El Confidencial's Innovation team set out to test a new way to consume the news: a short AI-generated audio summary, embedded inside the article, that a subscriber could play instead of reading. I designed the component end to end. In the first proof of concept the result was good but improvable: almost everyone who pressed play finished the summary, and they spent more time on the site, but fewer people pressed play than we'd hoped. So I iterated the design through an A/B test aimed at making the player easier to find.
Reading is the only way into a news article, and it shuts out anyone who can't look at a screen right then: commuting, doing chores, multitasking. The team's bet was an audio summary, generated automatically from the article text and short by design (36–46 seconds), playing from a small player embedded in the piece. I designed that player — controls, states, how it sat inside the article — and we launched it to a group of subscribers.
The first proof of concept ran as an A/B test, and it validated the format: 84.5% of everyone who started listening finished the whole summary, and the group with audio spent 18.7% more time on the site per session than the control group. The number left to improve was the play rate, 0.85% — the people who pressed play stayed, but not enough were pressing it. A low click rate on a component people clearly like points at one thing: visibility. The player blended into the article, so the lever was its design — making it register as something worth pressing.
I designed four new takes on the player, each pushing visibility a different way — all coherent with the design system and all clearly readable as audio (never mistakable for an ad). To keep the A/B test simple, we picked the strongest one — the dark player — and ran it head-to-head against the original.
Press play and the player expands into full controls — progress, time, and playback speed. Keep scrolling into the article and it collapses into a floating bar pinned to the screen, so you can read on without losing the audio or the controls.
The test is paused while one technical limitation gets solved before a wider rollout: the audio loaded with the article rather than only on play, which made it costly to serve at scale. The fix is scoped, and the direction the data points to next is commercial — a short summary for everyone, with the full audio behind a subscription CTA, built on the redesigned component.
I designed an AI audio component, saw the format validated by real behavior — 84.5% completion and +18.7% session time — and then treated its weakest number as a design problem. A visibility hypothesis, four new versions, and a head-to-head test against the original turned that into a measurable win: +56% click-through over the original. A clear design decision, backed by data.