Which kinds of AI personalisation pay off for a growing business, which rarely do, and how to keep tailored marketing helpful rather than intrusive.
For two decades, marketing software has promised a website tailored to every visitor, emails that reflect what each reader has done, and ads that speak to each person's exact problem. The promise outran reality, because every tailored version had to be written, designed and maintained by someone. AI has cut the cost of producing those versions, and that changes the calculation.
The practical reality is uneven. AI personalisation pays off today in email and in product recommendations, where there is plenty of behaviour to learn from and a clear measure of success. Website personalisation can work but is harder than the demos suggest, and tailoring based on who someone is, rather than what they have done, often feels intrusive and sits uneasily with European data protection rules.
Getting it wrong costs money in two ways: tools and content that never pay for themselves, and customers who feel watched and leave. This article sets out where personalisation earns its cost, what each type needs before it works, how to prove it is helping, and where the line between helpful and intrusive runs.
The most dependable personalisation is also the oldest: emails triggered by what someone has done. A welcome series for new subscribers, a reminder when a basket is abandoned, a follow-up after a first purchase and a nudge when a product is likely to be running out all arrive when the person is already thinking about you. In business-to-business marketing, the equivalents are a relevant case study after someone downloads your price list, or a check-in when a promising lead goes quiet.
AI adds a layer on top of these flows. It can write subject line variants, choose a send time for each subscriber, pick the products shown in each email and adjust the copy to past purchases. These features are now common inside email platforms, so the real question is rarely whether you can use them, but whether your basic triggered emails are in place first.
One caution about measurement. Apple's Mail Privacy Protection automatically loads emails, including the tracking pixel, for many Apple Mail users, so plenty of recorded opens never involved a person reading anything. Any AI feature that optimises for opens, including subject line testing and send-time optimisation, is partly learning from noise, so judge email personalisation by clicks, orders, replies and unsubscribes instead.
For online shops, recommendations are the other proven area. Models that suggest products based on what similar customers bought, and newer ones that understand product descriptions well enough to suggest a matching item with no sales history, can raise the value of each order. They need data to learn from, though, and many smaller shops do not have enough of it.
A useful decision rule: if you sell 40 products and take 300 orders a month, hand-picked rules will usually do as well as AI and cost far less. Someone who buys an espresso machine should see descaler and a milk jug, and you can set that up in an afternoon. AI recommendations start to earn their place with large catalogues, frequent repeat purchases and thousands of orders, where no person could keep the rules up to date.
Measure recommendations by revenue per visitor and average order value across the whole shop, compared with a group of visitors who do not see them. Clicks on the recommendation block prove very little, because people often click on things they were going to buy anyway.
Showing different website content to different visitors has the most attractive promise and the hardest path. In a demo, a visitor from the food industry sees food-industry case studies and a retailer sees retail ones. In practice, four things need to be in place first.
To see why traffic matters, say a landing page gets 2,000 visitors a month and 3 per cent of them enquire, which gives you 60 enquiries. Split it into five personalised versions and each one receives about 400 visitors and 12 enquiries a month. At those numbers you would need many months to tell a real improvement from chance, and your market will have moved on by then.
The cheapest and most dependable form of website personalisation needs no data about the person at all: matching the landing page to the ad. If an ad promises faster delivery for restaurants, the page it leads to should open with faster delivery for restaurants, not your general homepage. One focused page per campaign theme is personalisation at the level of the campaign, and it is usually the first change worth making.
Personalisation has a threshold. Below it, tailoring feels like good service; above it, it feels like being watched. 'Welcome back, here is the jacket you looked at' is fine for a logged-in customer, but 'Hello Anna from Müller GmbH, here is pricing for your 45-person team' on a first visit is not, even if the guess is right.
The safer approach is to tailor on behaviour rather than identity. What someone has read, clicked or bought on your own site feels like relevance, because the visitor knows they did it. Who they are, inferred from outside data or from their IP address (the number that identifies their internet connection), feels like surveillance.
The law points the same way. Under the GDPR, the EU's General Data Protection Regulation, profiling people needs a lawful basis and clear information, and EU cookie rules require consent before non-essential cookies are used, whichever kind of tailoring they support. Tailoring based on the ad someone clicked, or for a logged-in customer who has agreed to it, is far easier to justify than a guess about a stranger's identity.
Two further lines are fixed by law rather than taste. If you show different prices to different people based on automated decisions about them, EU consumer law requires you to tell them the price has been personalised. Inferences about health, such as a supplement shop guessing that a visitor is pregnant, can count as health data under the GDPR, which carries much stricter rules.
Before launching any personalisation, apply one test: would the customer be comfortable if they saw exactly how you decided what to show them? Personalisation is only worth having when the customer feels understood, not watched. That is an editorial judgement as much as a data question, and it belongs with whoever looks after your brand.
Personalisation is only worth having when the customer feels understood, not watched.
Personalisation always changes results; the question is whether it improves them. The only reliable way to know is a holdout group: a share of visitors or subscribers, say 10 to 20 per cent, who keep seeing the standard version. Compare enquiries, orders or revenue per person between the two groups over a period you fix in advance.
Count the costs as well as the gains. Personalisation tools, the extra content and the time to maintain it all cost money every month, and they have to be paid for out of margin, not revenue.
Say your average order brings €25 of margin after product and delivery costs. A personalisation tool at €1,000 a month needs to add 40 extra orders a month just to break even, before anyone's time is counted. If your holdout group shows 15 extra orders, the tool is losing money, however impressive its dashboard looks.
Watch the warning signs as well: more unsubscribes, more complaints, or customers remarking that your emails seem to know too much. These are early signals that you have crossed the threshold, and they are far cheaper to act on than a fall in sales.