Are these real model answers or simulated data, and how can a buyer verify that?

They are real answers. Every Miaowa GEO observation comes from a real browser asking the prompt through the normal web entry point of each AI product, following the same path a buyer would. We do not substitute API output for web results and we do not generate simulated data. Each run stores the answer text, a page screenshot and the cited sources; raw collection records are hashed with SHA-256; tasks that fail are recorded as failed rather than dressed up as observations. You do not need to take this on trust: the checklist at the end shows how to verify it yourself.

What makes the collection real

• Collection runs in a browser environment outside mainland China through the web entry points of ChatGPT, Gemini, Perplexity, Google AI Overviews, Microsoft Copilot and Claude, so the page we capture is the page an overseas buyer sees.
• No API stand-ins: model configuration, web retrieval and source display often differ between an API and the consumer product, so API output would misrepresent what buyers actually see.
• On ChatGPT each prompt runs in a temporary chat with personalisation switched off, so memory does not apply and one question cannot contaminate the next; every prompt gets its own session.
• Each prompt runs once per platform by default against a prompt set frozen at setup, which is what keeps cycles comparable.

Why the screenshot and answer text count as evidence

Every run stores the answer text, page screenshot, cited sources (URL and title), timestamp, region, language, platform entry point, prompt version and exception status. Raw collection records are hashed with SHA-256, and exports prefer the verified raw record. Every number in a report therefore traces back to a hashed original rather than to a summary table assembled afterwards. Evidence and reports are reviewed separately by people; only reviewed observations enter the denominator, and anything that cannot be verified is flagged as pending rather than stated as fact.

Failures are not hidden

Timeouts, CAPTCHAs, expired sessions and UI changes all cause individual tasks to fail. Failed, blocked and unknown tasks are excluded from business metrics but remain on record, and nothing is re-asked outside the cycle once it closes. A monitoring dataset with only successes and no failures at all is the one that deserves a follow-up question about where it came from.

A buyer's verification checklist

1. Ask for screenshots: pick a few observations at random and compare the screenshot with the stored answer text.
2. Ask for source URLs: check that each cited source is stored with its URL and title, not just a citation count.
3. Ask for failure records: request one cycle's failed, blocked and not-triggered tasks and confirm they are outside the denominator.
4. Ask for task IDs: confirm every observation has a stable task ID and that technical retries reuse it instead of creating a new row.
5. Compare with /en/report: the public sample uses the same status vocabulary, so review it first and then require the formal delivery to match that shape.

APIs are faster and cheaper. Why not use them?

Because the object of measurement is the answer a buyer sees in the web product. API output differs in retrieval and source display, and using it would strip mention rate and citation rate of their business meaning.

Could a screenshot be edited after the fact?

Raw collection records are hashed with SHA-256 and exports prefer the verified record. The screenshot, answer text, sources and timestamp sit on one task record; altering any of them breaks the hash.

Can collection run from inside China instead?

No. Collection runs in a browser environment outside mainland China because overseas buyers use the overseas entry points of these platforms. Miaowa GEO also does not monitor Doubao, DeepSeek or other Chinese platforms.

Last updated 2026-09-12. The capabilities page and the pricing page are authoritative where they differ.