What Personalisation Looks Like at Scale: Lessons from 20 Years

Twenty years is a long time to be wrong about things and slowly learn from it. That's basically what e-commerce personalisation has been for me. Two decades of getting things wrong, watching what worked anyway, and trying to work out why.

I built my first e-commerce site in the late 1990s. TSD turned 23 this year. Long enough to have watched trends come and go, technologies overpromise and underdeliver, and simple approaches beat clever ones more often than I'd have guessed back then.

So before I close out this series, I wanted to step back from the tactics. Here's the long view. What actually works when you've been doing it long enough to see the patterns compound, what I got wrong, and what I'd tell myself if I could go back.

What "scale" actually means

I don't mean enterprise software scale. I mean what happens when personalisation stops being a project and becomes part of how a business runs.

Our work with Joe Davies is one of the clearest example I've got. Ten-plus years, thousands of wholesale accounts, personalisation woven into how the site works day-to-day rather than bolted on as a campaign. You can't fake your way through a decade. Either the logic was sound or it wasn't.

The composite banner approach I wrote about earlier in this series (the "Death of the Hero" post) is another one. We first designed it for Pritchards Menswear of Hereford, then built it for Chalkboards UK back in 2011, refined it through Festive Sparkle, and it's still the architecture we reach for today. Fifteen years of the same underlying idea, getting better as data and platforms caught up with it.

That's what I mean by scale. The things that stop being surprising because they keep working.

The things I got wrong

Honest reflection is the only part of this post that really matters. Plenty of agencies will tell you what they've learnt. Fewer will tell you what they got wrong, so here goes.

Early on, we overcomplicated segmentation before there was enough data to support it. I'd build ten, fifteen, twenty segments for clients whose traffic couldn't sustain three. Each one so narrow the sample size made every result noise. More segments is not the same as better personalisation, and I learnt that the slow way.

I underestimated the cold start problem in personalisation, which is why I wrote a whole post about it earlier in the series. For years I assumed personalisation would start delivering from month one. It doesn't. You need data density before the models have anything useful to say, and we weren't warning clients about that clearly enough.

For most of the first decade, I measured the wrong things. Open rates. Click-throughs. Engagement metrics that looked great in reports but didn't map cleanly to revenue. It took longer than it should have before we committed to measuring actual sales movement as the only number that earns its keep.

I also assumed clients would maintain personalisation rules once we'd set them up. They don't. Personalisation needs an owner on the client side. Someone who cares enough to keep the rules updated as the business changes. Without that person, the rules drift, the relevance decays, and the results fade. We build that expectation into every engagement now.

And one I held onto too long: the belief that more personalisation is always better. It isn't. Over-personalisation feels intrusive, especially in B2B where wholesale buyers want efficient, not cosy. Getting the dial right is half the job.

The things that consistently work

After 20 years, a small number of approaches always move the needle. Not because they're clever, but because they match how buying actually happens.

Timing. Reaching someone just before they need to reorder. Boring, reliable, and it works. The 19% conversion rate we see on reorder nudges isn't a trick. It's what happens when you get out of your own way and send the right message at the right moment.

Relevance over creativity. The right product recommendation in plain text beats a beautifully designed banner pushing the wrong product. Every time. I've watched teams spend weeks on creative and lose to a simple rule engine that just knew what the customer actually wanted.

B2B personalisation. Treating wholesale buyers as humans with preferences, not procurement systems with credit limits. The sector timing work we did (31% conversion lift when we realised different retail sectors plan their year on completely different calendars) came from the basic recognition that B2B buyers are people too.

Search data as demand intelligence. Probably the single insight that's changed more of our clients' businesses than anything else in this series. Not flashy. Genuinely useful. Your search log tells you what customers want before your stock system knows you don't have it.

And starting simpler than feels comfortable. The projects that compound are the ones where we did three things well in year one instead of thirty things badly.

Where this is heading

A bit of forward-looking, without the hype.

AI is already making personalisation faster and cheaper to implement. Genuinely useful. But the data foundation problem doesn't go away. Poor data plus AI equals faster poor decisions, delivered with more confidence. The businesses that quietly invested in clean first-party data over the last five years are about to pull further ahead. The ones hoping AI will rescue messy data are in for a harder ride than they expect.

First-party data becomes structurally more valuable as third-party cookies fade and privacy regulations tighten. If you've been building clean data habits for years, you're about to look very clever. If you haven't, the catch-up is going to hurt.

B2B personalisation is where the biggest gains still are. Most B2B sites are five to ten years behind their D2C equivalents. The buying behaviours are different, the tooling is different, but the principles transfer across. Anyone prepared to invest in this now has an open goal in front of them.

The cold start problem gets easier as platforms get smarter at inferring from limited data, but it doesn't go away. New products, new segments, and new markets will always need time before the models have anything useful to say.

And the composite banner approach will be table stakes within three to five years. Businesses implementing it now are ahead of that curve. The ones still running a single hero banner in 2029 will look the way single-page catalogue sites looked in 2015.

What we've learnt

Not a sales pitch. A statement.

We've built personalisation into platforms that barely supported it, because the logic always mattered more than the technology. We've learnt what not to build, which turns out to be worth as much to clients as what to build. Our best personalisation work is the stuff clients forget is there. It just works, quietly, year after year.

We've made mistakes. Those mistakes are worth something to clients starting fresh, because they're mistakes you don't have to make yourselves.

There's a parallel I come back to, partly because we've been renovating a 15th-century farmhouse since 2013 and it's taught me more than I expected about the rest of what I do. The quick fix and the proper job look much the same on day one. They look very different on year five. Personalisation is the same. The work that compounds over years is always the work that was set up carefully and left to do its job. Rushed personalisation ages badly. Patient personalisation just gets more useful.

Where to start

This is the end of the series, so it feels right to offer something concrete rather than a generic "get in touch."

If anything here has changed how you think about your own business, a personalisation audit is probably the most useful next step. We look at your data, your customer behaviour, and your current setup, then tell you plainly where the real opportunities are. No fluff, no generic playbook. Just a clear starting point you can actually act on.

You can get in touch at tsd.digital/contact-us. If 20+ years has taught me anything, it's that the businesses who ask the right questions early tend to pull ahead of the ones who wait to be sure.

Thanks for reading the series. Now go and do something useful with your customer data.