• Failory
  • Posts
  • From Social Media to Collage Tool

From Social Media to Collage Tool

Inside Zeen’s pivot that pleased creators but stalled scale.

Author

Nicolás Cerdeira
August 14, 2025

Hey — It’s Nico.

Welcome to another Failory edition. This issue takes 5 minutes to read.

If you only have one, here are the 5 most important things:

This Week In Startups

🔗 Resources

📰 News

Threads now has more than 400 million monthly active users.

Nvidia unveils new Cosmos world models, infra for robotics and physical uses.

NASA has sparked a race to develop the data pipeline to Mars.

💸 Fundraising

Datumo raises $15.5M to take on Scale AI, backed by Salesforce,

US rare earth magnet startup raises $65M to scale up production.

Fail(St)ory

Social Media or Creator Tool?

Three weeks ago, Zeen, a startup that began as a social media app and later pivoted into a creator tool, shut down. What’s interesting here is the fork they hit: try to spark a new network or build a focused utility that rides the platforms everyone already uses. 

That choice exposes the real trade-offs—growth loops, funding fit, and who controls distribution. Let’s unpack how the pivot unfolded and what it says about picking between social media dreams and tool reality.

What Was Zeen:

Zeen started life as Landing, a collage app and social media platform launched in 2021. Think mood boards, outfit inspo, and visual scrapbooks—very 2010s Polyvore vibes—but refreshed for Gen-Z. 

It picked up a pocket of enthusiastic creators who used it to assemble digital collages from product links and images, then exported those boards to Instagram, TikTok, and newsletters. 

Soon the team saw a pattern: people wanted to shop from each other’s boards, and fashion content dominated. Collages plus affiliate links looked like a smart wedge— fashion creators could create collages of different outfits, and then monetize through affiliate links.

Meanwhile, building a durable social network was (unsurprisingly) proving to be extremely hard. So the company rebranded to Zeen and leaned fully into utility rather than building yet another social network. 

The product became a creator-side tool for making visual, shoppable collages that travel well across platforms and newsletters.

At its core, Zeen was still a collage tool. The canvas didn’t change—drag items in, arrange, crop, make it look good. But Zeen stripped out the feed and wired the canvas for shopping. Paste a product URL and it pulled the image and details, turned tiles into live links, and auto-generated a tidy product list you could drop into a newsletter. 

@myzeenapp

Turn your email into a magazine with shoppable collages from the Zeen App on Kit #zeen #kit #collage #moodboard #ootd #fyp

Compared with Landing, where the goal was “bring people here,” Zeen’s goal was “help creators publish everywhere and get paid.” That made it especially useful for fashion and lifestyle writers—many on Substack—who ship “shop this look” posts on a schedule. The tool collapsed a messy workflow (screenshots, spreadsheets, link managers, design apps) into one pass: build the board, attach links, export, done. 

However, despite raising $9M and iterating quickly, growth didn’t reach the bar needed for a venture-scale business, and the team chose to wind down. 

The Numbers:

  • 📅 Founded: 2019

  • 💰 Funding: $9M

  • 🧭 First product: Landing (launched 2021)

  • 🔁 Pivot: rebranded to Zeen; shifted from social network to creator utility.

  • 🧷 Monetization: premium tier at ~$6/month plus affiliate-driven use cases.

Reasons for Failure: 

  • VC scale vs. tool reality: A utility that sells for the price of a latte needs either massive user volume or high ARPU (average revenue per user). Zeen’s model sat in the “useful but inexpensive” zone. That’s great for creators; tougher for a venture-backed startup that needs either rapid growth or high margins to satisfy investors.

  • Narrow wedge, crowded neighborhood: The collage-to-shoppable pipeline is a smart wedge, but it lives next to entrenched affiliate platforms and design tools. When your category overlaps with Canva, Link-in-bio stacks, and affiliate dashboards, differentiation must be undeniable or distribution must be unfair. Zeen had a clean workflow, but not a moat.

  • Late pivot, short runway: Landing’s community had some spark, but turning off the network to pursue a tool meant resetting growth while burn and investor expectations stayed high. Pivots buy focus at the cost of time; if the runway is already short, even a smart new direction can’t compound fast enough.

