Retailers rarely suffer from a shortage of customer data, but customer data personalisation fails for a different reason. A typical ecommerce business can record page views, searches, product impressions, clicks, basket additions, purchases, returns, voucher use, message opens and device changes. The seductive conclusion is that collecting more of these events must produce better personalisation. It does not.
Data is evidence, not understanding. A large event store can describe what happened without establishing what one customer is likely to buy next. The commercial work lies in distinguishing durable preference from temporary curiosity, recognising when an old signal has lost meaning, relating behaviour to the products actually available and converting that interpretation into a decision that can be tested against sales.
This is why personalisation projects often become technically impressive but commercially disappointing. They count everything, unify everything and expose everything in a dashboard, yet still send broadly similar products to broadly similar groups. The organisation has improved its memory without improving its judgement.
Collection and interpretation are different disciplines
Collection answers a factual question: what events can the retailer observe? Interpretation asks a harder one: what does this particular sequence of events imply now?
A customer may view a dress because she wants it, because she is buying for somebody else, because she is comparing fabrics, because an advertisement brought her to the wrong page or because she has already bought something similar elsewhere. Ten product views are not automatically stronger evidence than three. Repetition may indicate growing intent, but it may also indicate uncertainty. A basket addition can be a strong commercial signal, yet it can be superseded by a purchase, invalidated by unavailability or weakened by repeated price checking without progress.
Glossy’s reporting on AI, product discovery and personalisation in luxury retail describes the difficulty of turning additional customer information into useful decisions. It also records a warning from Van Cleef & Arpels against allowing every new data point to trigger more communication. The distinction is decisive: a system earns value by improving the relevance of a decision, not by achieving the largest possible customer profile.
Recency changes meaning
Older behaviour should not retain equal influence indefinitely. A customer who spent an evening comparing winter coats in January should not still be treated primarily as a coat prospect in June. The data remains historically true, but its value for the next decision has changed.
Recency is not a universal expiry date. Its importance varies by product and person. Interest in replenishable skincare may decay differently from interest in furniture. A regular wine buyer may have a recognisable interval between orders; a gift purchase may never recur. The same elapsed time can therefore strengthen one prediction and weaken another. Yehuda Koren’s research on temporal dynamics similarly distinguishes changing individual preferences from short-lived fluctuations; simply discarding older observations can throw away useful evidence.
A useful model must learn these differences rather than apply one arbitrary look-back window to everybody. SwiftERM’s explanation of context in ecommerce personalisation shows why a current situation can be more informative than a static description of the customer. More history is useful only when the system can decide how much of it still matters.
Repetition is not a declaration
Frequency is often treated as certainty: the more often somebody views an item, the more they must want it. That interpretation is sometimes right and sometimes dangerously crude.
Repeated attention can indicate attraction, hesitation, affordability concerns, uncertainty over size, comparison with alternatives or difficulty understanding the product. The surrounding sequence provides meaning. Did the customer return directly to the same item? Did they explore closely related attributes? Did they move steadily towards purchase, or repeatedly leave after seeing delivery terms? Did another purchase satisfy the apparent need?
This is where raw counts fail. Fifteen disconnected views are not necessarily more valuable than a short, coherent progression through related products. Autonomous individualisation should recognise a behavioural pattern, not reward whichever database field contains the largest number.
Sequence reveals changing intent
Order matters. A customer who browses running shoes, reads a size guide, compares two models and returns to one product has created a different commercial narrative from somebody who encounters the same four pages in an unrelated order over several months.
Purchases are particularly important because they change the interpretation of what came before. Once an item has been bought, continued promotion of it may be irrelevant. The next useful product could be a replenishment, a complement, a different category or nothing at all for some time. A system that merely accumulates positive signals will keep reinforcing yesterday’s conclusion. A system that interprets sequence can revise it.
Research by Balázs Hidasi and colleagues examines recommendations based on a whole session rather than isolated item associations. Its experiments found advantages from modelling session behaviour in the datasets studied. That is supporting evidence for taking the surrounding sequence seriously; it is not proof that any particular retail implementation will achieve the same result. The event itself is not the personalisation.
Changing product affinity must be allowed to contradict the past
People change. They discover new brands, move home, acquire different responsibilities, alter budgets, develop interests and abandon old ones. Even within a stable category, colour, material, fit, style and price preferences can move quickly.
Static profiles resist that change. Once a customer has been labelled a premium buyer, a discount shopper or a particular category enthusiast, later evidence is often interpreted through the label. The result looks personalised because it uses customer history, but it may simply preserve an obsolete opinion.
Individualisation requires provisional beliefs. New behaviour must be capable of strengthening, weakening or replacing an earlier affinity. The commercial importance of this principle is explored in why customer segmentation is not personal. The point is not that historical data lacks value; it is that history must remain open to correction.
Isolated events create false confidence
An isolated search, click or product view is easy to record and easy to overstate. It can trigger a journey, assign a segment or start a retargeting sequence before the retailer has enough evidence to understand it.
The danger grows when separate tools each react independently. The search platform sees a query. The email platform sees a click. The advertising platform sees a visit. Each treats its own event as an expression of intent, producing several confident responses to one ambiguous action. More data has then created more noise.
