What evidence is stored for every observation?
Nine items per observation: answer text, page screenshot, source URLs with titles, timestamp, region, language, platform entry point, prompt version and exception status. They sit under one stable task ID that technical retries reuse; raw records are hashed with SHA-256, and evidence and reports are reviewed separately.
Version 1.1 · Last updatedEvery observation stores nine items: the answer text, a page screenshot, the cited sources (URL and title), the sampling timestamp, region, language, platform entry point, prompt version and exception status. They are bound to one stable task ID, which technical retries reuse; raw collection records are hashed with SHA-256, and exports prefer the verified raw record. Evidence and reports are reviewed separately by people, and the workspace shows an evidence-completeness percentage for each cycle.
What each field answers
| Field | Content | Question it answers |
|---|---|---|
| Answer text | The full text of the AI answer | Was the brand mentioned, and how was it described? |
| Page screenshot | The page as it appeared when asked | Does the text match the page; has anything been altered? |
| Cited sources | Every source URL and title attached to the answer, not deduplicated across prompts | Whom is the AI citing; which pages should remediation start from? |
| Timestamp | When the sample was taken | Which cycle it belongs to; whether it can be compared with other cycles |
| Region and language | Where the collection ran and in which language the prompt was asked | Is this the version an overseas buyer sees? |
| Platform entry point | Which platform and which web entry point was used | Does the result correspond to the web product rather than an API? |
| Prompt version | Version of the frozen prompt set | Are cycles asking the same set of questions? |
| Exception status | success / failed / blocked / unknown / refused / not triggered | Can this record enter the denominator? |
Task IDs and technical retries
Each observation carries a stable task ID; in the workspace you can drill from a cycle down to a single task and see its sampling time, status and screenshot. A task that hits a technical exception is retried automatically under the same ID, without consuming prompt quota and without a fresh success row appearing to cover the earlier failure. A task that still fails after retries is recorded as failed or blocked, with the exception type, timestamp and any captured screenshot.
Review and evidence completeness
Evidence and reports are reviewed as two separate steps: the evidence review checks that screenshot, text, sources and status agree with each other; the report review checks metric definitions and denominators. Only reviewed observations enter the denominator, and anything that cannot be verified is flagged as pending. The monitoring log shows evidence completeness per cycle, meaning the share of planned tasks that have complete evidence, alongside the runtime-exception rate. Reading that number before the mention rate is the safer habit.
What the public sample redacts
The public sample at /en/report uses the same status vocabulary but hides client identity, full URLs, answer text and screenshots. All four are present in a formal delivery: the workspace shows them per task, evidence files are stored per task and downloadable individually, and bulk packaging is agreed per delivery scope.
- How long is evidence kept?
- If a subscription lapses, the workspace and history remain available for viewing and export for 30 days. After an engagement ends, exports are available for 30 days from the end of the last subscription month; within the following 90 days client data and monitoring evidence are deleted or anonymised.
- Could the evidence contain personal data about our staff?
- Not from our side. The only data that leaves mainland China is the project's prompt text plus answers, screenshots and sources obtained from the AI platforms' public pages. Clients should make sure their prompts and brand materials contain no personal data.
Last updated . The capabilities page and the pricing page are authoritative where they differ.
Help centre
Back to the index, or email contact@miaowageo.com.