Aloom

Measure product visibility in official AI Web answers.

Fixed prompt suites · Repeatable browser collection · Provider-level reports

English · 简体中文

Aloom is a self-hosted GEO detection and recurring monitoring tool for Doubao, DeepSeek, Yuanbao, and Qwen. It uses fixed, versioned prompt suites, real official-Web collection, sample-level checkpoints, structured answer analysis, and comparable reports. Model APIs are never substituted for official-Web monitoring samples.

What Aloom measures

• Product mention, candidate, and recommendation rates.
• Absolute rank, sentiment, target share, and competitor share.
• Blind discovery versus aided brand recognition.
• Visible source exposure and extracted citations.
• Factuality against confirmed public reference pages.
• Stability across repeated samples.
• Collection completion and analysis success as separate metrics.

Every failed, blocked, or unattempted checkpoint remains in the report denominator. not_exposed means that the provider page did not expose extractable links; it does not mean that the answer had no underlying sources.

Scoring model

Aloom reports two explicit score levels:

• Answer Performance Score: the mean of each analysed answer's visibility, rank, sentiment, and recommendation scores, weighted 25% each. Missing or failed planned samples contribute zero to the series mean.
• Aloom GEO Score v1: a series-level score that combines six independent measurement layers and is calculated separately for the overall series and for each provider.

Series layer	Weight	What it measures
Visibility	25%	Measured brand appearance across all planned samples.
Evidence	20%	Visible source exposure and claim support.
Factuality	15%	Accuracy against the confirmed fact ledger.
Competition	20%	The target's observed share against extracted competitors.
Stability	10%	Agreement across repeated, mode-matched samples.
Governance	10%	Collection, analysis, and required-dimension completeness.

An unavailable layer is not silently scored as zero. Aloom normalizes the score across assessed layers, displays the assessed weight as scoring coverage, and marks the result provisional whenever coverage is below 100% or the formal series is incomplete. These are Aloom's independently versioned product weights.

Every provider report includes its own weighted score, layer breakdown, Answer Performance Score, completion, analysis rate, mention, recommendation, absolute rank, sentiment, visible-source exposure, target share, competitor share, mode cohorts, failures, prompts, raw answers, citations, and conversation evidence.

Detection workflow

1. Scan the official product website.
2. Confirm the brand, category, products, audiences, regions, and competitors.
3. Select a detection suite, language, providers, official-Web modes, and sampling depth.
4. Review every rendered prompt, total sample count, run count, and cadence, then click Start detection.
5. Aloom records the selected configuration and runs each prompt in a fresh official-Web conversation.
6. Inspect the report by provider, mode, locale, intent, stage, exposure, entity, or individual prompt.
7. Choose the total run count and a daily, weekly, or monthly cadence in New Detection. Every planned run stays in the same comparable series.

Formal detection does not accept arbitrary custom prompts. Historical custom and legacy prompts remain linked to their original samples but are excluded from new formal series and reports.

Aloom GEO Detection Pack v1.1

The built-in pack contains 54 fixed templates per language: nine intents across six decision stages.

Suite	Matrix	Core prompts per language
Quick Scan	All intents × Awareness and Evaluation	18
Discovery	Information, Recommendation, Scenario, Alternative × first three stages	12
Competitive Position	Recommendation, Comparison, Alternative × Screening, Evaluation, Purchase	9
Trust & Risk	Risk, Price, Brand Validation × Screening, Evaluation, Purchase	9
Buyer Journey	All intents × Awareness through Purchase	36
Full Matrix	All nine intents × all six stages	54

Advanced dimensions can narrow a suite by language, intent, decision stage, brand exposure, product, competitor, audience, or region. Entity assignment is deterministic, and the coverage planner adds only the minimum missing questions. It never creates a Cartesian product.

The run plan is independent from prompt coverage. Choose 1–30 total runs and, when more than one run is needed, a daily, weekly, or monthly cadence. The first run is queued immediately; later runs remain in the same series so the report can measure answer stability without splitting the result across unrelated detections.

Official-Web modes

Aloom records requested and verified modes separately. Search-enabled cohorts are never merged with non-search cohorts.

Provider	Supported text cohorts	Search rule
Doubao	Fast, Expert	Office-agent modes are excluded from GEO text baselines.
DeepSeek	Fast, Expert, DeepThink, Fast + Search	Search is only valid with Instant/Fast. DeepThink + Search is rejected.
Yuanbao	Default, Deep Thinking, Search, Deep Thinking + Search	Search must be explicitly selected from Tool > Search.
Qwen	Auto, Fast, Thinking, plus search-enabled variants	Search is selected through + > More > Web search; Aloom verifies the separate composer marker and never treats the general Tools switch as proof of search.

If a platform changes or a switch cannot be verified, the sample fails with a mode-specific error instead of being mislabeled.

Browser collection and privacy

The local collector runs task-bound headless Camoufox with one persistent profile per provider. Login, cookies, local storage, browser identity, and profiles remain on the user's computer.

