English | 简体中文

Self-hosted web and AI crawler analytics for Generative Engine Optimization (GEO).

Jointly published by 猎看科技 · 北京日报社智源GEO

Quick start · Deployment · AI crawler analytics · Architecture · Documentation · Contributing

GEOInflow brings website visitors, AI search referrals, server crawl logs, and content inventories into a shared website context. It helps teams understand which AI crawlers visit their sites, which pages they request, which important pages have no observed crawl activity, where requests fail, and how AI search contributes visitors and conversions.

The system includes browser tracking, account permissions, GEO reports, crawler evidence, log collection, and analytics acceleration. The current product version is v1.2, with build version 1.2.0. Application source code is distributed under the MIT license.

Quick start and deployment guides are available in English and Simplified Chinese. Other linked reference guides are currently in Simplified Chinese.

GEOInflow is designed for company websites, content sites, media publishers, developer documentation, and GEO teams managing multiple sites. Operators can work at the website and page level, while maintainers can inspect ingestion, queues, source quality, and calculation methods.

The built-in catalog covers OpenAI, Anthropic, Perplexity, Google, Microsoft, ByteDance, Baidu, Kimi, Tencent Hunyuan, and other providers. It also retains observations of traditional search crawlers and other bots.

The catalog records platforms, products, and request identities separately. Training, AI search, and user-triggered access describe request purposes; AI workbenches and personal agents describe product types. These dimensions can overlap. Active entries depend on the administrator-published catalog and each website's configuration.

Declared User-Agent strings, matches against official IP ranges, and available DNS or trusted edge evidence are stored independently. A User-Agent claim alone is insufficient to confirm a source. Historical records without verification evidence retain an unknown status.

Review requested-page rankings, observed pages, repeated requests, directory distribution, leading crawlers, recent activity, and content freshness. Import a content inventory to measure observed crawl coverage and AI blind spots against a known set of pages.

Identity, purpose, and product-type filters apply before aggregation and pagination. Cards, tables, comparisons, and exports use the same conditions. Path observations and inferred results retain their estimate labels.

HTTP responses, caching, latency, response bytes, collection continuity, and risk findings are shown separately. Teams can record handling status, notes, deferrals, and resumptions, then review the relevant time windows and evidence.

Access policies support robots / Nginx configuration exports. Supported Cloudflare controls require separate Zone credentials, a preview, explicit confirmation, background execution, and a read-back of the remote state. Provider plans, permissions, and available fields must be checked in the deployed environment.

Browser tracking recognizes supported AI search referrals and combines them with landing pages, events, goals, and conversions. Crawler requests, referral sessions, and conversions are measured separately. Cross-source associations retain their comparability limits.

Crawl logs establish observed access. Indexing, model training, answer citations, recommendation rankings, and customer acquisition each require their own evidence.

See the AI crawler analytics guide for metrics, workflows, and limitations.

The application supports Simplified Chinese and English. The main date ranges are Today / 7 days / 30 days, and the default analytics timezone is Asia/Shanghai. The first full sign-in confirms the country and city timezone; later requests use the account's saved timezone to define today. Realtime visitors cover the most recent 30 minutes, and retention uses calendar months.

Browser metrics share a common traffic-quality policy. Empty data, missing integrations, unknown values, estimates, and unsupported capabilities are shown as distinct states. Demo websites are explicitly labeled. See page coverage for the feature matrix and date ranges and timezones for date definitions.

flowchart LR
  A[Website browser script] --> C[Application and ingestion APIs]
  B[Server / CDN logs] --> C
  D[sitemap / CSV / JSON inventory] --> C
  C --> E[(PostgreSQL)]
  E --> F[Log and analytics workers]
  F --> E
  E --> G[Reports and published results]
  H[(Redis cache)] --- G
  G --> I[GEO / visitor / content analytics]
  G --> J[Permission-controlled sharing and exports]

The browser script collects pageviews, events, and enabled extended data. Server and CDN logs add crawler requests that do not run browser scripts. Content inventories provide a known set of pages. Each input can be connected independently, and the corresponding analysis states its limitations when a source is missing.

Log processing retains sources, classifications, and verification evidence. Analytics workers use pre-aggregation, exact visit facts, and durable results to reduce repeated calculations. Standard dashboards and heavier analysis have separate admission controls. Frequently used ranges are prewarmed, and published results remain available during background updates. See the architecture guide.

