500px Home Feed
Redesigned the 500px home feed to fix a systemic like inequality problem, surface new talent, and give every photographer a real chance to be discovered.
- 500px's Pulse algorithm was concentrating attention; top photos came from a small, established cohort of photographers
- The majority of users were gravitating toward the same few categories of photos, creating a feedback loop that made everything else invisible
- New users weren't staying because their photos weren't getting attention and they weren't finding content that inspired them
- Used Wizard of Oz prototyping to emulate the ML model before it was built, validating the concept with real users on sacred ground
01
Problem
A like inequality loop was concentrating attention on a narrow slice of content and photographers
02
Approach
Wizard of Oz prototyping to validate feed personalization before any ML investment
03
Design
Redistributed attention, restored map discovery, and surfaced new talent without alienating existing power users
04
Results
Increased Weekly Active Users and improved platform stickiness
01
Problem
A positive feedback loop was making 500px smaller, not bigger.
Photos that reached the top 50 of Pulse, 500px's proprietary quality score, came almost exclusively from a small cohort of established photographers. This wasn't a bug; it was a structural problem baked into how content was being surfaced. Users who found something they liked kept seeing more of it, which meant the same few categories dominated the feed regardless of what else was being uploaded. New users couldn't find content that moved them, and their own photos weren't getting the attention needed to make the platform feel worth staying on. Retention among new users was declining as a direct result.
02
Approach
We validated the concept before building anything.
Because the home feed is the most visited and most contested surface on the platform, we used Wizard of Oz prototyping, manually curating a prototype with photos we knew each user would respond to, to simulate what a personalized ML-powered feed would feel like. This let us test the emotional response and engagement before committing to the engineering investment.
03
Design
Two features came out of the research that hadn't been in the original plan.
Managing change was part of the work. A vocal minority pushed back hard against any feed changes, a classic "Who Moved My Cheese" dynamic. We listened carefully, used the data to separate signal from noise, and held the direction without dismissing the concern.
Photographers consistently flagged that losing the map had been a real loss; location was how they organized their own work and discovered others. We brought it back on rollover rather than making it a persistent UI element. And to address the like inequality problem directly, we adjusted how attention was distributed across the feed, so newer and less prominent photographers had a genuine chance of being seen.
04
Results
Redistributing discovery changed who the platform worked for.
Weekly Active Users increased as more photographers found content that connected with them and felt the platform returning the investment they put into it. Stickiness improved as a result: users who found new talent they cared about had a reason to come back.
