for Headlait
Headlait is an AI-based personalized news recommendation service. Pre-A funding raised: 3 billion KRW. The question the team was trying to answer: why weren't users coming back? The app had content and a working recommendation system, but retention was not moving the way the numbers needed to move for a pre-Series A company trying to prove its model.
The team did not start with a solution. They started with a signal. As the sole designer on a team of approximately 10 people, the role spanned all UX, UI, and the logic connecting them across every screen that shipped, collaborating with PO and developers in the same room, where design was not upstream from decisions but part of making them.
Group interviews, 1:1 interviews, and survey data. Four user groups identified:
Key pain points: information overload, irrelevant content, bias concerns, difficulty finding related/deeper articles, stressful comment sections.
Hypothesis: Personalizing the category tab order would increase user Engagement.
Result: Inconclusive. Provided convenience to ~20% of users already using tabs; no measurable effect on non-tab users. Possible negative impact on session time for diverse-interest users.
Hypothesis: Expanding daily circulated news from 3,000 to 10,000 would increase session time.
Result: Circulating more mainstream media news did not significantly impact product value. Monthly average grew 3,000 → 7,000 items; retention (65–75%) did not improve meaningfully.
Hypothesis: Adding IT, culture, entertainment content beyond standard news categories would broaden user choice and improve session time.
Mainstream news only: session 476s / D1 retention 34.3%
With diverse topic content: session 508s (+6.7%) / D1 retention 36.4% (+5.7%)
~300 diverse-topic articles daily (3% of total) = significant 6% improvement in both retention and session time.
Observation: Users read an article, returned to the list, and left, not because they were done, but because nothing was pulling them forward.
Hypothesis: Surfacing related news at the bottom of the article detail page would keep the reading loop alive.
Session time: +5.0%
Retention D4–D7: +2.3%
Views: +5.5%
Time per article: +1.7%
Personalization alone did not fix retention. Content volume did not either. What the data kept pointing toward was a more fundamental question about the reading experience itself. The problem was not completely solved, but a methodology was built: ship, measure, learn, redesign. In a pre-Series A environment, that discipline is the product.
Thanks for Watching
Back to work