Back to Blog

Privacy in AI testing: map the data flows first

Map data flows before an AI visibility test: public questions, confidential documents, providers and retention.

Editorial review:

Reviewed factual claims and sources; added distinct examples and acceptance criteria. Observations, methodological recommendations and unconfirmed causes are distinguished.

TL;DR: Key Takeaways

Protection needs depend on the data actually transmitted. Hosting, model processing, logs and optional analytics need separate entries in the data map.

  • Public service question: Deliberately choose the receiving search surface.
  • Synthetic test contract: Label as test data; do not copy actual client details.
  • Actual contract file: Agree approval, roles, recipients and deletion period first.

A public question about a company’s visibility is different from uploading a contract containing personal data. Before a project, record what is transferred, who receives it, the purpose and retention period. Even a public question can contain personal information.

Context and scope

No-training statements need an appropriate contractual and technical basis. A Swiss server location does not automatically describe every external service’s processing. Review model access, hosting, email, analytics and logs separately.

A proxy can make the network connection to a downstream service. It does not anonymise names or contract details in the text. Public service questions or clearly labelled synthetic examples are often sufficient for testing. Client documents belong only in an appropriately approved workflow.

SOURCE/01 records these flows and open questions. It does not issue a blanket data-protection certification. GlasBox’s own privacy policy describes hosting and processing at Infomaniak in Switzerland and separate retention periods for logs, email leads and browser storage.

A data matrix before a visibility test

This template distinguishes four data types. Enter the actual providers and retention periods for the project before use. “Switzerland” without identifying the processing, or “session only” without a defined period, is insufficient technical documentation. The template does not replace a project-specific review of legal duties.

Review pointExample / subjectInterpretation
Public service questionName and public offerDeliberately choose the receiving search surface.
Synthetic test contractInvented information that cannot identify a personLabel as test data; do not copy actual client details.
Actual contract fileConfidential or personal dataAgree approval, roles, recipients and deletion period first.
Technical logTime, identifier, error and possibly request contentDocument access and actual retention.

How to apply this

  • Minimise the information needed for each test.
  • Record hosting and downstream recipients separately.
  • Define retention, deletion and the responsible person.
  • Review production privacy information after configuration changes.

Sources and further reading

The next step

Next step

The free short check provides an initial assessment of one domain, three buying questions and three observations. An introductory call establishes whether deeper analysis is useful.

Request an introductory call

Frequently Asked Questions

Does a proxy anonymise data?

No. Personal content remains personal data even when a server forwards it.

Does Swiss hosting guarantee exclusively Swiss processing?

No. External model, analytics or communication services need separate review.

How does GlasBox retain enquiries?

Under the confirmed operating policy, request logs are retained for 14 days and email leads for 5 years. Browser identifiers have separate periods, detailed in the privacy policy.