AI Marketing Personalisation: Recommendations, Content and Customer Data

An online shop recommending accessories for a recent purchase is using a different signal from a shop recommending products after one browsing session. AI-powered personalisation uses patterns in customer activity to rank these options, but a recommendation still needs to match available stock, the customer’s situation and the business’s permission to use the underlying data.

A purchase also leaves room for uncertainty. Someone buying a gift has supplied evidence of a transaction rather than a reliable description of their own interests. Repeatedly promoting similar products to the same person risks turning a useful suggestion into an irrelevant reminder.

What AI-Powered Personalisation Means in Practice

AI-powered personalisation uses models to select or adapt content for a person or audience based on available information. Inputs include purchase history, recent interactions and preferences supplied directly by the customer. Outputs range from a ranked product list to an email featuring a particular category.

Rules, Predictions and Generated Content

A fixed rule such as showing delivery information for a selected country does not require AI. A predictive model estimates an outcome, such as the likelihood of someone interacting with a product. Generative AI produces material such as text or images.

These functions serve different purposes. A recommendation model ranks products. A language model writes a description of those products. Neither replaces the checks confirming the products are available or the description is accurate.

For example, a marketer testing personalised email content should decide separately which products qualify for inclusion and how those products are described, because an accurate recommendation loses value if the accompanying message invents a discount or delivery promise.

Personalisation Beyond Demographic Segments

Demographic segments group people using characteristics such as age or location. Behavioural segments use actions such as repeat purchasing or browsing a category. Individual recommendations go further by ranking options for a particular user rather than giving everyone in a segment the same selection.

A more detailed model does not automatically produce a better campaign. A simple category preference supplied by the customer provides a clearer starting point than a complicated profile built from ambiguous clicks. Compare the two approaches before paying for additional modelling and integration.

From Customer Data to a Marketing Decision

Separate Observations From Assumptions

An order record confirms a purchase. A page view confirms an interaction with a page. Neither proves satisfaction, future intent or the reason behind the visit.

Keep these distinctions visible in the data used for personalisation. For a clothing retailer, repeated visits to a size guide suggest a different question from repeated purchases in the same category. The first example supports testing more prominent sizing information. The second supports testing related product recommendations.

When a model assigns someone to a high-interest audience after several visits, inspect the underlying events before launching a campaign because duplicate tracking or an unresolved support problem creates activity without demonstrating an intention to buy.

Prediction Is an Estimate

Predictive analytics uses patterns in existing data to estimate future outcomes. A propensity score ranks someone as more or less likely to take a defined action. The score does not confirm the person wants a message or prove the message will change their behaviour.

This matters when choosing whom to contact. A customer already likely to buy sometimes purchases without another email or discount. Compare contacted customers with a suitable control group to measure the additional effect of the campaign rather than crediting every subsequent order to personalisation.

New Visitors Need a Fallback

A new visitor has little individual history. Prepare useful defaults such as popular products, the category currently being viewed or a short preference selector. Keep the website usable without requiring a detailed profile.

Amazon Personalize illustrates the distinction between new and known users. Its documentation describes personalised recommendations for new users drawing on existing users’ early interaction histories, with the new user’s own interactions providing further evidence as activity accumulates.

Five Ways to Apply AI Personalisation

1. Rank Product and Content Recommendations

A recommendation system ranks eligible products or articles for a user. Apply business restrictions before displaying the selection, including availability, service area and any relevant product compatibility rules.

For example, a retailer selling cameras should check accessory compatibility rather than assuming every popular lens suits every camera owner. Amazon Personalize supports filters for controlling recommended items, including exclusions based on previous purchases. The appropriate exclusion depends on the product: a durable purchase and a regularly replenished item need different treatment.

Keep some routes outside personalised recommendations. Category navigation and site search let visitors explore beyond their previous activity instead of repeatedly seeing a narrow selection.