Start Docker Engine / Docker Desktop and Docker Compose V2, and ensure that image and dependency download sources are reachable. After obtaining the source code, run this command from the project root:

sh scripts/liekan-deploy.sh

The script generates a private .env and random secrets, builds the application image, runs migrations, and starts the application, log worker, analytics worker, PostgreSQL, and Redis. The default address is:

http://127.0.0.1:3002

For a fresh installation, the initial username is admin and the initial password is liekan. Change the password after signing in, then create a website using the ingestion guide. Upgrades preserve existing customized passwords and accounts.

To try the bundled example after the application starts, run:

docker compose -f docker-compose.liekan.yml exec app node scripts/liekan-seed-demo.mjs

The example is named “猎看官网（示例数据）”. It initializes from the processed dataset included with the source. On first import, its timestamps are shifted to make the data convenient to explore. Repeated runs preserve existing data and deletion state. See the default demo guide.

Production access requires a domain, HTTPS, and a trusted reverse proxy. Browser PV collection can continue without a trusted IP ingress, while identity metrics such as UV require reliable ingress evidence. Automatic log ingestion and optional AI analysis have separate configuration requirements.

The deployment guide covers services, configuration, proxies, backups, and upgrades. The quick start explains how to connect the first website.

The main Today dashboards read published results every 30 seconds and show the actual data cutoff. Reads pause while the page is hidden or offline. Realtime visitors use their own refresh interval, while heavier analyses such as heatmaps and retention use longer background windows.

Completed results use browser, process, Redis, and PostgreSQL persistence. Common date ranges and pages share query conditions. Capacity safeguards govern eviction of older windows, idle failed tasks pause automatic retries, and initial calculations can finish through durable background tasks. Updates retain statistics permissions and checks for source quality, dates, and rule versions.

Initial calculations, cache hits, and background updates have different costs. Actual performance depends on data volume, coverage construction, the database, and cache state. After deployment, the Analytics Acceleration page shows tasks and coverage. See analytics acceleration and caching.

• Core website analytics and crawler identification can run on your own infrastructure without a model API key.
• Browser tracking is cookieless by default. Optional recording, event properties, and business integrations use the data scope configured for each deployment.
• AI analysis is disabled by default. When enabled, users choose model connections and website materials. See the AI configuration guide for the data sent to models.
• GEO Weight is in Beta. It uses qualified AI content access from the most recent 30 complete calendar days, with a documented formula, evidence, and limitations. The weight guide defines the fixed policy and applicable scope.
• GEO extensions, supplemental journeys, complete exports, and analytics acceleration primarily target PostgreSQL. Retained ClickHouse, link, and pixel endpoints follow their existing protocols; support for each GEO extension must be checked separately.
• Automatic server collection requires public HTTPS, an ingestion secret, and signed installation assets. Cloudflare integration also requires OAuth, permissions, and the relevant provider capabilities.
• Search impressions, actual rankings, model-answer citations, training inclusion, and article completion rates require additional data. The detailed guides define the observations and scores this project provides.

Development uses Node.js 22.23.3, pnpm 12.3.4, and PostgreSQL. Prepare an independent database and private configuration using the deployment guide, then run the following commands. This example uses the separate development port 3302:

pnpm install --frozen-lockfile
pnpm build:db
pnpm db:migrate
pnpm liekan:migrate
pnpm liekan:queue-bootstrap
pnpm build:openapi
pnpm build:packages
pnpm build:tracker
pnpm build:recorder
pnpm liekan:build-worker
pnpm dev --port 3302

Start the log and analytics workers in separate terminals as described in CONTRIBUTING.md. Development workers and the web application use the same private environment configuration.

Product pages live in src/components/liekan, data contracts in src/lib/liekan/contracts, and upstream query and permission adapters in src/lib/liekan/upstream. New structures use incremental project migrations. Applied migrations remain immutable. The liekan technical identifiers, collection protocols, and existing workspace package names retain compatibility.

For issue reports, contributions, scoped verification, and upstream synchronization, see Contributing, the documentation index, and upstream maintenance.

GEOInflow application source code and project documentation use the MIT license, permitting use, modification, and redistribution, including commercial use, subject to its notice requirements. The full license text governs these permissions.

GEOInflow is jointly published by 猎看科技 and 北京日报社智源GEO. Original copyright and license text are retained with the source. See the upstream notes for the project's origin and compatibility relationship.

Dependencies, fonts, and geographic datasets retain their own licenses. See third-party notices. The default geographic dataset is DB-IP City Lite, with its attribution and CC BY 4.0 license retained. Supplied MaxMind or other geographic databases must comply with their respective data terms. Runtime configuration, real visitor data, database backups, private keys, and privately supplied geographic databases are outside the source release.