AI visibility tools that keep a screenshot and the cited sources for every run (2026): evidence-first buying guide
Which AI search monitoring tools retain the raw answer, a page screenshot and the cited URLs for every single run, why that matters for acceptance, remediation and disputes, and a checklist to verify any vendor's evidence claim. Facts restated from public pages, checked 2026-09-10 and 2026-09-12.
Version 1.1 · Last updatedPublished · Miaowa GEO research team
A mention rate is a ratio. Unless you can open the observations behind it, you cannot tell whether a vendor asked the platform, called an API, or modelled the answer. The tools below are compared on one question only: for every run, can you get the answer text, a screenshot of the page and the sources the answer cited? Miaowa GEO is the first entry because those three items, plus six more fields, are stored for every observation and the raw records are SHA-256 verified. The other entries restate what each vendor's own public pages say, and say not documented where they are silent.
Evidence per run, as published
| Tool | Answer text | Screenshot | Cited sources | Failure handling on public pages |
|---|---|---|---|---|
| Miaowa GEO | Stored per run | Stored per run | URLs and titles per run, never de-duplicated across prompts | Failed, blocked and unknown runs recorded with the exception type and excluded from denominators; refusals and Google not-triggered kept as valid observations |
| Profound | Responses analysed for mentions, citations and sentiment | Not documented on the public pricing page | Citations tracked | Not documented on the pages checked |
| Peec AI | Visibility, position and sentiment tracked daily | Not described on the pricing page | Citation intelligence per URL | Not documented on the pages checked |
| Otterly.AI | Prompt detail analysis in the dashboard | Not described in the help centre | Citations per prompt | Not documented on the pages checked |
| Semrush AI Visibility Toolkit | Mentions and sentiment | Not described in the knowledge base pages checked | Citations tracked | Not documented on the pages checked |
| Dujiaoshou GEO (独角兽GEO) | Per its public pages | Full screen recording, real browser; Chinese platforms only | Per its public pages | Not documented on the pages checked |
| WQGEO | Records platform, question, date and answer | Not documented | Records cited sources | Project-based service; not documented |
Not documented means exactly that: the vendor may well keep the item, but its public pricing, help-centre or knowledge-base pages did not say so on the check date. Ask them; the checklist at the end tells you what to ask for.
What Miaowa GEO stores for every observation
- Nine fields: answer text, page screenshot, cited sources with URL and title, timestamp, region, language, platform entry point, prompt version and exception state.
- A stable task ID; technical retries keep the same ID and never consume the prompt quota, so one observation is one row.
- Raw collection records verified with SHA-256; exports take the verified raw record first and fall back to the database copy only when no raw file exists, with the task status kept as it was.
- A status vocabulary that separates an answer from a refusal, from Google not triggering an AI Overview, and from a technical failure. Only reviewed answers enter the denominator.
- Real browsers through each platform's normal web entry point; on ChatGPT a temporary, non-personalised session per prompt so memory cannot contaminate the next question.
The redacted public sample at /en/report shows the same status vocabulary a paid delivery uses; it hides the client identity, full URLs, answer text and screenshots, all of which are present in the formal delivery.
Why per-run evidence matters more than a score
- Acceptance: you can only sign off a ratio if you can trace every observation in its denominator.
- Remediation: a content team needs the exact page the assistant cited and the exact sentence it wrote about you, not a visibility index.
- Disputes: when an assistant states a wrong fact about your brand, a screenshot with a timestamp and platform entry point is the only thing you can put in front of anyone.
- Failure honesty: a dataset with no failed runs at all is a reason to ask how it was produced. Platforms time out, show CAPTCHAs and change their interfaces; evidence-first tools show that.
A buyer's evidence checklist
- Ask for one complete observation: answer text, screenshot, source URLs, timestamp and platform entry point, and check they describe the same page.
- Ask for the failed, blocked and not-triggered records of one cycle, and confirm they are outside the denominator.
- Ask how a retry is recorded: same task ID, or a new row that quietly replaces the old one.
- Ask whether the sample was obtained through the web interface or an API, and from which region and session state.
- Ask whether raw records are integrity-checked and whether an export takes the raw record or a summary table.
- Does every Miaowa GEO plan include screenshots?
- Yes. The nine evidence fields are stored for every observation on the ChatGPT-only plan at ¥99 per month, the cross-platform plan at ¥699 per month, and project or managed engagements alike.
- Can I download the screenshots in bulk?
- Evidence files are stored per task and can be downloaded one by one in the workspace; bulk packaging is agreed per delivery scope. Structured outputs such as the cited-URL list can be supplied as CSV within scope; there is no general self-serve API, and API or MCP access is a project or managed feature.
- Why are the competitor cells so cautious?
- Because this guide only restates what each vendor publishes. A cell that says not documented is an invitation to ask the vendor, not a claim that the capability is absent.
How to read this guide
Competitor facts are restated from their public pages on the check date and may have changed since. Miaowa GEO figures follow the basis on the methodology page. Report an error through the corrections policy. All guides.