EU AI Act transparency duties apply now; high-risk duties from December 2027. Check your exposure
Insights About us Careers
Contact us
AI for Marketing and Growth

Personalisation Engine Implementation

Deciding what each visitor sees, in real time, and proving through a controlled holdout that it earned more than showing everyone the same thing — which is not a given.

10 to 18 weeks
Typical build
Fixed scope
Commercial model
Holdout
Always retained

Personalisation is assumed to work, which is why so few programmes measure it against a proper holdout. When one is retained, the results are frequently smaller than expected and occasionally negative, because a badly personalised experience is worse than a well-designed generic one.

In one paragraph

A personalisation engine decides which content, offer, product or message to show a given visitor, based on their behaviour, attributes and context, in real time — with a control group held back so the incremental effect can be measured rather than assumed.

The problems that decide whether it works

Cold start, which is most of your traffic

A large share of visitors are new or unrecognised, and they are exactly the audience an untuned system serves worst. Contextual signals — entry page, referrer, device, time, campaign — carry the first visit, and the personalised experience improves as evidence accumulates rather than starting from nothing.

The narrowing problem

Optimising purely for immediate engagement narrows what a person is shown until the experience becomes claustrophobic and they disengage. Deliberate exploration is built in, and it costs a measurable amount of short-term performance to protect the longer-term one.

Latency, because personalisation happens in the render path

A decision that takes 300ms in a page load has to earn that delay. Where it cannot, the answer is precomputation and caching rather than a slower page, since the speed cost frequently exceeds the personalisation gain.

What may be used, under what basis, and what happens for visitors who decline. A system that assumes full tracking and degrades badly without it will underperform across a large share of European traffic. See data privacy for AI.

The creepiness boundary

Some accurate personalisation is unwelcome — inferring sensitive circumstances, referencing behaviour the visitor did not realise was tracked, or being right in a way that feels like surveillance. We agree the boundaries explicitly, including inferences you will not make, before the system can make them.

Worth knowing

Keep a permanent holdout, not just a launch test

A share of traffic never receives personalisation, permanently. It is the only way to know what the programme is worth a year later, when the initial test is long forgotten and everyone assumes the uplift is still there. The cost of the holdout is small; the cost of not knowing is a programme nobody can justify or defend.

What gets personalised, and what should not

SurfaceTypical decisionNote
Homepage and landingWhich message and proof to lead withHighest impact; also highest risk of getting it wrong
Product discoveryWhich items to surface and in what orderSee recommendation engines
Email and lifecycleContent, timing and frequencyFrequency is often the biggest lever, and is usually ignored
Offers and pricing displayWhich promotion to showDifferential pricing raises fairness and legal questions
Search resultsRanking adjusted to the personMust not override explicit intent
Support and self-serviceWhich answers to surface firstImproves resolution without personalising the answer itself

The pricing row needs care. Personalising which promotion is shown is ordinary practice; varying the price itself by inferred willingness to pay is a commercial, legal and reputational decision that belongs with your executive and legal teams rather than inside a model.

How we build it

  1. Define the outcome the system optimises for — and be honest that optimising engagement, conversion and lifetime value produce different systems.
  2. Establish the baseline. Current performance of the generic experience, which is what everything is measured against.
  3. Design the signal set within consent constraints, including what is available for unrecognised visitors.
  4. Build decisioning to the latency budget, precomputing where real-time inference cannot meet it.
  5. Retain a permanent holdout and report incremental effect rather than the performance of the personalised group alone.
  6. Monitor for narrowing and for segment-level harm, because aggregate uplift can conceal a segment whose experience got worse.
Process

How the engagement runs

A holdout is designed in from the start, because retrofitting one is rarely done.

Weeks 1 to 3

Baseline and design

Generic experience baselined; outcome, signals, consent constraints and creepiness boundaries agreed.

Weeks 4 to 6

Data and infrastructure

Signal collection, profile assembly and the decisioning path built to the latency budget.

Weeks 7 to 12

Model and decisioning

Models built with cold-start handling and deliberate exploration; integration into the surfaces.

Weeks 13 to 16

Controlled rollout

Launched with a holdout, incremental effect measured, segment-level outcomes checked.

Weeks 17 to 18

Handover

Monitoring, retraining, holdout reporting and the boundaries documented for the owning team.

Deliverables

What you receive

A decisioning system with a permanent holdout, so its value can still be stated next year.

01

Personalisation engine

Real-time decisioning within the latency budget, with precomputation where needed.

02

Cold-start handling

Contextual decisioning for new and unrecognised visitors, who are most of your traffic.

03

Consent-aware design

Defined behaviour for each consent state, degrading sensibly rather than failing.

04

Permanent holdout and measurement

Incremental effect reported, not the personalised group's performance alone.

05

Boundary documentation

Inferences the system will not make, agreed before it could make them.

06

Monitoring

Narrowing, segment-level outcomes and model health, with retraining and alerting.

Fit check

Is this the right engagement?

Worth being direct. Personalisation Engine Implementation is the wrong spend in some situations, and those are listed rather than buried.

Good fit if

  • Traffic is high enough for controlled experiments to reach significance.
  • Content or product range is wide enough that choosing matters.
  • Behavioural data exists and is reliable.
  • A generic experience is known to underserve distinct segments.
  • An existing personalisation programme has never been measured against a holdout.

Choose something else if

  • Traffic is too low for experiments to conclude anything.
  • There is little to choose between, so personalisation has nothing to decide.
  • The requirement is product recommendations. See recommendation engines.
  • Consent constraints leave no lawful signal to personalise on.
Questions

Frequently asked questions

Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.

Does personalisation actually increase revenue?

Sometimes substantially, sometimes marginally, and occasionally not at all — which is why we retain a permanent holdout rather than assuming. A badly personalised experience is worse than a well-designed generic one, and programmes that never measured against a control usually cannot say what they are worth.

What about visitors we know nothing about?

They are most of your traffic and they are where untuned systems perform worst. Contextual signals — entry page, referrer, device, campaign, time — carry the first visit, and the experience becomes more tailored as evidence accumulates. A system designed only for known visitors will disappoint on the majority of sessions.

How do we avoid being creepy?

By agreeing the boundaries before the system can cross them: which inferences you will not make, which behaviour you will not reference back to the visitor, and which categories are off limits entirely. Accuracy is not the test — some correct personalisation is unwelcome, and that is a judgement your brand makes rather than a model.

Can we personalise under GDPR?

Yes, within consent and lawful basis constraints, and the system has to be designed for that from the start. What fails is a design assuming full tracking that degrades badly without it, since a large share of European traffic will not grant it. See data privacy for AI.

Should we personalise prices?

That is a commercial, legal and reputational decision rather than a technical one, and it belongs with your executives and counsel. Personalising which promotion is shown is ordinary practice. Varying the price itself by inferred willingness to pay carries fairness, regulatory and trust risks that we will set out rather than quietly implement.

Is this the right engagement?

Tell us what you are trying to build. If a different service fits better, or if you do not need us at all, we will say so.