Personalisation has a credibility problem. Too many implementations amount to a “recommended for you” carousel of products the visitor already bought, or a popup that addresses them by the wrong name. Done well, though, on-site personalisation is one of the most reliable ways to lift conversion — because it removes effort for the shopper rather than adding noise. The trick is being specific about what you personalise, for whom, and how you know it worked.
What personalisation actually does for conversion
Personalisation converts when it reduces the work a shopper has to do to find the right thing. Every irrelevant product, every generic message, every extra click is friction. Surfacing the right category, the right size, the right reassurance at the right moment shortens the path to purchase.
That framing matters because it rules out a lot of vanity personalisation. Showing someone their name in a banner does nothing for effort. Showing a returning visitor the three products they viewed last session, with current stock and price, removes a real step. Aim for the second kind.
What to personalise
Not everything is worth personalising, and over-personalising a thin signal looks creepy and performs worse than a good default. In our experience the highest-return surfaces are:
- Homepage and landing modules. Lead with the category or use-case a visitor has shown interest in, rather than a static hero that suits no one in particular.
- Product recommendations. “Complete the look”, “frequently bought together” and “because you viewed” placements, when driven by genuine behavioural signals, consistently outperform best-seller blocks. Our guide to personalised recommendations covers the algorithm choices in detail.
- Search and merchandising. Reordering results and category pages by inferred intent is often the single biggest lever, because site search is the highest-converting channel on most sites.
- On-site messaging. Delivery thresholds, returns reassurance and stock urgency tuned to the visitor’s context — a first-time visitor needs trust signals; a returning cart-holder needs a reason to finish.
What rarely earns its keep: heavily personalised copy on every page, dynamic pricing visible to the shopper (a trust hazard), and personalisation that depends on data you can’t reliably collect.
For whom — segments and signals
You don’t need one-to-one personalisation to see results. Most of the value sits in a handful of well-defined segments.
Start with intent, not identity
The most useful signals are behavioural and available in the session: entry keyword, first category viewed, search terms, device, and referral source. These tell you what someone wants right now, which is more predictive than who they are. A new visitor arriving on a product page from a comparison query needs different help from a logged-in repeat buyer browsing the homepage.
Layer in known-customer data where you have it
For logged-in or recognised visitors, past purchases, lifecycle stage and category affinity sharpen recommendations. Build this on a clean foundation — fragmented data produces confident but wrong personalisation. Our piece on eCommerce data foundations explains why this groundwork pays off before any model does.
A pragmatic segment set
A workable starting point is four or five segments: new high-intent, new browsing, returning unconverted, returning customer, and lapsed. Each gets a distinct default experience. You can refine into finer clusters later, but this set captures most of the conversion upside with manageable complexity.
Doing it without crossing the privacy line
Personalisation lives or dies on trust, and the regulatory and reputational stakes are real. A few principles keep you on the right side:
- Minimise data. Collect what you’ll actually use. Storing everything “just in case” is a liability, not an asset.
- Be transparent. If a recommendation is based on browsing, it’s fine — even reassuring — to say so. Hidden inference feels manipulative when discovered.
- Respect consent. Behavioural personalisation that relies on tracking needs a lawful basis and honoured opt-outs. Build the experience so it degrades gracefully for visitors who decline.
- Avoid sensitive inference. Never personalise on inferred health, financial distress or other sensitive categories.
Privacy-respecting personalisation isn’t only compliant — it tends to perform better, because shoppers engage more when they trust the experience. For the regulatory specifics, see GDPR and AI in eCommerce.
Measuring it honestly
This is where most programmes deceive themselves. A personalised carousel will always show high click-through — but click-through isn’t the goal, and the visitors who engage with it were often going to convert anyway.
Test against a holdout
The only credible measure is a randomised holdout: a slice of traffic that sees the non-personalised default. Compare conversion, revenue per visitor and average order value between the personalised and control groups. If the personalised group doesn’t out-earn the control on revenue per visitor, the personalisation is decoration.
Watch the right metrics
- Revenue per visitor is the headline; it captures conversion and basket value together.
- Conversion rate by segment, to see where the lift concentrates.
- Guardrails: return rate, unsubscribe rate, and complaint volume. A lift that comes with rising returns isn’t a lift.
For a fuller treatment of the headline metric, see revenue per visitor.
A rollout that works
- Fix the data. Unify identity and behaviour so signals are reliable.
- Pick two or three surfaces with high traffic and clear intent — search results, the homepage hero, the recommendation block.
- Define your segments and the default experience for each.
- Ship with a holdout from day one, never retrofitted.
- Read revenue per visitor against control after a full business cycle.
- Expand to new surfaces only once a current one proves out.
Common pitfalls
- Recommending what’s already owned or in the basket. Filter aggressively.
- Personalising on stale or sparse data, which produces confident nonsense.
- No control group, so you can never separate lift from selection effect.
- Over-fitting to clicks instead of revenue.
- Letting the experience break for opted-out visitors rather than degrading to a sensible default.
Personalisation that lifts conversion is quieter than the marketing suggests: better defaults, relevant recommendations, the right reassurance, all measured against an honest control. It’s a core part of any serious conversion optimisation effort.
If you want help deciding which surfaces are worth personalising — and proving the lift rather than assuming it — get in touch.