A better approach examines relationships between signals: recency, repetition, sequence, product attributes, transactions, availability and the customer’s own prior rhythms. It should also admit uncertainty. Not every observation warrants a message, and restraint is itself a decision.
Product intelligence matters as much as customer intelligence
Customer data cannot produce relevant product selections if the catalogue is poorly understood. A recommendation needs dependable identifiers, current availability and enough product detail to distinguish meaningful attributes. Two products may share a category while serving very different tastes, budgets or occasions.
For this reason, adding behavioural events while neglecting catalogue quality can reduce performance. The model becomes more certain about the customer but remains unable to map that understanding to the right available item. Personalisation is a matching problem: both sides of the match need substance.
The practical operation is described in how SwiftERM works: behavioural and transactional evidence is continually related to the product catalogue so that each customer’s ranking can change as the evidence changes. The objective is not a richer dashboard. It is a more commercially useful next-product decision.
Identity resolution does not equal insight
A customer data platform may assemble thousands of fields into one profile and still leave the retailer to define segments, triggers and offers. The technical achievement is real, but it is upstream of personalisation. Somebody—or something—must still decide which evidence matters, how it interacts, what has changed and which available product now has the strongest purchase probability.
NIST’s AI Risk Management Framework provides a useful structure for assessing the trustworthiness and risks of an AI system, but adopting a framework is not evidence that its product decisions work. The retailer still needs to establish how reliable inputs become relevant individual selections and how those selections are tested. The distance between a profile and a decision is where many personalisation programmes stall.
More fields can make weak rules look sophisticated
In practice, the rule becomes brittle. It reflects assumptions made when it was written, cannot evaluate every interaction across a large catalogue and needs constant maintenance. It also treats customers who pass the same conditions as interchangeable. The system may use more data without becoming individual.
Autonomous product selection takes a different approach. It recalculates the relative relevance of products for each identifiable customer rather than waiting for a marketer to anticipate every useful combination. That does not remove governance; it directs governance towards data quality, commercial boundaries and measurement instead of endless workflow construction.
Consent and minimisation improve commercial discipline
Collecting more personal data also creates responsibility. The Information Commissioner’s Office explains that data minimisation means personal information should be adequate, relevant and limited to what is necessary for the stated purpose. This is not merely a compliance constraint. It is a useful design test.
Commercial measurement decides whether interpretation is good
No amount of model complexity proves that personalisation works. The final test is whether the decisions create incremental commercial value.
Clicks and opens can help diagnose a communication, but they are not equivalent to additional orders. A model can become extremely good at attracting curiosity while recommending products that customers do not buy. It can also claim revenue from customers who would have purchased without the intervention.
Microsoft Research’s work on controlled experiments explains how randomisation and a sound experimental design help establish whether a change caused an outcome. Applied to ecommerce, that means preserving an appropriate holdout and comparing commercial results between equivalent groups. More customer data is worthwhile when it helps produce better decisions whose effect survives that test, rather than merely improving the attribution attached to existing sales.
What retailers should ask before collecting another field
A useful evaluation begins with the decision, not the database. Retailers should ask:
- Which commercial decision will this evidence improve?
- Can the system distinguish recency, repetition and sequence?
- Can a purchase or new behaviour overturn an earlier conclusion?
- Does the interpretation vary appropriately by product and individual?
- Is the catalogue detailed and current enough to support the match?
- Can uncertainty result in no communication?
- Will the outcome be tested against incremental orders and turnover?
- Is the collection necessary, lawful and proportionate to its purpose?
These questions expose the difference between a data project and a commercial decisioning system.
The advantage comes from learning, not accumulation
Customer data is indispensable, but volume is not the objective. The advantage comes from continually revising what the evidence means for one person, one catalogue and one moment.
Good individualisation notices that recent actions can outweigh old ones, that repetition can express doubt, that sequence changes interpretation, that purchases invalidate assumptions and that people do not remain fixed inside the categories assigned to them. It also accepts that the most intelligent response is sometimes to wait.
Retailers therefore need not choose between sparse insight and indiscriminate collection. They need a disciplined connection between first-party evidence, evolving individual interpretation, available products and measurable commercial outcomes. That is what turns customer data from a growing cost of storage into a source of additional sales.
References and accreditations
- Danny Parisi, Glossy, How Van Cleef & Arpels and The Wires are using AI for product discovery and personalization — 29 September 2026; retailer interviews on the useful interpretation of customer information and restraint in communication.
- Yehuda Koren, Communications of the ACM, Collaborative Filtering with Temporal Dynamics — 2010; changing preferences, temporary effects and the limits of indiscriminate time-window approaches.
- Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas and Domonkos Tikk, Session-based Recommendations with Recurrent Neural Networks — 2016 version; experiments concerning recommendations from session behaviour rather than isolated item associations.
- Information Commissioner’s Office, Principle (c): Data minimisation — official UK guidance on adequate, relevant and necessary personal information; guidance notes that it is under review following legislative changes.
- Microsoft Research, Online Experimentation at Microsoft — 2009; randomisation, experimental design and the causal evaluation of changes.
- National Institute of Standards and Technology, AI Risk Management Framework — a framework for assessing AI trustworthiness and risk, rather than an endorsement of a particular supplier or performance claim.
