Chinese AI Visibility Analyzer

See whether your brand appears in current, web-grounded answers from four major Chinese AI families:

• Moonshot AI — Kimi K2.6
• Z.ai — GLM 4.6
• DeepSeek — V3.1
• Alibaba — Qwen 3.5 Flash (02-23)

Enter a brand and up to five buyer-intent questions. The Actor asks every selected model the same neutral question, requires web search and structured citations, then returns comparable brand mentions, positions, competitors and sources. No model API key is required.

The Actor also keeps its original import workflow for teams that already have AI answer datasets.

Live analysis

The default input runs all four families:

{
  "mode": "live",
  "brand": "Nike",
  "brandAliases": ["Nike, Inc.", "耐克"],
  "competitors": [
    { "name": "Adidas", "aliases": ["阿迪达斯"] },
    { "name": "New Balance", "aliases": ["新百伦"] },
    { "name": "On", "aliases": ["昂跑"] }
  ],
  "queries": [
    "Which sustainable running shoe brands should a US buyer compare?"
  ],
  "models": [
    "moonshot-kimi",
    "zai-glm",
    "deepseek",
    "alibaba-qwen"
  ],
  "market": "United States",
  "language": "English",
  "maxModelConcurrency": 2
}

Each query runs once per selected family. One query with all four models creates up to four paid observations; five queries create up to twenty.

Model family keys stay stable even when a tested provider endpoint changes. Each output row records the exact provider, modelFamily and modelId used. Kimi is identified as a Moonshot AI model rather than appearing as an unrelated provider.

Neutral prompting and branded vs. organic queries

The model is never told which brand or competitors you are tracking. Each model call receives only the query, market, language and neutral JSON/citation requirements — the tracked brand, its aliases and its competitors are applied afterwards, deterministically, against the returned answer text. This keeps mentionRate honest: a model cannot inflate it just because the prompt hinted at the brand it should mention.

Every output row also carries a queryType of branded (the query itself names the tracked brand, an alias or its official domain) or organic (it does not). Set the optional brandWebsite input to the brand's official site so branded-query detection and source-backed evidence can use the real domain; if omitted, a URL-shaped brand alias is used instead. Organic queries are the more meaningful signal — they show whether a model recommends the brand unprompted.

Evidence contract

A live observation is delivered only when the dependency returns:

1. a complete object with answer and ordered brandsMentioned;
2. confirmation that web_search was used;
3. at least one structured HTTP or HTTPS citation.

Missing citations, malformed answers and model failures are free diagnostics in QUARANTINE. They are never converted into a false “brand not mentioned” result. A Run is complete only when every requested model produced a grounded observation; otherwise it is partial, partial_budget or dependency_failed.

Every mentioned row additionally carries a sourceBackedMention verdict, based on scanning the model's cited pages (official domain, or brand text in a cited URL/title/page content — never on trusting the model's own claim):

• true with mentionEvidence: "official-domain" — a cited page belongs to the brand's own domain;
• true with mentionEvidence: "citation-text" — a cited third-party page names the brand;
• false with mentionEvidence: "answer-only" — the brand appears in the answer text but no cited source backs it up, after a full scan;
• null with mentionEvidence: "unknown-import" — only for rows imported from another source without enough citation detail to decide either way (never asserted for a fresh live observation from this Actor);
• mentionEvidence: "not-mentioned" whenever the brand is not mentioned at all.

Summary and Dataset

The default Dataset contains successful observations with:

• deterministic brand mention and position;
• queryType (branded/organic) and sourceBackedMention/mentionEvidence/ supportingCitedUrls;
• provider, model family and exact model ID;
• detected competitors;
• cited URLs and domains;
• grounding status and quality flags;
• bounded evidence and child Run provenance.

Four Key-value store records support review:

• SUMMARY — completion, coverage, mention rate by model, source-backed mention rate, a branded-vs-organic breakdown, competitors and cited domains for live runs;
• OUTPUT — run status, billing counts and child Run receipts;
• COVERAGE — normalization and grounding counters;
• QUARANTINE — bounded free diagnostics.

Mention rate uses grounded observations as its denominator. A failed model is shown as missing coverage and does not reduce the brand's score.

SUMMARY reports both the legacy mentionRate (renamed answerMentionRate, kept for backward compatibility — the brand appears anywhere in the answer) and the stricter sourceBackedMentionRate (a citation actually backs the mention up). organicSourceBackedMentionRate narrows that to organic queries only — the cleanest read on unprompted visibility. byModel and byQueryType break both rates down further.

Import existing observations

Set mode to import and provide inline JSON, CSV or up to ten buyer-accessible Apify Datasets:

{
  "mode": "import",
  "brand": "Acme Running",
  "competitors": [
    { "name": "StrideLab", "aliases": ["Stride Lab"] }
  ],
  "inlineRows": [
    {
      "sourceActorId": "apify/chatgpt-search-scraper",
      "query": "Which running shoe brands should beginners compare?",
      "text": "Acme Running and StrideLab are options to compare.",
      "sources": [
        { "url": "https://example.com/running-guide" }
      ],
      "checkedAt": "2026-09-15T00:00:00.000Z"
    }
  ],
  "market": "United States",
  "language": "English",
  "maxRows": 100
}

Older inputs that omit mode and contain inlineRows, inlineCsv or sourceDatasetIds are automatically treated as imports.

Supported adapters include Bulk LLM Runner, ChatGPT Search Scraper, Google AI Overviews, Perplexity, Gemini, the owned visibility collectors and ai-observation-1.0.

Pricing

This Actor uses Pay Per Event:

Tier	Actor start	Imported observation	Grounded live observation
Free	$0.005000	$0.005000	$0.040000
Bronze	$0.004750	$0.004750	$0.038000
Silver	$0.004500	$0.004500	$0.036000
Gold	$0.004250	$0.004250	$0.034000
Platinum	$0.004250	$0.004250	$0.034000
Diamond	$0.004250	$0.004250	$0.034000

At the Free tier, one query across all four models costs at most $0.165: $0.005 for the Run plus four grounded observations at $0.04 each.

The Actor checks the buyer's remaining charge limit before starting model work. It caps the child dependency at $0.008533 per attempted observation. Errors, quarantine, coverage and summary records do not emit result events.

Limits

• Live mode accepts 1–5 unique queries and 1–4 fixed model families.
• Imported mode accepts up to 1,000 source rows.
• Only structured source citations count as grounding.
• The Actor measures visibility and evidence structure. It does not determine sentiment, factual correctness or commercial impact.
• Provider availability can change. Exact model IDs remain visible in every result and failed paths remain explicit.