Why It Matters: 

  • A $6/month tool that lives inside other platforms can delight creators yet miss venture-scale math—great product, mismatched capital.

  • Shifting from Landing’s budding social graph to a utility reset compounding growth right when runway mattered most.

  • The “shoppable collage” wedge never became an owned surface (marketplace, proprietary data, or checkout), so value flowed to affiliate networks and the feeds it exported to.

Trend

GPT 5

This week, OpenAI’s long-teased GPT-5 finally arrived.

After more than two years of development (and a couple of missed release windows), the company rolled out what it calls its “most capable model yet.” It’s not the sci-fi leap to human-level intelligence some were hoping for, but it does move the goalposts in important ways.

Six days in, the hype is giving way to actual hands-on use—and the early feedback is… complicated. GPT-5 nails some things it set out to do, but the launch also showed just how hard it is to please a user base that’s already deeply attached to older models.

Why It Matters:

  • Bigger brains, fewer hallucinations: GPT-5 is better at reasoning, coding, and complex problem-solving, while making far fewer factual errors.

  • AI as a work tool, not a toy: OpenAI clearly tuned GPT-5 for professional use—law, logistics, sales, engineering—pushing it further into high-value business tasks.

  • The personality problem: Technical improvements don’t always translate to user love, especially when they come with changes to tone and style.

What’s New

On paper, GPT-5 is a clear step up. It posts state-of-the-art results across core areas that actually matter for work: 94.6% on AIME 2025 (math), 74.9% on SWE-bench Verified and 88% on Aider Polyglot (real-world coding), plus strong marks in multimodal reasoning (MMMU 84.2) and health (HealthBench Hard 46.2). 

One big architectural change: GPT-5 runs as a unified system with a real-time router that decides whether to answer fast in chat mode or slow down and “think” when the task is hard. This means that you no longer need to pick between several different models like you used to before.

In addition to this, GPT 5 is way more reliable. OpenAI reports materially lower hallucinations—about 45% fewer errors than GPT-4o with search, and ~80% fewer than o3 when thinking—plus steep drops on open-ended factuality tests. It’s also less prone to “fake confidence” on impossible tasks. In short: more “I can’t do that, here’s why,” less bluffing. 

Style and control saw tuning too: reduced sycophancy, better instruction following, and responses that feel more like a candid coworker than a people-pleasing assistant. Meanwhile, the model gets more done with less “thinking” output—beating o3 while using 50–80% fewer tokens—and, in internal evaluations of economically valuable knowledge work, meets or beats experts roughly half the time.

The Launch

GPT-5’s rollout wasn’t quiet—it replaced GPT-4o entirely in ChatGPT overnight. For a brief moment, everyone logging in was funneled into the new model, no opt-out, no switch-back. That’s when the cracks showed.

First, the simple issue: the new real-time router—the system meant to decide when to respond fast or think deeply—wasn’t working for a good chunk of launch day. Without it, GPT-5 defaulted to the quicker, lighter reasoning style, making it seem less capable than GPT-4o in complex tasks. Users were quick to call it “dumber” and “sloppier.” OpenAI fixed the bug within hours, but the damage to first impressions was done.

Then came the trickier issue: personality. GPT-5’s tone is noticeably different—more direct, less chatty, more like a sharp coworker than a friendly assistant. For some, that’s exactly what they wanted. For others, it felt colder and less engaging, even if it was technically more accurate. Loyal GPT-4o fans complained they’d lost the “human-like” warmth they enjoyed, framing the change as a downgrade in user experience rather than an upgrade in intelligence.

The backlash got loud enough that OpenAI reversed course, restoring GPT-4o as an option within days. It’s a reminder that in AI, performance gains don’t guarantee adoption. People get used to how a tool behaves, and changing that—no matter how much better it is on paper—comes with risk. GPT-5’s next hurdle isn’t proving it’s smarter, it’s proving it’s the model people actually want to use.

Help Me Improve Failory

How Was Today's Newsletter?

If this issue was a startup, how would you rate it?

Login or Subscribe to participate in polls.

That's all of this edition.

Cheers,

Nico