Adaptive personalisation is the practice of changing what a user sees based on live signals like their behavior, location, device, or past choices. Unlike a static profile the user sets once, it updates itself as the person keeps using the product. It exists because one-size-fits-all interfaces waste people's attention.
Use cases
Get this wrong and you cross the line from helpful to creepy, or you trap users in a bubble they can't escape. The stakes are trust.
The filter bubble complaint: A news app keeps narrowing feeds until users only see one viewpoint. On Reddit and elsewhere, people describe feeling manipulated and delete the app.
The cold-start flop: A new user gets junk recommendations on day one because the system has no data yet, so they bounce before it can learn anything.
The "why did it change?" panic: The homepage reshuffles overnight with no explanation, and loyal users think the app broke.
How it's used in practice
Define which signals you'll act on (clicks, dwell time, location) and write down which ones are off-limits for privacy.
Design a sensible default state for users the system knows nothing about yet.
Add a visible "why am I seeing this?" affordance and an easy way to reset or correct the personalization.
Ship changes gradually and measure them, so a bad model doesn't tank the whole experience at once.
🪄
Pro-tip: Always give users a manual override that outranks the algorithm. The fastest way to lose trust is to keep "helpfully" showing someone something they've explicitly dismissed three times. A visible "not interested" control that actually sticks is worth more than a smarter model.
Challenges & limitations
The creepiness cliff: Too much accuracy feels like surveillance. People get uneasy when a product knows more than they told it.
Cold start: With no history, the system guesses badly, and first impressions are where you lose people.
Hard to debug: When output is personalized per user, "it looks fine on my machine" stops meaning anything, and QA gets messy.
