AI Services
AI Personalisation Engines
Works well on large catalogues with real traffic. On a small site it is expensive machinery producing recommendations a curated list would beat.
Personalisation adapts what a visitor sees based on their behaviour — recommended products, relevant content, tailored messaging. The Nexclick builds it where you have enough traffic for the patterns to be real, and stops short of the version customers find unsettling rather than helpful.
Is this you?
What usually prompts the call
- Your catalogue is large enough that customers cannot find what suits them.
- Every visitor sees the same homepage regardless of what they have looked at.
- You have recommendation widgets that suggest the item someone just bought.
- Your email sends the same content to segments that behave completely differently.
What we do
The actual deliverables
Things that appear on an invoice, not adjectives.
- Check you have the traffic for it
- Personalisation learns from behaviour. Below meaningful volume the patterns are noise, and a curated selection chosen by someone who knows the catalogue will outperform it.
- Start with rules, add learning later
- Rule-based personalisation — recently viewed, category affinity, returning customer — delivers most of the value and is explainable. Learned models come after the rules are working.
- Recommendation logic that makes sense
- Complementary rather than identical products, and never the item just purchased. Most recommendation failures are logic failures rather than model failures.
- On-site content adaptation
- Homepage, landing page and category content adjusted to segment or behaviour, with a sensible default for everyone the system knows nothing about.
- Email and lifecycle personalisation
- Content driven by actual behaviour rather than a broadcast with a first name inserted. Usually the highest-return personalisation available.
- A deliberate creepiness line
- What you can infer and what you should act on are different questions. Referencing something a customer did not tell you produces unease, and the line is agreed rather than discovered.
- Measured against a control
- A holdout group seeing no personalisation, so the uplift is evidenced. Personalisation is frequently assumed to work and rarely measured.
Comparison
Personalisation that helps, and personalisation that unsettles
The difference is whether the customer would be comfortable if you explained how you knew. Use that test on anything you are considering — it is more reliable than a rule about what data is technically permitted.
| Tactic | Verdict | Why |
|---|---|---|
| Recently viewed items | Helps | Obvious to the customer; they did it themselves |
| Complementary product suggestions | Helps | Genuinely useful, clearly derived from the basket |
| Content by category the visitor browsed | Helps | Relevant, and easy to explain |
| Returning customer greeting | Helps | Expected on any account-based site |
| Reordering a previous purchase | Helps | Convenient, obviously derived from their history |
| Location-based delivery information | Helps | Practical, and clearly inferred from IP |
| Referencing a page they viewed but did not buy | Borderline | Fine in email, unsettling on-site |
| Inferring life events from browsing | Unsettles | Frequently wrong, and disturbing when right |
| Referencing behaviour on another site | Unsettles | Customers do not expect cross-site tracking |
| Different prices for different visitors | Damages | Detectable, and it destroys trust when found |
| Fabricated scarcity personalised to a visitor | Never | Misleading under consumer protection rules |
How it works
Step by step, with timeframes
Timeframes are typical rather than guaranteed, and they assume we get account access and approvals when we ask.
- 01Week 1–2
Assess volume and data
Whether you have enough traffic and behavioural data for this to produce real patterns. An honest gate.
- 02Week 2–5
Rules first
Rule-based personalisation deployed with a holdout control, so the baseline uplift is known before anything more complex is added.
- 03Month 2–4
Add learning where it earns it
Behavioural models introduced only where they beat the rules against the control group.
- 04Monthly
Measure and refine
Uplift against the holdout, reported honestly including where personalisation performed worse.
What you get
Reporting and ownership
- An honest assessment of whether your traffic supports personalisation at all.
- A holdout control group from day one, so uplift is measured rather than assumed.
- Rule-based personalisation first, which is explainable and usually delivers most of the value.
- A written position on what you will and will not infer publicly about a customer.
- Monthly uplift reporting, including where personalisation underperformed the control.
Tools and platforms
- Your ecommerce platform or CMS
- GA4 and behavioural data
- Recommendation engines or custom models
- Email platform segmentation
- Holdout group testing
Timeline
How long this actually takes
Five weeks to first deployment, then measurement over two to three months. Two honest positions. Below meaningful traffic, personalisation is expensive machinery producing worse results than a curated list — the volume gate is real and we apply it. And personalisation is frequently assumed to work rather than measured; a holdout group costs nothing and regularly shows that a particular tactic performed worse than showing everyone the same thing. We build the control in from the start for that reason.
Pricing model
Fixed-price project
Fixed price for rules-based implementation with a control group. Learned models quoted separately, and only recommended once the rules have proved the baseline.
Questions
AI Personalisation Engines questions
How much traffic do we need for this to work?
Enough that behavioural patterns are real rather than noise — which in practice means a substantial catalogue and consistent traffic. Below that, a curated selection chosen by someone who knows the products will outperform any engine, and we will say so rather than build one.
Does personalisation actually increase revenue?
Sometimes, and less reliably than the category assumes. That is why we build a holdout control from the start. Measured honestly, some personalisation tactics underperform showing everyone the same thing, and you only find that out if you measured.
Where is the line on being too personal?
A useful test: would the customer be comfortable if you explained how you knew? Recently viewed passes easily. Inferring a life event from browsing does not — and it is frequently wrong, which is worse. The line gets agreed in writing rather than discovered through complaints.
Do we need consent for this?
It depends on the data and the mechanism. Behavioural personalisation using cookies generally requires consent under PECR, and profiling personal data carries UK GDPR obligations. It gets scoped explicitly rather than assumed, and it is documented in the governance pack.
Can we personalise with a small product range?
Rules-based, yes — returning visitors, recently viewed, complementary items. A learned recommendation engine has too little to work with. The rules deliver most of the value on small catalogues anyway, and they are explainable when someone asks why something was shown.
What if the recommendations are bad?
Usually a logic problem rather than a model one — recommending the item just purchased, or near-identical alternatives. Those are fixed with rules rather than more sophisticated learning, which is a large part of why we start with rules.
Last reviewed 28 July 2026.
Tell us what you are trying to fix
A 20-minute call, no pitch deck. The Nexclick will tell you what we would do, roughly what it costs, and whether we are the right people for it.