Machine learning reached e-commerce earlier and more thoroughly than almost any other commercial domain, for an unromantic reason: online retail generates exactly the kind of data ML needs. Every search, click, cart addition, abandonment and return is a labelled example, produced continuously and at volume, with an outcome attached.
That head start means e-commerce is a useful place to see which applications actually earn their keep — and which ones have absorbed a decade of investment without moving a number.
The applications that reliably pay
Search and ranking
The single highest-return application, and the one most often under-invested. A shopper who searches has told you exactly what they want; failing to return it is the most expensive mistake in the funnel. Improvements to query understanding, synonym handling, typo tolerance and result ranking tend to show up in revenue faster than anything else on this list, because they operate on traffic that had already decided to buy.
If you have limited engineering capacity, spend it here before recommendations.
Fraud and payment risk
Well-suited to ML on every axis: high volume, clear labels, immediate feedback, and a directly measurable cost to being wrong in either direction. It is also a domain where the model must be retrained continuously, because the adversary adapts — a fraud model is one of the few that genuinely degrades because someone is actively working to degrade it.
Demand forecasting and inventory
Less visible than recommendations and often worth more. Better forecasting means less capital tied up in stock, fewer stockouts on the items that sell, and fewer markdowns on the items that don't. The gains are measured in working capital rather than conversion rate, which is why this work tends to be championed by finance rather than growth.
Recommendations
Genuinely valuable, and the most oversold item on the list. Recommendations work well where a catalogue is large enough that browsing it is impractical and purchases are frequent enough to establish a pattern. On a small catalogue, or with infrequent purchases, a well-merchandised category page frequently outperforms a personalised carousel — and costs far less to maintain.
Where the money tends to go missing
Two patterns account for most disappointing e-commerce ML programmes.
Personalisation that personalises nothing. A recommendation engine trained mostly on popularity will show everyone roughly the same best-sellers while reporting healthy engagement, because best-sellers get clicked regardless. The measurement trap is that it looks like it is working. The only reliable way to know is to test against a non-personalised baseline and require an actual lift — many deployed systems have never been tested this way.
Chatbots deployed as a deflection metric. Support automation succeeds when it resolves the questions that are genuinely repetitive — order status, returns policy, delivery windows — and hands over cleanly when it cannot. It fails when its target is reducing contacts rather than resolving them, at which point it becomes an obstacle between a customer and a refund, and the cost lands in retention where nobody attributes it back.
The question is never "does the model perform well?" It is "did the business metric move against a baseline that didn't have it?"
The unglamorous prerequisite
Every application above depends on the same foundation, and it is where most programmes actually stall: event data you can trust.
- Consistent event definitions. If "add to cart" fires differently on web and mobile, every downstream model learns the discrepancy rather than the behaviour.
- Stable identity resolution. One customer across devices and sessions. Broken identity fragments the history each model depends on.
- Clean catalogue data. Categories, attributes and descriptions that mean the same thing throughout. Most "the recommendations are strange" complaints resolve to a mislabelled catalogue rather than a bad model.
- Reliable outcome data. Purchases, returns, refunds and chargebacks joined back to the session that produced them. Without the join there is no ground truth to learn from.
This is data engineering, and it is the majority of the work. Teams that skip it end up with sophisticated models trained on an inaccurate picture of their own business — a failure that is difficult to detect because the model still produces confident output.
A sensible sequence
- Fix search first. Highest intent, fastest measurable return.
- Instrument properly. Consistent events, resolved identity, joined outcomes.
- Add forecasting if you hold inventory — the working-capital gain is usually larger than the conversion gain.
- Then personalise, with a holdout group from day one so you can prove it is doing something.
- Automate support where the questions genuinely repeat, measured on resolution rather than deflection.
Each step is testable, each produces a number, and each builds the data foundation the next one needs.
The takeaway
Machine learning in e-commerce is a mature field with well-understood applications and well-understood failure modes. The gap between retailers who get value from it and retailers who don't is rarely about model sophistication — nearly everyone is using comparable techniques.
It is about whether the underlying event data is trustworthy, and whether anyone is measuring against a baseline honest enough to fail.
We build the event pipelines, search and forecasting systems that make retail ML worth the investment. Tell us where your funnel leaks.