2. Adapt Email and Website Content

Dynamic content selects different material for different recipients or visitors. An email template contains an approved block for first-time buyers and another for returning customers, but selecting between two fixed blocks does not by itself require AI.

An AI model adds a prediction or ranking step. For example, the model ranks product categories while an approved template controls the wording. Generative AI adds another option by drafting descriptions, subject lines or explanations.

Review generated claims against product records before publication. NIST identifies confidently presented false content as a risk of generative AI. Limit promotional messages to verified prices, approved product details and current terms rather than asking a model to fill gaps in missing information.

3. Build Behavioural Audiences

Behavioural segmentation groups customers by recorded activity. Examples include repeat purchasers, recent category browsers or customers approaching a normal replenishment interval. Machine learning adds value only if the resulting groups support a useful decision beyond an existing rule.

Define what changes for each group. A replenishment audience needs an appropriate reminder period. Someone already completing a purchase needs removal from the corresponding acquisition sequence. A model-generated segment with no distinct message or action adds reporting complexity without improving the campaign.

4. Support Customers With Relevant Information

Customer service personalisation uses available context to answer a specific question. For example, an authenticated customer asking about an order needs accurate order information rather than a generic sales recommendation.

A chatbot responding to a question is not automatically predictive customer service. Proactive support needs a relevant trigger, such as a recorded delivery exception. Separate service messages from promotional recommendations so a support interaction does not become an unwanted sales conversation.

Provide a clear route to a person when the system lacks the required information. Do not allow a generated answer to invent a refund decision, product specification or account update.

5. Respond to Activity During a Visit

Real-time personalisation updates a selection using recent interactions. For example, a visitor moving from general outdoor equipment to camping products supplies a recent signal worth testing against an older category preference.

The update speed depends on the implementation. Amazon Personalize distinguishes supported real-time interaction updates from bulk data imports requiring subsequent training. Asking a supplier about event processing and retraining gives a more useful answer than accepting an undefined claim of instant personalisation.

If the website fails to record a completed purchase promptly, a recommendation system still relying on earlier browsing activity risks promoting the same item again even though the customer’s most recent action calls for a different response.

Measure Sales Effects Without Assuming Success

Personalised campaigns need a comparison with a credible alternative. Higher click-through rates alone do not demonstrate additional profit, greater satisfaction or stronger loyalty.

Define the Expected Benefit

1. Conversion improvement: Compare completed purchases or qualified leads between personalised and standard experiences. Keep the offer consistent when testing the effect of product selection.
2. Return on investment: Include software, integration and review costs alongside additional contribution from sales. Account for discounts, returns and fulfilment costs rather than treating extra revenue as profit.
3. Customer loyalty: Track repeat purchases over a relevant buying cycle. Include unsubscribe rates and complaints so repeated contact does not appear successful solely because a small group buys again.

Use a Control Group

Randomly assign eligible customers to personalised and comparison experiences where practical. Define eligibility, the main success measure and the test duration before launch. The comparison group should receive a sensible standard experience rather than a deliberately weak alternative.

For example, comparing a personalised product email with no email measures the combined effect of sending a message and selecting its contents, while comparing otherwise similar personalised and standard emails provides a clearer test of the selection method.

Keep customers in a consistent test group during the experiment. Moving people between experiences complicates interpretation. Allow enough time for the normal purchase cycle and returns process before judging a revenue-based result.

Separate Monitoring From Automatic Changes

A dashboard reporting current results does not automatically optimise a campaign. Identify the specific decision the system changes, such as recommendation ranking, message timing or audience inclusion.

Avoid changing several major elements during the same test. If the offer, audience and landing page change together, the resulting performance difference provides little evidence about the contribution of personalisation itself.

Privacy, Consent and Customer Control

Identify the Legal Basis for Each Use

Under the GDPR, processing personal data requires an appropriate lawful basis. Consent is one of the available bases rather than a universal requirement for every processing activity. A business still needs to assess the proposed use and meet the associated obligations.

