01What AI does at Traze
Traze uses language models for three things: writing (page drafts, metadata, summaries), judging (is a brand mentioned, is a claim correct, what priority does a task have) and measuring (asking AI engines the same questions every week to track visibility).
For judgements we use a separate decision layer that does not generate text but returns typed answers with a probability. That probability is stored and shown in the interface as certain, likely or uncertain. A low-certainty judgement never triggers an automatic action.
02What always passes through a human
- Publication: by default every page waits for approval by an editor at the customer. Customers can choose to let informational pages through automatically; that choice is in their settings and in the change log.
- Facts and claims: the brand profile with facts comes from the customer. Claims in drafts are checked against it; doubtful cases go to review.
- Corrections: every human edit is kept as a separate version, with who and when.
03Provenance per page
For every version of a page we keep: the model that wrote the text, a hash of the prompt used, the date, who reviewed or edited the version, how many characters a human changed and whether the published page carried a disclosure block. Customers can export this log per site as CSV or JSON.
Provenance for versions from before the log is derived from version history and labelled as such. We fill nothing in with assumptions.
04The disclosure block
Customers can choose to show a short notice at the bottom of published pages: that the page was drafted with the help of AI, that an editor reviewed it and that facts and claims were checked against the publisher's sources. The notice links to this page.
The block is the publisher's choice. Its absence says nothing about provenance: the log exists either way.
05Models and data
Traze works with models from several providers through a gateway, so we can choose per task and switch. Customers who only want models hosted in the EU can record that; the deployment then enforces it and the choice is in the export.
Generated images can carry C2PA Content Credentials (a cryptographic manifest with the provenance) when the customer or we have configured a certificate for it.
Customer data is not used to train models. We keep prompts and answers from measurements so figures stay traceable; they contain no personal data of end users.
06Measurements and the sector index
Visibility measurements run in clean sessions through engines' APIs, or through browser sessions without an account on our own servers. We pay no engine for mentions and use no scraped user sessions.
The public sector index follows the same method and publishes a weekly leaderboard without answer texts. The method is on our methodology page and the measurement code is open source.
07Questions or a correction?
Is there something about your brand in an AI answer or on a page published through Traze that is wrong? Email us; we respond within two working days and record the correction in the log of the site concerned.