Why Your UX Feels Wrong Even When the Data Looks Right

Brands across Southeast Asia are spending serious money on data infrastructure — customer data platforms, real-time segmentation, AI-driven personalisation — and then watching users bounce anyway. The instinct is to interrogate the model. The problem is usually the map.

When Correct Isn’t the Same as Believable

Here’s a design truth that most analytics teams resist: statistical accuracy and perceived accuracy are different products. UX Collective’s deep dive into Tetris history makes this uncomfortably clear. The original game generated pieces truly at random — mathematically clean, probability-perfect. Players hated it. They experienced long droughts of the pieces they needed, then sudden floods of the same shape, and concluded the game was broken or rigged. It wasn’t. Human intuition simply cannot read a probability distribution from a sequence of events.

The fix? Tetris began deliberately weighting piece selection — ensuring variety felt distributed even when pure chance wouldn’t guarantee it. The game started cheating to feel fair.

For digital teams, the implication is uncomfortable: your recommendation engine can be statistically impeccable and still feel arbitrary to users. A Shopee product carousel surfacing items with a genuinely high purchase-probability score can read as random noise if there’s no visible logic threading the selections together. The model isn’t broken. The UX is not doing the explanatory work the model can’t do itself.

Your Information Architecture Is the Real AI Problem

Adrian Levy’s analysis on UX Collective draws on Kevin Lynch’s 1960 study of how people mentally map cities — the finding that humans navigate by landmarks, edges, and paths, not coordinates. Levy’s argument: the same cognitive structure governs how users navigate AI-powered interfaces. A large language model can retrieve and reason with precision, but if the information architecture surrounding it is poorly structured, users will misinterpret responses, lose context, and distrust outputs that are actually correct.

This lands hard for Southeast Asian digital products where the stakes are high and the surfaces are small. LINE’s mini-apps, Grab’s super-app ecosystem, Lazada’s in-app storefronts — these are environments where users make fast, low-trust decisions on 6-inch screens. If the IA isn’t providing clear landmarks — obvious categories, consistent navigation patterns, predictable content hierarchy — AI-generated recommendations get mentally filed under “suspicious,” regardless of their quality.

Practically: before your next AI feature ships, run a card-sorting exercise specifically on how users expect that feature’s outputs to be organised. Not what they want to see, but how they expect it to be structured. The gap between those two answers is where trust breaks down.

---

The Perception Gap Is a Revenue Problem

This isn’t a philosophical concern. Segment’s 2024 State of Personalisation report found that 62% of consumers say they’ll stop buying from brands that deliver poorly personalised experiences — but the definition of “poor” in user minds is rarely about irrelevance. It’s about unexplained surprise. A recommendation that’s right but feels random is experienced as wrong.

For conversion-focused teams, that perception gap has a measurable cost. If your personalisation layer is driving a 15% uplift in click-through but a 22% drop in add-to-cart, the model is likely working fine — the interface around it is eroding trust before the decision moment. Audit the micro-copy, the sequencing logic visible to users, and the consistency of recommendations across sessions. Perceived consistency outperforms actual accuracy at the point of purchase.

Brands like Tokopedia have invested in “reason codes” — small contextual labels that explain why a product is being surfaced (“Because you bought X” or “Popular in your area”). These add no algorithmic value. They add perceived logic, which is the thing users are actually evaluating.

Designing for Human Intuition, Not Model Performance

The through-line across both Tetris and Lynch’s city maps is that humans evaluate systems by the stories those systems appear to tell, not by the underlying mechanics. Your users are not reading your model’s confidence scores. They’re reading sequences, patterns, and structures — and making fast judgments about whether the system understands them.

For design teams managing AI-adjacent products, three implementation priorities follow from this:

First, audit your empty states and low-data moments. When personalisation can’t fire confidently — new users, sparse browsing history, fresh sessions — what does the interface show? A generic grid reads as random. A clearly labelled curated section reads as intentional. Same content, different trust outcome.

Second, pressure-test your IA against your AI outputs. Map the five most common model outputs against your current navigation and content hierarchy. If the outputs don’t map cleanly to existing mental models users already hold, you’re adding cognitive friction at the exact moment you need clarity.

Third, consider mobile-first sequencing specifically. On desktop, users can scan a wide layout and construct their own narrative. On mobile — where the majority of Southeast Asian commerce happens — they read top-to-bottom, linearly. The sequence becomes the argument. If your AI surfaces three unrelated items in a row, mobile users experience that as chaos. Grouping logic must be visible, not just present in the backend.

---

The irony is that the better your models get, the more this problem compounds. Higher personalisation precision means more surprising outputs — and more explanatory work for the interface to carry. The question worth sitting with: as AI gets smarter about what users want, are your design systems getting smarter about making that intelligence legible?

At grzzly, we work with growth teams across Southeast Asia who have solid data pipelines but are losing users at the interface layer — where model outputs meet human perception. If your analytics say one thing and your conversion metrics say another, that gap is usually a design problem wearing a data costume. Let’s talk