Personalization: why most of it doesn’t work

Personalization has a branding problem. The word conjures either something magical or something creepy, and the reality is neither. Most of it is unremarkable, some of it is genuinely useful, and a surprising amount of it is quietly worse than showing everybody the same thing.

The plain definition: personalization means changing what someone sees based on something you know about them. That’s the whole concept. A returning visitor sees different content to a first-timer. A customer in Manchester sees different stock to one in Dubai. Someone who bought a printer gets shown ink rather than printers.

The reason it fails so often isn’t the technology, which mostly works. It’s that teams start from the tool rather than from a reason, then measure whether it ran instead of whether it helped. This piece is about doing it the other way round.

What it actually means

Every piece of personalization has three parts, and if any one is missing the whole thing collapses.

There’s a signal, which is the thing you know. Where someone is, what they’ve browsed, whether they’re a customer, which campaign brought them in, what they’ve bought before, whether it’s their first visit or their tenth.

There’s a rule or a model, which decides what to do with that signal. It can be as simple as “if they’ve bought before, hide the new-customer discount.” It can be a model that ranks products by predicted relevance. Simple rules cover far more ground than people expect.

And there’s a change, which is what the visitor actually sees. Different headline, different products, different order of results, a different call to action, or sometimes nothing at all, because deciding not to show someone an irrelevant banner is personalization too.

Worth separating from two neighbours it gets muddled with. Segmentation is grouping people; personalization is what you do with those groups, so segmentation is an input rather than the thing itself. And customisation is when the user chooses their own settings, which is a different beast entirely: they’re in control, not you.

The ladder, from easy to hard

Personalization isn’t one capability, it’s a ladder, and most teams should climb it in order rather than jumping to the top rung.

The first rung is context. You change things based on what’s observable right now with no history required: device, location, language, time of day, the referring source. Cheap, immediate, and often the highest return per hour spent. Showing local currency and stock is not glamorous and it reliably works.

The second rung is behaviour in the session. What has this person done in the last few minutes? Viewed three items in one category, searched for something specific, put something in the basket and stalled. You still need no long-term profile, just attention to the current visit. This rung is badly underused.

The third rung is known attributes. Now you need a profile that persists: are they a customer, what did they buy, what plan are they on, which industry are they in. This is where you start needing your data joined up, and where a CDP or similar becomes relevant.

The fourth rung is predicted. Models rather than rules. Recommendation engines, propensity scores, next-best-action. Powerful when you have volume, and mostly disappointing when you don’t, because models need data to learn from and a low-traffic site never gives them enough.

Most of the value for most organisations sits on the first three rungs. The fourth gets all the conference talks. I’d push back gently on any plan that starts there.

What good and bad look like

Same scenario, two versions, because the difference is easier to feel than to define.

Someone buys a laptop from you on Tuesday.

The bad version: for the next three weeks your site and emails show them laptops. The same laptop, sometimes. The system saw an interest signal and kept firing it, and nobody asked whether interest survives purchase. This is the single most common personalization failure and everyone reading this has experienced it.

The good version: after Tuesday they see accessories, a setup guide, and an extended warranty offer. The purchase changed their state, and the experience changed with it. If they come back in two years, laptops become relevant again.

Notice the good version needed no machine learning. It needed someone to think about what a purchase means, then write a rule. The bad version probably came from a more sophisticated system that nobody had told about purchases.

Second illustration, quieter but more valuable. An enterprise buyer arrives from a search for compliance documentation. The bad version shows them the standard small-business pricing hero. The good version notices the entry point and leads with security and compliance material. No profile needed, just attention to what brought them there.

How to actually start

Six steps, in order, and the order genuinely matters because most failures come from skipping to step four.

Start with a business problem, not a capability. “Our checkout drop-off is high” or “returning customers see the same generic homepage” are problems. “We want to use personalization” is not. Write the problem down before anyone opens a tool.

Then find the signal you already have. Look at what you can observe today without new infrastructure. Referral source, device, session behaviour, purchase history if it’s accessible. Most teams have more usable signal than they think and go shopping for data they don’t need.

Next, write the rule in plain English. If someone has bought in the last thirty days, hide acquisition offers and show support content. If you can’t express it in a sentence, it’s not ready to build, and a rule nobody can explain is a rule nobody will maintain.

Then decide how you’ll know it worked. Pick the measure before you launch. Conversion rate, add-to-basket rate, return visits, whatever fits the problem. Crucially, decide what would make you turn it off. Personalization that nobody evaluates just accumulates.

