AI personalisation and customer insight:
how it works in practice
AI personalisation uses machine learning to decide, for each person, which message, product, offer or moment is most likely to be relevant. It learns from behaviour and purchase data rather than from rules a marketer writes by hand. That lets it personalise for thousands of people at once and adjust as their behaviour changes.
This article explains how that works in practice. It covers the data AI needs, the segmentation techniques and recommendation engines behind it, dynamic and real-time content, and how to tell whether it pays off. For a broader introduction to AI in marketing, start with the power of AI marketing.
How does AI personalisation work?
Every AI personalisation system follows the same four steps, whatever the channel:
- Collect signals. Clicks, page views, searches, purchases, email engagement and profile data are gathered per person.
- Learn patterns. A model looks for what similar people did next: which products they bought together, how often they returned, when they lost interest.
- Decide. For each person and moment, the model scores the options, such as products, content blocks or send times, and picks the most likely to work.
- Deliver and learn again. The chosen content is shown, the response is recorded, and the model updates.
The fourth step is what separates AI from rule-based personalisation. A rule stays the same until someone changes it. A model keeps learning from every response.
What data does AI personalisation need?
Mostly data you already hold. The most useful signals are first-party: what customers told you, what they did on your site and in your emails, and what they bought. Our article on first-party versus third-party data explains why first-party data has become the foundation.
- Profile data: location, language, role or company for B2B, preferences people chose themselves.
- Behaviour: pages and products viewed, searches, email clicks, time since the last visit.
- Transactions: what was bought, how often, at what value, and what was returned.
- Context: device, time of day, the campaign or channel someone arrived from.
Quality matters more than quantity. Duplicate profiles, missing purchase history or disconnected systems lead to confident but wrong recommendations. Bot activity matters too: automated email opens and clicks can make uninterested contacts look engaged, as our guide on spotting bot opens and clicks explains.
Which AI segmentation techniques are most useful?
AI segmentation groups people by what they are likely to do, not only by who they are. These are the techniques you will meet most often:
Recency, frequency and monetary value (RFM) segmentation is a useful, non-AI baseline. If a predictive model cannot beat a simple RFM split on your own data, it is not earning its keep. Our guides explain how predictive AI works and how to use it for customer lifetime value optimisation.
How do recommendation engines work?
Recommendation engines predict which products or content a person is most likely to want. Three approaches do most of the work:
- Collaborative filtering: people who bought or viewed what you did also bought these. It needs plenty of behaviour data.
- Content-based filtering: products similar in attributes, such as category, brand or price, to the ones you liked. It works with less history.
- Hybrid models: a combination of both, often with business rules on top, such as excluding out-of-stock items or favouring margin.
Recommendations increase sales because they shorten the path to the right product and surface items people would not have found. They also have known weak spots. New visitors and new products have no history- the so-called cold-start problem- so a good engine falls back on bestsellers or trending items. Recommending what someone just bought is a classic mistake, and a sign that systems aren’t sharing purchase data.
Measure recommendations on click-through from recommendation slots, the share of revenue they influence and their effect on average order value. Ideally, compare against a control group that sees bestsellers instead. Our article on searchandising shows how search and recommendations work together on site.
What is dynamic content optimisation?
Dynamic content optimisation means building an email or web page from blocks that change per person, with AI choosing which version each person sees. A hero image, an offer, a product row or a headline can each have several variants.
It differs from a classic A/B test. An A/B test splits traffic evenly, waits for a winner and then sends everyone the winner. Dynamic optimisation shifts traffic towards better-performing variants as results come in, and can pick different winners for different segments. You lose less to underperforming versions while you learn.
Keep the variants meaningfully different, and keep enough people on each to learn from. Ten near-identical headlines teach a model very little.
What is real-time personalisation?
Real-time personalisation reacts to what someone is doing now, within the same visit or session. A visitor who reads three articles about one product category sees that category featured on the homepage. A shopper who adds an item to the basket sees accessories that fit it. An email opened tonight shows tonight’s stock and prices, not this morning’s.
It needs event data that flows quickly between systems and decisions made in milliseconds. That is worth the effort on high-traffic websites and in ecommerce. For a B2B site with a long sales cycle, near-real-time updates usually deliver most of the value with far less complexity. Daily scores feeding the next email are often enough.
How does AI personalise email?
- Content: product rows, articles or offers chosen per subscriber.
- Timing: send time adjusted to when each person usually engages.
- Audience: propensity scores decide who receives a campaign at all, or who gets a softer version without a discount.
- Frequency: send fewer emails to people showing signs of fatigue, before they unsubscribe.
Our guide how we turned simple data into an AI-powered email shows a worked example.
How do you personalise across channels and the customer journey?
By deciding at the customer level, not the channel level. A unified profile lets AI choose the next best action across email, website, WhatsApp and ads. It can send the replenishment reminder by email, show the accessory on site and stop the retargeting ad for the product someone just bought.
That requires one view of each customer, which is what a customer data platform provides. Without it, each channel personalises on its own partial data, and customers notice the gaps. Our guide on enhancing the customer journey with data-driven personalisation covers the journey side, and our article on hyper-personalisation in retail covers the retail benefits.
How do you know AI personalisation is working?
Compare it with what would have happened without it. Keep a small, random control group that receives the non-personalised version, and measure the difference in conversions, revenue per recipient or retention. That difference, the incremental lift, is the honest measure.
Look at results per segment as well as overall. Personalisation can lift results for engaged customers while doing nothing for new ones. And give models time: their value builds over several campaigns as they learn.
Where does AI personalisation go wrong?
- It feels intrusive. Personalisation that reveals more than people expect you to know backfires. Our guide to the personalisation paradox explores how to strike the right balance.
- It narrows too far. Showing people only what they’ve already bought hides the rest of your range. Build in some discovery.
- It learns from bad data. Wrong or incomplete data produces wrong recommendations, at scale.
- It forgets context. A gift purchase is not a lasting preference. Let customers correct what you think you know.
- It ignores consent and transparency. You must explain profiling for marketing in your privacy notice, and people can object to it.
Frequently asked questions
How much data do you need to start?
Less than you might think for simple uses. Recommendations based on product similarity and bestsellers work from day one. Predictive scores such as churn or lifetime value need enough purchase history to learn from, typically at least several months of transactions.
Is AI personalisation allowed under GDPR?
Yes, with conditions. Marketing personalisation is profiling, so you need a lawful basis and must explain it clearly. Under GDPR, people have an absolute right to object to profiling for direct marketing, and you must stop when they do.
AI personalisation with Spotler
Spotler Activate brings customer data into one profile, so every channel personalises on the same view. Its predictive AI add-on groups customers by predicted lifetime value in dynamic segments. Search and merchandising help shoppers find products faster with smart recommendations, and email marketing automation turns segments and product data into personalised emails. Our article on customer loyalty and AI shows where it leads. Book a demo to see it with your own data