Changing customer intent is the reason a returning visitor can be completely familiar to a retailer and still be misunderstood. The customer record may be identical to the one seen yesterday. The customer is not. A purchase has been made, a need has disappeared, a new occasion has emerged, the available budget has changed, or the relative appeal of the products now in view has shifted.
That is not an edge case. It is the normal condition of ecommerce. A customer profile is a useful history, but history does not freeze the next decision. Retailers that treat a person as a permanent segment confuse what has been observed with what will matter now.
The commercial problem is easy to miss because the data can look reassuringly complete. The retailer knows the customer’s identity, previous orders, browsing, returns, favourite category and average spend. Yet none of those facts, alone, answers the only question that matters at the next interaction: which product is becoming most relevant to this individual at this particular moment?
Changing customer intent begins between visits
Consider the same customer across four visits. On Monday she is replacing something practical. On Thursday she is researching a gift. A month later she is considering a purchase for an event. At Christmas she is buying for several other people. Her account, address and order history remain the same. The purpose of each visit does not.
Even within one visit, intent develops. A broad search becomes a comparison; a comparison reveals a constraint; a constraint changes which attributes matter. Price may dominate at the beginning and availability at the end. A colour that appeared irrelevant can become decisive once an alternative product establishes a reference point. Changing customer intent is therefore visible inside a session as well as between visits.
Research into real-world choice supports the underlying principle. A study reported in Scientific Reports found that evaluations and choices were influenced by the surrounding choice set, rather than being fixed properties of the options alone. In practical terms, what a customer wants cannot always be separated from what they are comparing it with at that moment.
This is why a static label such as “premium buyer”, “discount shopper” or “category enthusiast” is an inadequate operating model. It compresses a changing person into a retrospective average.
History is evidence, not an instruction
Purchase history is valuable because it records demonstrated preference. It can reveal replenishment patterns, category affinities, price tolerance and the interval between related purchases. The mistake is to turn that evidence into a permanent campaign rule.
A customer who bought running shoes does not remain a running-shoe prospect indefinitely. The purchase may have satisfied the need. It may also alter the next likely purchase: specialist clothing, recovery equipment, a gift, or nothing from the category for a considerable period. Recommending more of what has just been bought is often the clearest sign that a system has learned the event but not its consequence.
The same limitation applies to browsing. A product view may express desire, comparison, curiosity, doubt or a failed search for something else. A product left behind is not automatically an abandoned intention. It may have helped the customer establish that they wanted a different material, size, price or use altogether.
SwiftERM’s distinction is important here. Next-purchase prediction is not the same as arranging recommendations around a previous item. It considers the individual’s changing probability of buying each product, rather than treating an old action as a command to repeat the category.
The choice set changes the customer’s apparent preference
Ecommerce teams often talk about a person “having” a preference as though it were stored intact, waiting to be retrieved. In reality, preferences are frequently constructed during the act of choosing. The available alternatives, their order, the attributes made easy to compare and the time available for deliberation all influence the result.
A review of preference construction indexed by the US National Library of Medicine notes that preferences are sensitive to context and can be formulated during judgement and choice. Goals, experience and cognitive constraints affect that process. The implication for retail is not that customer data is unreliable. It is that the meaning of the data depends on the decision now being made.
This matters at catalogue scale. A person may have an enduring liking for a brand or aesthetic while the best next product changes as stock, season, ownership, price, occasion and competing options change. An individual can be consistent in taste and variable in purchase intent at the same time. That distinction is central to recognising changing customer intent without discarding durable preferences.
Segmentation struggles with that distinction. A segment describes similarities between people; it does not resolve the changing probabilities within one person. Increasing the number of segments only produces narrower averages. It does not create an individual decision.
More data does not solve a stale decision model
The instinctive response is to collect more data. That can improve observation, but it does not guarantee a better decision. A customer data platform can assemble a richer record while the activation layer continues to apply fixed rules: bought X, send Y; viewed category A, enter journey B; crossed threshold C, receive campaign D.
These rules are understandable and auditable, but they inherit the assumptions of the person who wrote them. They decide in advance which event matters, what it means, which products belong together and how long the conclusion should remain valid. The customer can change faster than the rulebook.
That is why more customer data does not automatically produce better personalisation. The decisive question is how the system converts evidence into a fresh product decision for each individual, not how many fields are present in the profile. A large record interpreted through a stale rule remains a poor response to changing customer intent.
Data quality still matters. Identity, consent, transactions, returns and catalogue information must be dependable. But once that foundation exists, commercial improvement comes from better inference, not indefinite accumulation.
Context and intent must be interpreted together
Google’s retail guidance has long made a useful distinction between intent and context. Intent is what a shopper wants in the moment; context includes the situation in which that want is expressed. Device, location and time can change what is useful, even when the retailer already knows the customer.
