How we use AI

Our agents only propose. Each proposal is an “action” that shows the current state, the proposed change and the reason for it.

• No agent can approve an action. Only a signed-in person with the Approver, Administrator or Owner role can, and this rule is enforced on our servers and tested.
• Nothing is written to your website before approval. New content is created as a draft in your CMS by default.
• Every deployed change keeps a snapshot of what was there before, so it can be rolled back. If the live check after deployment fails, the change is rolled back automatically.
• Every human and agent action is recorded in an append-only audit log that you can read in the app.
• One exception, off by default: you can pre-approve two low-risk action types — image alt text and meta descriptions — for a given project. You must switch it on yourself, per project.

• Your expertise first: drafting is blocked until the brief contains at least one fact specific to you — a figure, a case, a review or a photo.
• Sources: at least three verifiable sources per draft. Any figure without a source is flagged and blocks approval.
• Uniqueness: a draft more than 85% similar to an existing page of your site, or to another draft of the same project, is blocked.
• Volume: publications are capped per project and per week. We do not mass-produce pages for query variations.
• Information gain: the reviewer sees what the draft adds compared with the top results, with an explicit warning when it adds too little.
• Excluded sectors: we refuse projects and content related to gambling, alcohol, interest-based finance and adult content.
• Labelling: exported content carries machine-readable metadata stating that it was produced with AI assistance and reviewed by a person (EU AI Act, article 50).

• The application and its database are hosted in the European Union (Hetzner).
• We ask model providers for zero or minimal data retention wherever they offer it; this is being confirmed contractually with each of them.
• Pages, search results and AI answers we collect are treated as data, never as instructions to our agents. Prompts keep them strictly separate.
• For debugging, the full input and output of each model call is kept for 30 days, then deleted. We keep the metadata (model, tokens, cost, prompt version) to measure cost and quality.
• Your connection credentials (Search Console, CMS…) are encrypted (AES-256-GCM) with keys held outside the database.

• AI answers change from one day to the next. We ask each question several times and show visibility with a confidence interval, and we favour trends over single values.
• We measure the effect of each change at 14, 28 and 56 days against a control group of comparable pages. A result is shown as a gain only when it is positive and its 80% confidence interval excludes zero; otherwise it is marked “not conclusive”.
• Results that overlap a Google update are flagged “interpret with caution”.
• Models can be wrong. That is why every change is reviewed by a person and every claim with a figure needs a source.

A language is opened to customers only after it passes evaluation sets reviewed by native speakers: search intent (at least 85% agreement), keyword clusters (at least 90% coherent), drafts (average score of at least 4/5 with no factual error) and reading of AI answers (at least 90% accuracy on brand mentions). Any change of model or prompt re-runs the relevant sets.