market is cn / global / both. Nine steps run automatically: crawl → audit → derive facts/competitors/questions → sample every engine → tickets → assets → report → auto-verify → delivery package, landing in work/<slug>/delivery/<date>/.

1. Onboard — python3 scripts/geo.py ui, follow the 3-step wizard (URL + market), hit "create & auto-bootstrap". The first cycle runs in the background.
2. Review the foundation (important, ~10 min) — facts are extracted only from your site copy; unknowns are marked "unconfirmed". Check Brand Facts (fix wording, add aliases — missing aliases understate mention rate) and the Question Bank (do questions read like real user queries?).
3. Status — Overview for the one-line verdict and health score; Engines to drill into each engine and replay raw answers (log factual errors on the spot); Competitors to see each rival's strongest engine — that's the channel to go build.
4. Diagnosis — Site audit (click a missing block → its fixing ticket) → Gap diagnosis → Channel map (open any channel for its build plan).
5. Execute — Take tickets P0-first in Action Plan (click a title for why/how/acceptance); write in the Workbench (pre-check ≥ B, publish as final, then work through the distribution checklist); deploy Assets (llms.txt to site root, JSON-LD into <head>, snippets into templates — see DEPLOY.md).
6. Verify — Settings → "Auto-verify" re-crawls and judges tickets; after the next sampling round, check per-question before/after in Verification.
7. Operate — enable scheduled re-runs (7/14/30 days); generate monthly reports and client packages in Reports & Delivery.

No site is fine — brand-owned sites account for only 1.37% of CN citations; AI visibility lives on external channels anyway:

python3 scripts/geo.py init --no-site --name "Product" --materials brief.md --market cn

Your brief replaces site copy as the basis for brand facts, competitors and the question bank (omit --materials and a template is generated for you). Only three things differ: crawl/audit are skipped, llms.txt and JSON-LD are not generated (they need your own domain), and own-site citation rate shows "n/a" instead of 0. Everything else runs as usual.

python3 scripts/geo.py sample-sheet  --slug <project>                      # full sheet
python3 scripts/geo.py sample-sheet  --slug <project> --intent buyer --limit 20   # weekly buyer-intent check
python3 scripts/geo.py sample-import --slug <project> --file <sheet>

Faster: install the Chrome sampling assistant (extension/README.md) and launch a clean sandbox with extension/sandbox.sh — queue, one-click extraction and upload, ~10 minutes per weekly round. Review imported samples in the dashboard's Samples page (machine parsing is correctable there; corrections recompute metrics instantly).

Every command has --help.

Q: AI answers differ every time — how can sampling results be stable?

A single AI answer is inherently stochastic, so GeoLook never reads a single answer as a metric. Stability comes from four layers:

1. Aggregation — mention rate and friends are ratios over dozens of questions × multiple engines; per-question jitter averages out.
2. Fixed variables — each engine's sampling model is pinned (visible and changeable in Settings), the question bank is fixed, and the same set is reused across rounds; the only thing that changes is time.
3. Repeat rounds — sample --repeat N samples each question multiple times when your budget allows.
4. Attribution discipline — the UI explicitly labels single-round movement as observational ("a drop isn't necessarily a regression — sampling is noisy"); only two consecutive rounds moving the same way count as a trend, and ticket verification relies on deterministic signals (re-crawling the site) rather than sampling alone.

Honestly put: it measures a distribution, not a fixed value — which is exactly what real users experience, since every real query to an AI engine draws a random sample too.

All six audit dimensions are anchored in public empirical data; scripts/audit.py implements references/method.md:

• High-impact pages average 1,943 words; low scorers just 170 (11.4×)
• Numbers +61.6%, definitions +57.3%, comparisons +55.3%, how-to +41.2% citation-probability lift
• Pure Q&A formatting is −5.7% — looking like an FAQ doesn't help
• Topical relevance is the strongest predictor (r = 0.432), above authority
• Brand-owned sites get only 1.37% of CN citations — your site is the fact source; external channels are the citation sources

• Single-machine, self-hosted: stdlib http.server on 127.0.0.1; no DB, no accounts; data is plain files
• Never fabricate: facts only from site copy; inventing competitor names is forbidden; AI drafts must pass lint + human review
• Verification is the product: anything auto-verifiable never relies on someone saying "done"
• Publishing is always manual: channel credentials in local .env (mode 600); every publish is an explicit click; WeChat/WordPress go to drafts only

This repo doubles as a Claude Code skill (SKILL.md): drop it into your skills directory and tell Claude "do GEO for example.com". Claude is optional — every script is a plain CLI.

scripts/          All logic (geo.py CLI · dashboard.py server · ui.html single-page UI · service.sh)
extension/        Chrome sampling assistant + sandbox.sh + its own e2e tests
references/       Methodology: sampling discipline, content patterns, attribution, citation structures
tests/            Unit tests
work/<slug>/      Per-project data (gitignored, never leaves your machine)
docs/             Screenshots and the 40-second demo video

• @yaojingang

Questions, ideas or collaboration — email bingqiang2008@gmail.com, or open an issue.

MIT