Is one-to-one personalisation worth it compared with segments?

Sometimes, but less often than it sounds. It needs accurate, lawfully usable data per person and content that works in every combination. Well-chosen segments deliver much of the relevance at lower risk. Individual fields, such as a name or nearest store, are worth adding where the data is reliable and the value to the customer is obvious.

What does GDPR mean for AI-personalised marketing?

You need a lawful basis for the profiling, clear information about it and an easy way to object; for direct marketing, an objection must always be honoured. Avoid inferring sensitive traits. Minimise the data each template uses, and make sure no personal data reaches a model or provider that retains or trains on inputs.

How do you stop personalised content becoming creepy?

Use data the customer knowingly gave or would expect you to hold, such as their market, purchases or stated preferences. Avoid inferences about health, money trouble or life events, and avoid referring to single browsing moments. If a customer would be surprised by how you knew something, leave it out, however well it tests.

AI Personalisation at Scale

Personalisation at scale, within data and consent limits.

Personalisation at scale means producing versions of content for segments or individuals without producing a mess. For most brands the return comes from well-chosen segments, such as market, life stage, product owned or intent, more than from one-to-one messages. The constraints are consent and data law, template quality across every variable, a holdout group to prove the effect, and a clear line past which personalisation feels like surveillance.

Segments before individuals

One-to-one personalisation sounds better than it usually performs, because it needs data that is accurate, current and lawfully usable for each person, and content that makes sense in every combination. Segment-level personalisation, by market, language, product owned, life stage or lapsed versus active, captures much of the relevance with far fewer ways to go wrong. AI changes the economics: a distinct creative for each of forty segments is now practical, where it once meant forty shoots. Start with segments whose needs genuinely differ, and personalise individual fields, such as a name or nearest store, only where the data is reliable.

Data and consent boundaries

Personalisation runs on personal data, so the law sets the edges. Under GDPR you need a lawful basis for the profiling behind it, and people can object to profiling for direct marketing at any time, with no balancing test. Inferring sensitive traits such as health, religion or sexuality from behaviour can create special category data and should be off limits for marketing. The EU Digital Services Act adds that online platforms may not show ads based on profiling with special category data, nor profiling-based ads to users they are reasonably certain are minors. Use only the fields a template needs, and never pass personal data to a model that retains inputs.

Templates with variables, tested at the edges

Personalised content is a template plus data, and the failures live at the edges of the data. Test every variable at its extremes: the longest name, a missing field, a name with an apostrophe or accent, a right-to-left language, a product with no image, a store that has closed. Define a fallback for each field, so a missing first name produces a sensible greeting rather than an empty gap. For generated imagery and video, lock what must not vary, such as product, logo and legal lines, and let only the variables change. Review a sample drawn from every segment, not just the default version.

Measurement and the creepiness line

Prove the effect with a holdout: a random group that receives the generic version, so any lift can be credited to personalisation rather than to timing or the offer. Then watch the creepiness line. Using a customer's city or last purchase feels helpful; referencing something they browsed once, a life event they never shared or a price they almost paid feels like being watched. A useful test is whether the customer would understand how you knew. In the studio, batch generation can run from a product feed with the brand hub's approved claims and banned words enforced at generation, so versions stay on brand as they multiply.

Updated 25 September 2026 · General information, not legal advice. Rules change, so check the current text with your legal team before relying on it.