Content Benefit

Content Strategy | UX Writing | Content Design | Globalization | Communications

Personalization is tiring

Personalization is tiring

What interfaces that adapt on their own actually do

The idea has been circulating for years, though it has changed scale within a few months. Language models no longer limit themselves to suggesting content inside fixed containers: they build the container itself, screen after screen, based on what somebody has just written, searched for or tapped. Home pages that reorder themselves, checkout flows that shorten or stretch, alerts calibrated on the hour when a click usually arrives.

Arin Bhowmick, on UX Collective, places the real-time generation of tailored screens among the ten shifts of 2026, and immediately afterwards adds an entry that reads like a corrective: adaptation that respects boundaries, without intrusive surveillance. Both items sit in the same list because the industry has already sensed an open problem. The technical workings of those moving layouts appeared in Generative UI: always adapting.

91% call at least one profiling practice intrusive

The freshest evidence comes from a ThoughtSpot survey run in June 2026 across 4,833 adults in the United States and the United Kingdom. Fully 91% of respondents flag at least one of the techniques currently used in digital retail as intrusive. In detail: tracking that follows people from one site to another collects 57.6% of the complaints, overly frequent recommendations 56.7%, real-time geolocation 48.8%, references to delicate subjects 48.7%, and the same share goes to the habit of anticipating a need before anybody expresses it.

The picture worsens on the trust side. Only 13% say they believe suggestions generated by artificial intelligence, 47% openly reject them, and 66% report having received proposals that were off target. Some 74% reckon companies already gather an excessive quantity of personal information. Whoever designs tailored experiences, in short, works in front of a sceptical audience.

The creepiness factor described by Nielsen Norman Group

Kim Flaherty and Kate Moran had photographed the dynamic back in 2019, in a study whose title spoke plainly about the perils of excessive individualization. Their fieldwork surfaced four unwanted consequences: boredom with a stream forever identical to itself, the echo chamber that returns only opinions already held, the unpleasant feeling of being watched, and lost occasions for discovering fresh interests.

The most telling passage concerns that discomfort. One research participant found it sinister that a social platform seemed to know about searches made elsewhere. Other people abandoned applications turned monotonous, because the mechanism had shut them inside a narrow niche. Nielsen Norman Group itself, in work signed by Aurora Harley in 2018, had recorded a very different climate: benign resignation, tolerance for poor suggestions, willingness to swap data for convenience. Setting those interviews against the 2026 figures shows how quickly the credit ran out.

Why continuous adaptation adds effort instead of removing it

Here sits the paradox that makes the subject worth examining. The commercial promise of aggressive profiling consists in cutting options down: less noise, less wasted time, less decision paralysis. The theoretical reference is well known, and it still stands. Sheena Iyengar and Mark Lepper, back in 2000, demonstrated that a stall with six jams sells far better than a stall with twenty-four, and that whoever chooses from a narrow range remains happier with the decision taken.

Except that a shifting digital environment moves the load onto a different plane instead of lightening it. Anyone opening an application reorganized since yesterday first has to work out what changed, then judge whether the reshuffle deserves any trust, then hunt for the feature that used to sit elsewhere, and finally decide whether to correct the system or let it run. None of those operations belonged to the original path, and each one consumes the scarce resource discussed in Attention please, if you will. A journal in the field, the International Journal of Human-Computer Interaction, published research in 2026 linking the perception of being watched to cynicism about data protection and to self-limiting behaviour: whoever feels observed stops looking certain things up. Hard to imagine a worse outcome for a commercial product.

An experience that keeps changing shape becomes unmanageable backstage

Then comes the part that rarely enters the enthusiastic presentations, and it concerns the daily labour of whoever keeps the product running.

Testing loses its object: checking a screen that never appears twice in the same arrangement resembles photographing the sea. Accessibility gets harder, because the memory of positions and the consistency of routes matter enormously for people navigating with a screen reader or a keyboard, as noted in Accessibility by default. Customer support stops seeing the same page as the caller, so every conversation restarts from a description of somebody else’s display.

On the editorial front the affair turns thorny. Microcopy grows out of revisions, trials and calibration against contexts of use, as recounted in The discipline of UX writing. If wording gets composed on the spot by a model, that chain of control snaps, and with it terminological consistency, reviewed translation, and the voice that tells one brand apart from another. Add compliance: automated profiling means legal bases, comprehensible notices, the right to object. Nothing on that list can be improvised in production. And a badly calibrated overhaul costs plenty in goodwill, as I argued for The unneeded Trenitalia redesign.

Personas give back a readable perimeter

Jakob Nielsen wrote a sentence in 1998 that has aged remarkably well: automatic personalization often works as an alibi for failing to design decent navigation. People resist the idea of being pigeonholed by a machine, their requirements shift from visit to visit, and no static profile withstands that variability.

The way out I prefer remains segmentation by personas, meaning small numbers of hand-designed, documented groups. Kim Flaherty, in a 2020 piece on how wide such models should be, points to the trade-off worth keeping in mind: the further the perimeter stretches, the thinner the data supporting it becomes. Put differently, five profiles built on genuine research beat five million micro-segments generated by an algorithm nobody can inspect.

How to set up targeting that survives contact with reality

A workable method runs through a handful of explicit choices.

Define between three and six profiles born from interviews and usage data, never from marketing assumptions. Write the textual variants by hand, review them, translate them with care. Keep the activation rules readable, so that anyone on the team can explain why a given profile sees a given version. Leave every delicate signal out of the criteria: health, orientation, economic standing, instantaneous position. Offer a visible control for understanding where a suggestion came from and for returning to the shared version. Measure results group by group, and maintain a baseline identical for everybody, so that support, testing and documentation retain a stable reference.

The questions to put on the table before switching on real-time adaptation

Anyone weighing up a project of this kind can start from three queries. First: does this variation solve a problem observed during research, or does it justify a technology investment? Second: can somebody explain out loud, in thirty seconds, why a particular screen looks the way it does? Third: if a customer asked tomorrow how their data gets handled, would the answer fit on one page or require an internal investigation?

When the replies to those queries come out laboured, the persona route remains the sturdier one. Less spectacular, granted, yet manageable, explainable and open to improvement over time.

Related sources

Chiara Bonifazi