Menu

繁简EN

AI 可見度報告

How we test

This report is scored with the PALU four dimensions (Presence / Accuracy / Listing completeness / Understanding).

Each AI engine answered each prompt once. Results are a snapshot from the day of asking; AI answers can change each time, and this is not a long-term audit.

How the four scores are counted: P is how many of the six approved engines mention you on industry prompts that do not name the brand (weekly monitoring uses five); A is correct checkable facts ÷ (correct + incorrect); L is complete, matching readable sources ÷ readable assessed sources; U is answers about you that get the business and service area right ÷ answers about you. Results are shown per dimension (P/A/L/U) with no overall total.

The free brief is not scored with the PALU four dimensions (P Presence / A Accuracy / L Listing completeness / U Business understanding). It only checks whether your brand is present in the AI world: whether a single engine mentioned your brand on the fixed question set for this run. This is a quick check using a fixed question set, and it is not a full visibility analysis. For data gaps, confusion, and suggested changes, see the detailed report.

Engines and models

Engines are fixed by PALU, not selected by the customer. The HK detailed offer covers Perplexity, ChatGPT, Gemini, DeepSeek, Grok and Kimi. Each report lists the AI engines actually used for that run. Only enabled engines are tested; the names below do not imply all six were available on a past test date. Models are recorded for each response and may change between test dates.

  • Perplexity: A search-focused AI with sources, representing users researching information and comparing options.
  • ChatGPT: The world's most widely used AI assistant, representing the general public.
  • Gemini: Google's AI, connected to Google Search, representing Google AI and Search users.
  • DeepSeek: No web search in this test; answers reflect model memory.
  • Grok: xAI's AI; no web search in this test.
  • Kimi: Commonly used in mainland China, with strengths in reading long documents.
  • Perplexity: sonar
  • ChatGPT: gpt-5.1
  • Gemini (Google): gemini-flash-latest
  • DeepSeek: deepseek-flash
  • Grok: grok-4.7
  • Kimi: kimi-k2.6

Public samples use anonymous engine labels; paid reports show real engine names.

Definitions and formulas

PALU is the four dimensions GEOPALU uses to measure AI answers. L checks the required information in assessed website and platform sources, not whether AI cites their links. U checks business-category or service-area confusion in answers that recognise the brand, not its strengths or audience.

P Presence

P: Paid reports: distinct approved engines mentioning the brand in answers to industry questions that do not name the brand ÷ 6 × 100%. Weekly monitoring uses five approved platforms: mentioning engines ÷ 5 × 100%.

A Accuracy

A: Correct verifiable facts ÷ (correct + incorrect verifiable facts) × 100%. Missing or unverified facts are excluded. Price, hours, address, contact and website are checked; a correct official name earns no point, while an incorrect name counts as an error.

A and U count only answers identified as being about the customer: the AI must recognise the brand. Answers that do not mention the customer in the answer or its sources, or that mistake another business for the customer, are excluded from both A and U.

L Listing completeness

L: Complete, matching customer sources ÷ readable assessed sources × 100%. Required fields and cross-source agreement are checked. Unreadable sources are excluded; readable incomplete or misidentified listings count against completeness.

  • Markets outside mainland China: official website JSON-LD + Google Business (2 items).
  • Markets outside mainland China, online-only businesses: official website JSON-LD (1 item). Google Business is skipped only when the paid questionnaire answer specifies online-only operation. A missing answer or an old order still includes the two items above.
  • China market: official website + customer screenshot business listing (up to 2 items). Website JSON-LD always counts as one. Uploaded screenshots add one listing item, including online-only businesses. No upload: denominator 1. Listing requires a listed business, matching name and matching address or phone. Unrecognized screenshots earn no points. Based on customer-provided screenshots.

U Understanding

U: Answers about the customer without business or service-area confusion ÷ answers about the customer × 100%.

Limits and exclusions

Failed engines and engines with fewer than three successful unbranded questions are excluded from A and U. Missing keys, errors, timeouts and missing mainland-transfer consent are missing data. Paid report P needs all six approved engines; weekly monitoring P needs all five. Incomplete paid reports remain under review. Answers about another business or with no mention are excluded from A and U. Fewer than five answer samples suppress a percentage; A also needs at least three checked concrete facts. Misidentified brands are excluded from A/U. Answers identifying the brand but confusing its business or service area count in U as confused. Reports show P/A/L/U separately without a total score.

These are API tests with web search where supported. Consumer apps can use different models, search tools, personalization, location and conversation history, so their answers may differ. Source links are evidence from that test, not proof of continuing accuracy. Implementation code is not published.

One-off reports from different dates should not be compared with each other. AI answers change over time, and each report is a snapshot of its test date.