Can AI Truly Understand Your Shoppers? Breaking Down Intent Detection

'Understanding the shopper' has been the aspiration of retail marketing for as long as retail has existed. The merchant who knows which customers are serious buyers, which are browsing, which are price-sensitive, and which are loyal enough to tolerate a stockout — that merchant has a commercial advantage over one who treats every customer identically. The challenge has always been scale: the small shopkeeper who genuinely knows their regular customers cannot replicate that understanding across ten thousand daily website visitors.

AI intent detection claims to solve this problem at scale — to identify, from behavioural signals, what individual shoppers are looking for, how close they are to a purchase decision, and what kind of experience or intervention is most likely to serve them well. The claim is significant. The reality is nuanced. AI intent detection is genuinely powerful, meaningfully limited, and widely misunderstood in both directions — overstated by vendors making claims that exceed what the technology can deliver, and undersold by sceptics who compare AI intent detection to human understanding and find it wanting.

This post aims for clarity: what AI intent detection actually does, how it works, where it is genuinely accurate, and where its limitations matter for how retailers should use it.

What Intent Detection Actually Is

Behavioural Signal Processing, Not Mind-Reading

AI intent detection does not read minds. It reads behaviour. The distinction is important because it sets accurate expectations for what the technology can and cannot do. An intent detection system does not know what a shopper is thinking — it infers what a shopper is likely to want based on what they have done: the pages they have visited, the sequence in which they visited them, the time they spent on each, the searches they made, the products they added to and removed from a cart, and the patterns in all of that behaviour that correlate with specific outcomes in historical data.

The inference from behaviour to intent is probabilistic, not certain. A shopper who has visited the same product page four times and spent significant time on the size guide is probably close to a purchase decision for that product — the behavioural pattern is a reliable correlate of conversion in historical data. But 'probably' is not 'certainly.' Some proportion of shoppers who exhibit that pattern will not convert, for reasons that the behavioural signals do not reveal: they found a better price elsewhere, they changed their mind about the category, they were researching for someone else.

The Three Layers of Intent

Intent in a retail context has three layers, each more specific and more actionable than the last:

• Category intent: the shopper is interested in a general product category. Identified from search terms, category browse behaviour, and content consumption patterns. The least specific layer but the most broadly useful for targeting and personalisation.

• Product intent: the shopper is interested in a specific product or a small set of products. Identified from product page visits, comparison behaviour, return visits to the same products, and the narrowing pattern that characterises genuine evaluation. More specific and more actionable than category intent.

• Purchase intent: the shopper is close to a purchase decision. Identified from the combination of high product-level engagement, return visits, checkout initiation, and the specific behavioural sequences that in historical data most reliably precede conversion. The most specific and most commercially valuable layer.

Most intent detection systems operate across all three layers but with varying accuracy at each. Category intent detection is reliable. Product intent detection is reliable when the shopper has expressed sufficient signal. Purchase intent detection is reliable for the clearest cases and noisy for the marginal ones — and the marginal cases are often where the intervention would matter most.

How AI Intent Detection Works

Pattern Matching at Scale

The foundation of AI intent detection is pattern matching: comparing current session behaviour against historical session behaviour to identify which current sessions most closely resemble the sessions that historically preceded specific outcomes. A model trained on millions of historical sessions learns which behavioural patterns correlate with purchase, with category-level interest, with price sensitivity, with churn risk, and with dozens of other commercially relevant outcomes.

When a new session begins, the intent detection system processes its behaviour in real time and scores it against the learned patterns, producing intent signals that reflect how closely the current session resembles the historical patterns associated with each outcome. The more historical data the model has been trained on, and the more closely the current behaviour resembles patterns in that data, the more accurate the intent signals.

The Role of Context

Intent signals are not produced from session behaviour alone — they are produced from session behaviour in context. The same product page visit carries different intent signals depending on how the shopper arrived at it (organic search for the product name versus a broad category search), what they did before visiting it (a direct navigation versus a broad category browse), and what they do after visiting it (return visit versus switching to a competitor comparison).

Context is what separates sophisticated intent detection from simple page-view counting. A system that can read the sequence, the source, the depth of engagement, and the transitions between engagement types produces meaningfully more accurate intent signals than one that counts product page visits without context.

Real-Time vs. Batch Processing

Intent detection can operate in real time — processing session behaviour as it happens and updating intent signals continuously — or in batch mode, processing session data after the fact to inform future campaigns and personalisation decisions. Real-time intent detection is more technically demanding and enables in-session interventions: the personalised recommendation that arrives while the high-intent shopper is still deciding, the offer triggered by the exit intent signal before the shopper leaves. Batch intent detection is more accessible and enables campaign-level personalisation: the email sequence triggered by a session that exhibited high purchase intent without converting, the segment that receives a targeted promotion based on the category intent they expressed in the previous week.