Now test it against not doing it. This is the step that gets skipped and it’s the one that matters most. Run the personalized experience against the generic one and compare. Plenty of personalization performs worse than the control, and you will never find out if you don’t check.

Finally, keep it or kill it, and be willing to kill it. Rules that stop earning their place should be retired. A site carrying forty rules nobody remembers writing is slower, harder to change, and produces stranger experiences than a site with six good ones.

A planning worksheet

The six steps are simple to describe and easy to skip under pressure, particularly the measurement ones, which is precisely why they’re worth writing down.

We’ve turned them into a short planning worksheet: space for the business problem, the signal you’ll use, the rule in plain English, the change the visitor sees, the measure you’ll judge it by, and the condition that would make you switch it off. There’s also a small register for tracking which rules are live, when they were last reviewed, and what they’re currently doing to your numbers.

Fill one in per personalization you’re considering. If you can’t complete the measurement rows, that’s a useful signal in itself.

Where it actually pays off

If you’re deciding where to spend the effort, some places reward personalization far more than others, and the popular ones aren’t always the productive ones.

Search and listing pages are usually the best return and the most overlooked. When someone lands on a category with four hundred items, the order you show them in is enormous. Putting relevant things near the top is personalization, even though it rarely gets called that, and it acts on people who have already shown you what they want.

Post-purchase is the second big one, and it’s mostly ignored because teams are focused on acquisition. What someone sees after buying determines whether they come back, whether they need support, and whether they buy again. Changing state after purchase is cheap to implement and fixes the most visible failure mode.

Returning-visitor experiences pay well because you know something. A second visit is a strong signal on its own. Picking up where they left off, surfacing what they looked at, or simply not repeating the introductory pitch all help.

Empty and edge states are the quiet win. No search results, an empty basket, a first login with nothing configured. These moments are usually generic and usually the point where people give up. Making them contextual is often a few hours of work.

Now the places that reward less than people expect. Homepage heroes attract the most attention and move the least, because your homepage is often not where the decision happens. Broad demographic targeting tends to underperform behavioural signals, since what someone just did predicts far better than what category they belong to. And personalizing your way around a genuinely confusing interface never works. Fix the confusion.

The pattern across all of this: personalization performs best close to the point of decision and close to a fresh signal. The further you drift from either, the thinner the return.

Why most personalization underperforms

Five reasons, and technology is barely one of them.

Not enough traffic to learn anything. Split a modest audience across six variants and none of the results mean anything. Personalization needs volume before it can be evaluated, and evaluating it is the only thing that keeps it honest.

Stale signals. The laptop problem. Behaviour is treated as permanent when it’s often temporary, and states change. Purchases, cancellations, and support complaints should all change what someone sees, and frequently don’t.

Optimising the wrong thing. Personalizing a homepage hero feels satisfying and often moves nothing, because your customers’ real friction is three steps deeper. Look at where people actually drop off before deciding what to personalize.

Nobody owns it. Rules get created during a project, the project ends, and two years later nobody knows why an odd banner appears for Belgian visitors on Tuesdays. Someone needs a standing responsibility to review and retire.

The creepiness line. Using data people didn’t realise you had produces a reaction that no conversion lift compensates for. A rough test: would the person be comfortable if you explained the rule to them out loud? “We showed you this because you looked at it earlier” is fine. Anything you’d rather not say plainly is worth reconsidering, both because trust is expensive to rebuild and because privacy rules in most markets are moving in one direction.

I’d add one more thing that isn’t a failure exactly. Personalization has a ceiling. It rearranges and filters what you already have. If the underlying content is thin or the product isn’t right for the visitor, no amount of clever targeting rescues it. It’s an amplifier, and amplifiers work on whatever you feed them.

The stuff worth remembering

Personalization means changing what someone sees based on something you know about them. Signal, rule, change. That’s the whole idea.
Climb the ladder in order: context, then in-session behaviour, then known attributes, then predictive models. Most of the value sits on the first three.
Simple rules beat sophisticated models at most companies, because rules can be explained, maintained, and corrected.
Start from a business problem, write the rule in plain English, decide the measure before you launch, and always test against not doing it.
Be prepared to switch things off. Unreviewed rules accumulate and quietly degrade the experience.
If you would not comfortably explain the rule to the person it affects, do not ship it.
Personalization amplifies what you already have. It does not fix weak content or a bad fit.

Wondering where to start?

Personalization has a few close cousins worth reading: website personalization, hyper-personalization, and the warning signs of bad personalization.

The useful version of this conversation begins with where people actually drop off, not with what a platform can do. If you would like a second opinion on which one or two changes are worth making first, and which fashionable ones to skip, get in touch and we will work through it with you.