For autonomous individualisation, context is broader than a handful of environmental fields. It includes the individual’s sequence of behaviour, the catalogue presently available, what has recently been bought, changing product relationships and the movement of probability over time. The aim is not to invent a psychological story about the customer. It is to make the next product decision from current evidence rather than an inherited campaign assumption. Changing customer intent must be inferred from behaviour, not dramatised into an unsupported personal narrative.
Responsible use is equally important. The Information Commissioner’s Office guidance on profiling explains that predicting preferences and behaviour remains subject to data-protection principles, transparency and appropriate safeguards. Relevance is not a licence to obscure how personal data is used.
In ordinary ecommerce product selection, the goal should be proportionate: use authorised behavioural and transactional evidence to improve relevance, avoid unsupported inferences, and retain clear governance over what the system is allowed to do.
Campaigns remember the plan; autonomous systems reassess the person
A campaign is designed around a proposition chosen in advance. The audience may be refined, the content personalised and the send time optimised, but the commercial idea remains collective: this group will receive this treatment.
Changing customer intent exposes the limit. If the proposition stops being relevant to one individual, the campaign does not normally redesign itself around that person. It continues because the schedule, segment and creative have already been approved.
An autonomous system works in the opposite direction. It does not begin with a campaign that needs recipients. It begins with the individual and repeatedly determines which products are becoming most relevant to that person. Where no product has sufficient probability, the correct decision may be not to force a communication.
This is also why operating burden changes. Marketers no longer need to anticipate every useful combination of customer, category, timing and message. The system can consider the catalogue continuously at individual level. Our account of the first 30 days of autonomous individualisation explains how this operating model moves from data connection to measured live decisions without requiring a new campaign for every emerging opportunity.
What ecommerce leaders should test
The claim that intent changes should be visible in commercial evidence. Retailers can test it without accepting a vendor’s vocabulary on trust.
- Measure prediction freshness. Establish whether product probabilities are recalculated as new behaviour and transactions arrive, or whether the system merely refreshes a static segment.
- Inspect post-purchase behaviour. Check whether the system recognises that a purchase can satisfy a need and alter the next likely category, rather than continuing to promote close substitutes.
- Compare individual decisions with campaign rules. Sample customers whose recent behaviour differs from their historic category and examine which method produces the more plausible next-product choice.
- Include the whole eligible catalogue. A system cannot detect a changing intention if it only chooses among products preselected for a campaign.
- Use a holdout. Plausible recommendations are not enough. Compare treatment with an equivalent untreated group to determine whether the decisions create incremental revenue or gross profit.
- Review silence as a decision. If the system always finds something to send, it may be optimising contact volume rather than relevance.
These tests separate genuine individualisation from polished segmentation. They also create a practical due-diligence path for retailers evaluating whether a platform can respond to changing customer intent or only describe it.
The commercial value of treating every visit as new evidence
A returning customer is valuable partly because the retailer knows more about them. The advantage disappears when that knowledge becomes a stereotype. The best use of history is not to keep the customer in the same box. It is to improve the next inference while remaining willing to revise it as changing customer intent produces new evidence.
That creates a different commercial rhythm. Product selection can adapt after purchase rather than waiting for a marketer to amend a journey. Newly relevant items can surface without being manually assigned to a segment. Contact can occur when the individual opportunity strengthens, not simply when the campaign calendar reaches its next slot.
The result is not novelty for its own sake. It is a more exact allocation of attention. The retailer communicates when there is evidence of relevance, selects from the products the individual is becoming most likely to buy, and measures whether acting on that prediction produces additional value.
The same customer will return with the same name and a different decision. Changing customer intent is not an exception for ecommerce systems to manage; it is the condition they should be built around.
Questions ecommerce teams commonly ask
Does changing intent mean historic data is no longer useful?
No. Historic behaviour is essential evidence. The point is that it should inform a current prediction rather than become a permanent label or an instruction to repeat the last category.
Can segmentation respond to changing customer intent?
Segments can be refreshed, but they still apply a collective rule to a group. Autonomous individualisation recalculates product relevance for each person, allowing two customers with similar histories to receive different decisions and one customer’s decision to change over time.
How should a retailer prove the commercial effect?
Use a properly constructed holdout and compare revenue or gross profit per eligible customer. Attribution can show that a communication was present; a treatment-and-control comparison is needed to estimate whether it created an outcome that would not otherwise have occurred.
References and accreditations
- Otto, Devine, Schulz, Bornstein and Louie, Context-dependent choice and evaluation in real-world consumer behavior, Scientific Reports — evidence that real-world choices and valuations vary with the surrounding choice set.
- Warren, McGraw and Van Boven, Values and preferences: defining preference construction — a review of how goals, context, constraints and experience influence preferences during judgement and choice.
- Google, How Mobile Has Redefined the Consumer Decision Journey — the practical relationship between shopper intent, context and prior customer knowledge.
- Information Commissioner’s Office, Rights related to automated decision-making including profiling — the definition, governance and safeguards relevant to predicting preferences and behaviour.