AI Personalisation in E-commerce: Where Things Actually Stand in 2026

A colleague of mine (she works in product at a mid-sized UK fashion retailer) described her team's personalisation roadmap as "endlessly almost there." They've been investing in recommendation infrastructure for three years. Their click-through rates are up. Their customers, when surveyed, say they find the suggestions mostly irrelevant.

That tension is everywhere right now. The technology has moved substantially. Consumer patience has not.

What's actually changed on the technology side

The clearest shift of the past two years is in search. Traditional e-commerce search is built on keyword matching. You type "blue trousers" and the system looks for products tagged "blue" and "trousers." Type "something smart but comfortable for a long haul flight" and the system stares back blankly, returns chinos, and calls it done.

LLM-powered search interprets natural language to deliver results based on context, intent, and sentiment, not just the literal words in the query. That's not a marketing claim at this point; it's a fairly mundane engineering fact. The more interesting question is whether retailers have actually deployed it, and whether it's making a measurable difference.

Zalando's rollout of its AI assistant across all 25 of its markets (live since October 2024, in local languages) is the most concrete European proof point. Over 2 million customers had used it by March 2025. During the pilot of its deeper personalisation features, connecting the assistant to individual account history and preferences, Zalando observed a 40% increase in high-value interactions such as likes and adds to cart, compared to standard browse sessions. That's an engagement metric, not a conversion rate, but it's directionally significant.

ASOS's equivalent is "Styled for You," which uses an AI trained on more than 100,000 curated studio outfits to generate personalised outfit suggestions based on each customer's purchase and search history. The company's CEO has noted that customers who interact with the AI stylist save 50% more items for later viewing, a leading indicator that tends to track through to purchases.

Meanwhile, in UK bricks-and-clicks retail: Boots is running adaptive AI across its online search and trialling a ChatGPT-based chatbot for conversational product guidance. John Lewis is optimising product descriptions and editorial content for generative search queries. M&S has personalised homepages, body-shape quizzes, and in-house AI models doing recommendation work.

None of this is hypothetical. It's deployed. That part of the "promise vs. reality" story has closed.

The part that hasn't closed

Here's the number that should sit uncomfortable with anyone building personalisation products: 61% of UK consumers are now using or have used AI when they shop. But 68% of them can't name a single AI-powered retail experience that actually impressed them.

That's from CI&T's Retail Tech Reality Check, a December 2025 survey of 2,000 UK and Ireland consumers. The gap it describes is not a technology problem. The underlying models are capable. The gap is an experience design problem, and underneath that, a trust problem.

Scurri's research from September 2025, which surveyed 1,000 UK consumers, found that 72% of shoppers are apprehensive about AI making retail decisions without their input. And 94% say it's vital that AI tools are transparent about how they operate and how they handle data.

Ninety-four percent. That's not a vocal minority.

Rory O'Connor, Scurri's CEO, put it plainly: "Consumers welcome AI but on their own terms. They want the benefits of speed, personalisation and convenience, but they also want transparency, choice and control."

The retailers doing personalisation well right now are the ones who've figured out that "why am I seeing this?" is not a liability question. It's a conversion question. When customers understand the basis of a recommendation (you bought this in spring, here's what works with it for summer) they trust it more. When recommendations feel conjured from nowhere, they feel surveilled.

What collaborative filtering actually means now

The "customers who bought X also bought Y" framing is not dead. It's just not the whole stack anymore.

Modern recommendation systems layer behavioural signals (what someone clicks, how long they pause, what they add and remove from a basket) with contextual data (time of day, device, location) and LLM-interpreted intent signals. The system is trying to answer a harder question than "what do similar customers buy?" It's trying to answer "what does this customer need right now, that they might not have articulated yet?"

The gap between those two questions is where the interesting engineering is happening. It's also where the privacy tension lives. Answering the second question requires more data, held for longer, interpreted more deeply. Which is exactly what 72% of UK shoppers say makes them nervous.

What transparency-first personalisation actually looks like

The transparency obligation gets talked about in abstract terms quite often — "be transparent about how you use data," full stop. That's not actionable design guidance.

What it looks like in practice is closer to this: a customer opens a recommendations panel and alongside the product grid there's a one-line explanation. "Based on your recent Barbour jacket purchase and your browsing in the boots section this week." Or: "Because you've bought gifts for gardeners before — here are options under £40." Not a privacy policy link. Not a cookie consent modal. An actual, plain-language explanation of the inference.

M&S does a version of this in its app, where body-shape and style quiz inputs are surfaced back to the customer as explicit filters they can see and adjust. Boots does it through its Beauty Profile, where skincare recommendations are tied to stated skin type rather than opaque tracking. Neither is perfect, and neither is as sophisticated as what the AI models underneath them are actually doing. But they've moved in the right direction: making the recommendation legible to the person receiving it.

The gap between "the AI knows a lot about you" and "the AI helps you" comes down to whether that knowledge is visible. When it's hidden, it's surveillance. When it's surfaced, it's service. Most personalisation systems are still closer to the former than retailers would care to admit.

Where this goes

The next iteration that's already in pilot at a handful of European retailers is genuinely conversational personalisation: not a chatbot bolted onto a product grid, but a system that co-creates the shopping experience through dialogue. "I need an outfit for my sister's wedding, I'm the maid of honour, the ceremony is outdoors." The system holds that context across the session, adjusts as the customer narrows or expands, and can explain its choices.

The technology for this exists. The question is whether the experience design is good enough that it doesn't feel like talking to a form.

Zalando's assistant is the closest thing to a deployed version of this that's publicly documented. Whether it's a proof of concept or a genuine step change will depend on what the conversion data looks like over a full year, not just a pilot.

What I'd watch for in the second half of 2026 is whether UK retailers start publishing transparency-first personalisation features alongside the capability features. Not just "here's our AI assistant" but "here's what it knows about you and here's how to adjust it." The 94% who say transparency is vital are not going anywhere. The retailers who treat that as a UX requirement rather than a legal footnote are the ones who'll convert the engaged-but-unimpressed majority.

The technology is there. The trust architecture is still being built.