Personalisation
Personalisation is showing different content to different users based on what you know about them. The simple version is rule-based - if the user is logged in, show their name; if they’re returning, skip the explainer. The complex version is ML-driven - rank products, recommend offers, sequence content based on a model trained on past behaviour.
The argument for personalisation: average optimisation has a ceiling. The same headline that converts cold paid traffic kills brand search. The same product carousel that works for first-time visitors annoys repeat customers. Once the universal experience has been optimised, the next leverage is varying the experience by segment.
What gets personalised
Section titled “What gets personalised”The common surfaces:
- Product recommendations. “Customers who bought X also bought Y”, “Recommended for you”, “Trending in your size”. The most common form, often ML-driven on bigger sites.
- Homepage hero and modules. New vs returning visitors. Cold vs warm. Mobile vs desktop. Each surface state can be a different “default”.
- Email and lifecycle content. Same campaign, different copy per segment. The lowest-effort form because the variation lives in email, not the live site.
- Pricing and offers. Geographic pricing, tier-based discounts, retention offers. Sensitive territory - personalised pricing visible to multiple users in the same household is a trust risk.
- On-site copy and CTAs. Different headlines for different traffic temperatures, different awareness stages, different past behaviours.
How personalisation interacts with A/B testing
Section titled “How personalisation interacts with A/B testing”Personalisation and A/B testing pull in opposite directions. A/B tests want clean comparisons across uniform traffic. Personalisation wants different experiences per user. Done together they have to be sequenced carefully.
The usual pattern: A/B test the personalisation rule itself. “Does showing recommended products on the PDP lift conversion?” is an A/B test where the variant is the personalised experience and the control is the static version. The test measures whether personalisation is worth it on this surface, not whether the recommendations are right.
Once personalisation is live, evaluating individual recommendation quality is harder. Holdouts (a permanent slice that gets no personalisation) are the cleanest way to measure long-run effect.
Where it goes wrong
Section titled “Where it goes wrong”- Personalising before you’ve measured it. Shipping a recommendation widget because it “should work” without testing whether it actually moves conversion. Often it doesn’t.
- Over-personalising. When every visitor sees a different page, debugging gets hard and the brand experience fragments. Personalisation works best when it varies one or two clear axes, not everything at once.
- Building ML personalisation before the basics. A recommendation engine on a site with broken navigation is solving the wrong problem.
- Ignoring the data infrastructure cost. Personalisation requires reliable identity, behaviour history, and a real-time decision layer. Underbuilt infrastructure produces bad recommendations that hurt the experience more than no personalisation would.
When not to personalise
Section titled “When not to personalise”- Small audiences. Personalisation algorithms need data to train. Below a certain traffic threshold, the rules are operating on too little data to be reliable.
- First-time visitors. No history to personalise on. Better to use the static experience and earn the right to personalise on the second visit.
- High-stakes single decisions. Big-ticket B2B purchases where the buyer wants comprehensive information, not curated highlights.