Both modes have commercial value. The distinction that matters for deployment decisions is whether the use case requires in-session intervention (which requires real-time processing) or campaign-level personalisation (which batch processing can support).

Where Intent Detection Is Accurate — and Where It Is Not

High Accuracy: Clear Signal Cases

Intent detection is most accurate when the behavioural signal is clear and the historical correlates are strong. A shopper who has visited a specific product page multiple times across different sessions, spent significant time on the product details and customer reviews, initiated checkout and abandoned it, and returned to the product page in a subsequent session is exhibiting a behavioural pattern that very reliably indicates high purchase intent. The intent detection system's confidence in this case is warranted, and interventions targeted at high-confidence cases produce strong conversion lift.

Lower Accuracy: Ambiguous Signal Cases

Intent detection is less accurate when the behavioural signal is ambiguous — when the current session behaviour resembles multiple different historical patterns without clearly resembling any single one. A shopper who is browsing broadly across a category, visiting many products briefly without engaging deeply with any of them, may be at the beginning of a genuine evaluation, doing casual inspiration browsing, or researching for someone else entirely. The intent signal in this case is weak, and interventions targeted at ambiguous-signal sessions should be lower-friction and lower-cost than those targeted at high-confidence cases.

The New Shopper Problem

Intent detection based on historical session data has a structural limitation: it cannot produce accurate intent signals for shoppers whose behaviour does not resemble anything in the historical data. First-time shoppers from new acquisition channels, shoppers in newly launched product categories, and shoppers whose behaviour patterns are genuinely novel will produce weaker intent signals than shoppers whose behaviour resembles the historical training data. This is not a failure of the technology — it is an inherent property of any system that infers from historical patterns.

The practical implication is that intent detection should be treated as a continuously improving system rather than a deployed-and-done capability. As new sessions accumulate, the historical data base expands, and the accuracy of intent signals for previously ambiguous patterns improves. Intent detection deployed today will be meaningfully more accurate in twelve months than it is on day one, if the system is designed to learn from its own outcomes.

What Intent Detection Cannot Do

Intent detection cannot tell you why a shopper is behaving as they are — only what they are doing and how that behaviour correlates with historical outcomes. A shopper who has visited a product page multiple times but has not converted may be close to a purchase decision, may be waiting for a price drop, may be experiencing a payment friction they have not expressed, or may have decided against the product for reasons that are entirely internal to them. The behavioural signal is the same; the reasons are different; and the right intervention differs accordingly.

This limitation is most consequential for the treatment of high-intent non-converters — the shoppers who exhibit strong purchase intent signals without converting. A system that treats all high-intent non-converters identically will produce good results for some and suboptimal or counterproductive results for others. The most sophisticated intent detection deployments layer qualitative signals — survey responses, support interaction content, declared preferences — alongside behavioural signals to reduce this ambiguity.

Intent detection also cannot substitute for the fundamental quality of the product, the pricing, and the purchase experience. A shopper who correctly detects high intent and targets an intervention at a shopper who has then encountered a broken checkout flow, an out-of-stock notification, or a price that has just been significantly undercut by a competitor will not convert the session through better intent targeting. Intent detection identifies who is close to buying — it cannot create the conditions that make buying happen.

Using Intent Detection Well

• Calibrate interventions to signal confidence: high-confidence intent signals justify high-effort interventions (personalised recommendations, targeted offers, proactive support). Low-confidence signals justify lower-friction interventions (gentle personalisation, relevant content surfacing) that do not feel intrusive when the signal is wrong.

• Treat intent detection as an input, not an answer: intent signals inform decisions; they do not make them. Human review of intent-driven campaigns, regular accuracy assessment, and continuous model improvement are requirements for sustained performance, not optional enhancements.

• Build for real-time when the use case demands it: in-session intervention capability requires real-time processing infrastructure. If the most commercially valuable use cases involve reaching shoppers while they are still in the session, the infrastructure investment in real-time processing is necessary, not optional.

• Layer behavioural and declared signals: where feasible, supplement behavioural intent signals with declared information — purchase purpose, preference declarations, survey responses — that reduces ambiguity about the why behind the behaviour.

Conclusion

AI intent detection does not understand shoppers in the way a skilled retail associate understands a regular customer. It identifies patterns in behaviour that correlate with specific outcomes, and it identifies them at a scale and speed that no human process can match. That is a genuinely valuable capability — not because it is perfect, but because it is systematically better than treating all shoppers identically, and it improves continuously as more data accumulates.

AI does not read minds. It reads patterns. For retailers who understand the difference, that distinction is the foundation of deploying intent detection well — and getting commercial value from it that those with unrealistic expectations will not.