Transparency.
A measurement company should be easy to check. This page sets out what Visibian measures, what it sends to the AI providers, what it keeps, and where its numbers have limits.
The first rays of a sunrise on Earth's atmosphere, from the International Space Station above the Atlantic, 12 January 2020. Photo: NASA.
What we measure
For each brand a customer tracks, Visibian asks AI assistants the questions that brand's buyers ask, daily or weekly as the customer chooses, and records what comes back: whether the brand is named, where it appears in the answer, how it is described, which competitors are named beside it and which sources the assistant cites. It also flags claims about the brand that contradict facts the customer has recorded.
The assistants are ChatGPT, Google Gemini, Google AI Mode, Perplexity, Microsoft Copilot, Grok and Claude.
What we send to the AI providers
Collecting answers means asking questions of those seven assistants. Three rules govern that traffic, as our principles page sets out.
- Questions, not identities. Providers receive the questions being asked, and nothing about the customer, their account or their data.
- Retention switched off. Wherever a provider offers an opt-out from storage or training, we set it on every request, and our tests check that it is sent.
- Paid tiers where it matters. Where only paid traffic is excluded from model improvement, we pay.
What we keep, and for how long
Results stay for as long as the account does. Everything else expires on a fixed schedule, enforced by scheduled jobs. The main periods, from our privacy policy (section 11, updated 20 October 2026):
| Account, settings, collected answers | Life of the account, then removed within 90 days of closure |
|---|---|
| Deleted brand portfolios | 30 days, then permanently deleted |
| Provider request logs: prompt and response text | 90 days |
| Provider request logs: tokens, cost and timing | 730 days |
| Cached copies of cited public pages | 365 days |
| Server logs | 14 days |
| Backups | Nightly, kept 14 days |
| Billing records | 10 years, as the law requires |
What we hold for a customer never trains any model, and we never sell it.
Where our numbers have limits
AI answers vary. The same question can name different brands from one asking to the next, because assistants draw their words with some randomness and the pages they search change. Our explainers on temperature and how assistants pick answers show why.
So a reading built on too little data says so on the screen, and we would rather show a range than a false certainty. Every flag about a wrong claim points to the exact sentence and the recorded fact it contradicts.
We measure what assistants say. We cannot see inside them, and no assistant publishes why it chose one brand over another. When our reports describe causes, they say what the evidence is and how strong it is.
What we do not do
We do not place ads, buy mentions or promise positions in AI answers. We show what is true, why it is likely true, and which of a customer's own pages could change it.
Changes to this page
27 October 2026: first published.