Separate the activities involved: collecting website interactions, building profiles, sharing data with suppliers and sending marketing messages. Permission for one activity does not automatically cover every later use of the same information.

Tracking and Electronic Marketing Have Separate Rules

In Ireland, the Data Protection Commission identifies targeted advertising and conversion-tracking cookies as requiring consent. Strictly necessary cookies and cookies used solely for transmitting communications have limited exemptions. Do not treat an AI feature as a reason to bypass these rules.

Electronic direct marketing generally requires affirmative consent. A limited existing-customer exception applies under specific conditions, including marketing your own similar products or services, offering opt-outs at collection and in messages and meeting the applicable timing requirement.

Assess the full conditions before relying on the exception. A visitor browsing a product without purchasing does not qualify as an existing customer for this purpose merely because the website recorded an interest in the item.

Explain the Use and Honour Objections

Describe the information used for personalisation and the purpose of the processing in clear language. Avoid claiming the business understands someone’s needs when the system has only inferred a possible interest from browsing activity.

Individuals have a right to object to processing for direct marketing, including related profiling. Once someone objects, stop processing their personal data for those purposes. Include connected systems in the withdrawal process rather than updating the email platform alone.

For example, test an unsubscribe or objection through the same route a customer uses, then inspect the connected audience and campaign systems to confirm the person’s marketing status has changed wherever the relevant processing takes place.

Practical and Ethical Challenges

1. Incorrect or outdated data: Correct duplicate events, returned purchases and obsolete product information before using the records to select messages. Specify how quickly corrections reach connected systems.
2. Inappropriate inference: Do not turn a sensitive browsing topic into promotional wording exposing a presumed health condition, financial difficulty or personal circumstance.
3. Unequal treatment: Review which groups receive valuable offers or useful information. An overall performance improvement does not show every audience receives an acceptable experience.

Review Generated Content and Inferred Profiles

NIST’s generative AI guidance identifies risks involving inaccurate content, sensitive inferences and harmful bias. In a marketing workflow, review both the generated message and the information used to select the recipient.

A grammatically correct email still fails review if the offer is unavailable or the opening line implies knowledge the customer never supplied. Keep approved facts separate from predicted interests and give reviewers access to the records needed to check a disputed message.

Limit Data Sent to Suppliers

Before connecting a personalisation service, map the fields leaving the website or CRM. Remove information unnecessary for the task. A product-ranking test does not automatically require full support conversations or unrestricted customer notes.

Check supplier settings and contractual terms covering retention, access and use of submitted information for model training. Assign responsibility for reviewing those settings after a product update or a change in the business’s intended use.

Coordinate Personalisation Across Channels

Omnichannel personalisation requires website, email, app and store systems to exchange the relevant information. AI does not create this connection automatically. Define which system records purchases, preferences and marketing permissions before deciding where model outputs belong.

For example, a purchase recorded in a shop should remove the customer from a corresponding promotional reminder only after the purchase has been reliably linked to the correct customer record, without overwriting unrelated preferences or assuming every purchase ends interest in the category.

Use a shared contact policy to prevent overlapping campaigns from repeatedly messaging the same person. Specify which communication takes priority when a service issue, an active promotion and a recommendation sequence coincide.

Assess New Interfaces Against a Specific Task

Natural-language assistants add a conversational interface to product selection. AR adds a visual layer, such as a product preview. VR creates an immersive environment. None of these features demonstrates a need for individual profiling by itself.

For a proposed AR or VR project, identify the decision the experience helps a customer make and the information required to support the task. A useful product preview deserves evaluation separately from an additional personalisation feature.

Start With One Observable Decision

For an initial test, choose one placement such as a product recommendation block in a permissioned customer email. Keep an approved standard selection available when customer history is missing or the recommendation service fails.

Record which products were eligible, which were selected and which version the customer received. Retain enough information to investigate incorrect recommendations without collecting unrelated personal details. Set the review date around the buying cycle so the next decision rests on completed outcomes rather than early clicks.