THE SHORT ANSWER
AI can use available signals to rank options, estimate likely responses and adapt an experience. Personalisation is useful when it helps a customer accomplish something. More detailed inference is not automatically more relevant, and it creates questions about privacy, fairness and control.
Separate selection, prediction and generation
A recommendation system selects what to show. A predictive model estimates an outcome. A generative model creates material. They can work together, but they need not. A helpful recommendation does not require a freshly generated message for each person.
Google's recommendation-system teaching material describes candidate generation followed by scoring and ranking. The choice of scoring objective matters: a system arranged to predict clicks is solving a different problem from one intended to help someone find a suitable product.
| Approach | Illustrative use | Question to ask |
|---|---|---|
| Segment rule | Show setup guidance to new customers | Is the group definition still useful? |
| Predictive ranking | Order products by estimated suitability | What outcome trained the ranking? |
| Generated adaptation | Explain the same approved feature for a chosen context | Does every version preserve the actual promise? |
Evidence & context: Google Machine Learning education
A score is not a fact about a person
GA4's documented predictive metrics require sufficient qualifying data and model quality; they are not available universally. Treat a prediction as an estimate under particular conditions, not as knowledge of someone's intention.
Imagine a retailer whose customer buys a gift once. Repeatedly treating that purchase as a lasting personal preference can make the experience less useful. Let customers express what they need now, and avoid presenting an inference as if they explicitly told you.
Evidence & context: Google Analytics Help
Relevance has a boundary
A visitor may welcome a filter for products available locally but dislike a message implying that the brand has inferred a sensitive personal circumstance. Design from the benefit the customer can understand, then ask what information is actually necessary.
The UK ICO's AI guidance discusses data minimisation in terms of information needed for a specified purpose. This is jurisdiction-specific guidance, not permission to process a particular dataset. Get appropriate specialist review of lawful use, consent and data handling for your market.
Provide a useful default when a signal is missing or a person declines personalisation. Offer correction or preference controls where relevant. Do not make disclosure of unnecessary information the hidden price of a usable experience.
Evidence & context: UK Information Commissioner's Office
Test assistance, not just response
In a hypothetical product selector, measure whether people find a suitable item, understand its limitations and avoid returns caused by mismatch. Click-through rate can help diagnose the interface, but cannot settle whether the recommendation served the customer.
- Compare with a simple, well-designed default.
- Check whether outcomes differ across relevant groups and situations.
- Watch complaints, exclusions and repeated irrelevant recommendations.
- Keep a way to pause or reverse an adaptation that causes harm.
Use marketing analytics to distinguish a promising response pattern from evidence that the intervention helped.
Sources & further reading
- Recommendation systems: scoring
Google Machine Learning education. Conceptual explanation of scoring and ranking candidates, not a description of every current commercial recommender.
- GA4 predictive metrics
Google Analytics Help. Documents data and model-quality prerequisites. Checked 11 September 2026; prediction is not evidence of causal impact.
- How should we assess security and data minimisation in AI?
UK Information Commissioner's Office. UK regulatory guidance, checked 11 September 2026. Jurisdiction-specific context, not individual legal advice or permission for a particular use.
Examples and exercises are illustrative unless attributed to a source. No independent expert review is claimed.
A correction, a counterexample or an experience worth sharing?
Join the conversation ↗