The recommendations are rules. The model is optional.

Rule-based analysis runs on your data by default. Models stay off until you switch them on.

Content Recommendations is not a model

The recommendations in your dashboard come from rule-based detectors reading your own aggregates: a page converting well on little traffic, a page taking the traffic but not the conversions, a question your visitors search for that no page answers. Deterministic, inspectable, and the same input always gives the same card. No model is involved, and none is needed.

Off until you say otherwise

The model features — an Ask box, narration over those same rule-based cards, a natural-language funnel builder and a page-classification pass — are built, and they stay off until you turn them on. While the switch is off, no data reaches a model anywhere in the product: not a summary, not a sample, not a prompt. The rule-based analysis runs either way.

Your analytics data is not training material

It is not pooled across customers and it is not used to train a model. If you turn the model features on, what is sent is sent to answer your question and for no other purpose.

Seeing AI assistants in your traffic

Separately, and with no model involved: visits referred from ChatGPT, Claude, Perplexity, Google Gemini and Microsoft Copilot are recognised and reported as their own channel. That is referrer classification against a maintained list — useful precisely because this traffic is invisible as "direct" in tools that do not look for it.