• Every browser page belongs to a concrete collection task.
• All pages owned by a task are closed together when that task finishes.
• Every baseline prompt starts a fresh conversation.
• A confirmed submission is never sent again merely because extraction failed.
• Login expiry, QR confirmation, CAPTCHA, sliders, and security checks enter waiting_human.
• Human handling opens the same persistent profile in a visible window; Aloom does not bypass verification.
• Completed samples and checkpoints survive browser, collector, and challenge recovery.

AIHubMix is used only after collection to structure observed answers. A Full Matrix never places 54 answers into one model context: each answer receives one schema-constrained analysis call, and reports aggregate validated records deterministically.

Architecture

flowchart LR
  UI["Next.js control plane"] --> PG["PostgreSQL manifests and checkpoints"]
  UI --> CH["ClickHouse answers and analysis"]
  UI --> R["Redis and BullMQ"]
  R --> C["Local persistent Camoufox collector"]
  C --> D["Doubao Web"]
  C --> DS["DeepSeek Web"]
  C --> Y["Yuanbao Web"]
  C --> Q["Qwen Web"]
  C --> CH
  CH --> A["AIHubMix structured analysis"]

Loading

Local setup

Requirements: macOS or Windows, Node.js 20+, pnpm 10+, Docker Desktop, and Python 3.12.

git clone https://github.com/MYZ8088/aloom.git
cd aloom
cp .env.example .env
pnpm install
pnpm camoufox:setup
pnpm camoufox:doctor
pnpm local:background

Open http://localhost:3000, then connect each provider under Providers. The background runtime remains active after the launching terminal closes. Persistent profiles are stored under .aloom-storage/ and are excluded from Git.

For an installation created under an earlier project name, run pnpm brand:migrate-runtime once before pnpm local. The migration copies legacy Docker volumes and local browser state into Aloom names, verifies volume sizes, keeps the legacy volumes, and backs up incomplete target volumes.

Useful commands:

pnpm camoufox:doctor  # verify Python, Camoufox, browser, GeoIP, and executable
pnpm local            # run the Web app and collector in the foreground
pnpm local:background # start a terminal-independent local runtime
pnpm local:status     # check the runtime and Web endpoint
pnpm local:stop       # gracefully stop the background runtime
pnpm collector        # start only the local collector
pnpm test             # unit and fixture tests
pnpm typecheck        # workspace TypeScript checks
pnpm build            # language/privacy gates plus production builds
pnpm test:soak:same-day -- --workspace=<id> --user=<id> # scheduled acceptance smoke

Background runtime logs are written to .aloom-storage/logs/local-daemon.log. Interrupted BullMQ jobs use PostgreSQL checkpoints to skip terminal samples and resume only unfinished prompts.

Reliability rules

• A formal series cannot start until every expected prompt hash has a checkpoint source.
• Each provider has concurrency one. Selected providers run in parallel up to COLLECTOR_PROVIDER_CONCURRENCY (default 4), with an automatic reduction to 2 on low-resource machines.
• Structured analysis runs in a separate bounded background lane controlled by COLLECTOR_ANALYSIS_CONCURRENCY (default 2) and never blocks official-Web collection.
• Browser restarts reuse the same persistent identity and profile.
• Prompt echoes, login pages, challenge pages, provider errors, and empty answers are rejected.
• A submitted answer is re-read before any resend. If it remains structurally incomplete, Aloom may retry it once in a verified fresh conversation; extraction failures never trigger an unbounded or blind resend loop.
• Collection and analysis statuses are stored independently.
• Strict JSON Schema, balanced JSON extraction, local repair, Zod validation, a fallback model, and one targeted repair pass stabilize structured analysis.
• Trends compare only matching prompt hashes, providers, modes, languages, and sampling definitions.
• Offline collectors leave work in waiting_runner instead of dropping the series.
• Release acceptance uses a same-day two-round smoke: one immediate or seeded round and one timer-triggered round. The optional pnpm test:soak:72h profile is reserved for extended observation and does not block routine releases.

Project lineage and licenses

Aloom is an independent project, not an official OneGlanse or Yao GEO Skills release.

• Parts of the original codebase and self-hosted architecture are derived from OneGlanse, copyright 2025 Aryaman Todkar, under the MIT License. Aloom substantially reworks that foundation into a detection-only product with persistent local collectors, versioned run manifests, mode-aware sampling, and report-level checkpoints.
• Selected prompt-taxonomy, intent-mining, and Web-sampling quality concepts were informed by Yao GEO Skills, copyright 2026 Yao, under the MIT License. Aloom is not a complete implementation of that Skill: its prompt pack, runner, data model, scoring, and reports are maintained independently, and it does not depend on the Skill, Codex, OpenCLI, or the remote repository at runtime.
• Aloom itself is licensed under Apache-2.0 and is copyright 2026 MYZ8088.

See LICENSE and NOTICE for the complete notices.

Contact

For questions, feedback, or collaboration, contact z1250835369@gmail.com.