I run Zuidaima (“Most Code”), a Chinese source-code sharing site that launched in 2012. At its peak it felt like a small economy: upload code, earn points, spend points to download, cash out points. About 537k registered users, nearly 10k shared projects, and millions of behavioral events later, search traffic dried up—and so did uploads.

Recently I finally sat down with the production data (registrations, views, downloads, completed payments through mid‑2026) to answer a question I’d been avoiding:

Do people still want downloadable student/side-project code—or did AI make that whole category irrelevant?

Short answer: demand didn’t vanish. It got thinner, more seasonal, and more about “can I ship/demo this?” than “can I download a zip?”

The big picture: not dead, just an order of magnitude smaller

Annual new registrations (selected years):

Year New users vs 2018 peak
2018 87,578 100%
2020 48,114 55%
2022 30,566 35%
2024 7,776 9%
2025 2,124 2.4%

Completed real-money top-ups (unique users who finished WeChat Pay, not just opened a QR code):

  • 2018: ~8.3k
  • 2022: ~2.2k
  • 2025: ~130

Same story for views and downloads. This isn’t a cliff; it’s a long decline. Seasonality is still real—June 2024 spiked vs Aug/Sep, classic graduation / coursework season in China.

If you operate a niche learning marketplace, this pattern may look familiar: SEO-fed UGC economies shrink when the search pipe shrinks.

The funnel for users who joined after 2020 (~149k people)

Step Users % of signups
Viewed a project 117,030 78.5%
Downloaded code 54,837 36.8%
Paid download (spent points on priced projects) 52,640 35.3%
Completed cash top-up 11,088 7.4%
Uploaded a shared project 619 authors 0.4%

How I read this:

  1. Intent is still “find code.” Almost 4 in 5 new users view something. They’re not accidental tourists.
  2. Half bounce between view and download. The unresolved question is fitness + runnability: Is this right for my assignment? Will it start?
  3. Once they download, it’s usually a priced item. The points marketplace still works.
  4. ~7% convert to real money. That’s a durable niche, not a growth rocket. Important caveat: most “payment” events were unpaid QR sessions (status=0). Counting every payment intent as revenue would inflate payers by ~3–4×.

Later cohorts are colder:

  • 2020 cohort: ~43% downloaded, ~8.8% completed a top-up
  • 2024 cohort: ~30% downloaded, ~5.9% topped up
  • 2025 cohort: ~27% downloaded, ~5.6% topped up

What people actually pay for

Track A — cheap frontend “snacks” (volume)

Top paid downloads are still flooded with:

  • HTML/H5 mall & admin templates
  • login pages, food/travel static sites
  • cute animation / confession-page demos

Price band: mostly 1–5 points. High volume, low willingness-to-pay for code itself. This layer is exactly what AI coding tools commoditize first.

Track B — complete coursework systems (value)

What looks closer to “someone will top up for this”:

  • Spring Boot / SSM admin systems (jobs board, bookstore, hotel, property, warehouse)
  • projects that mention docs / thesis-style writeups
  • Spring Boot + Vue full-stack demos

Typical prices: 9–15 points, sometimes more.

Keywords that keep showing up: admin, management, student, mall, Spring Boot, Vue.

What’s almost absent from paid charts

  • Pure ML / “AI project” novelty packs
  • Trendy stacks with weak demo value

Users aren’t shopping for the frontier. They’re shopping for something they can present.

One-line product thesis from the data:

Use cheap templates for acquisition; make money on runnable, documentable full systems.

If you “modernize,” amplify Track B—don’t invent another Track A.

Supply died before demand finished dying

New shared projects / distinct authors:

Year Projects Authors
2017 1,113 655
2020 669 317
2023 171 91
2025 15 14

Among post‑2020 signups, only ~619 people ever became authors.

Classic death spiral:

less traffic → fewer downloads/top-ups → creators stop earning → uploads stop → catalog ages → SEO and conversion get worse.

Ironically, paid charts still feature years-old HTML/JSP/early Spring Boot repos. The inventory still prints some cash. The factory floor is empty.

The underrated job-to-be-done: make it run

Among 2024+ new-user views, a troubleshooting blog post about importing/running downloaded Java projects ranks near the top—above many source packages.

That single ranking told me more than a dozen brainstorms:

  • Friction after download is the product.
  • “One-click preview” and “paid help to deploy” aren’t side quests; users already vote with clicks.
  • In a world where AI can scaffold a Next.js app in minutes, the zip file is less scarce than a working demo and a human (or agent) who unblocks you.

Practical takeaways (for builders of similar products)

  1. Stop worshipping UV. Instrument graduation-season: preview → pay → assisted setup.
  2. Curate 30–50 admin / Spring Boot coursework apps with short-lived Docker previews and runbooks.
  3. Keep cheap HTML templates as bait, not as the P&L center.
  4. Win back creators before chasing new signups. A 7% pay rate with empty shelves is still zero.
  5. If you test a $1–$2 impulse SKU, sell “complete system + docs + run guarantee”, not “another generic starter repo.”

Closing

The old loop was beautiful in its simplicity:

share → earn points → spend points → withdraw.

It worked when search brought endless students and creators could see money move.

What’s broken now isn’t mainly the loop—it’s distribution, delivery (still “here’s a zip”), and creator incentives.

The data doesn’t say “shut it down.” It says:

Evolve from a code download site into a coursework-grade app delivery surface—previewable, explainable, and help-to-run. Keep the points economy if it still clears. Put AI in the middle as infrastructure, not as a chatbot sticker on a dying homepage.

If you run an aging developer community or digital goods marketplace, ask yourself the same question our charts forced on us:

Are your users buying files—or buying a result they can show on Friday?

Method notes (for the pedants)

  • View / download / pay mapped from first-party event types in our legacy Java app.
  • Paid download = non-author download of a project with price > 0 (current price field; historical price edits can add minor noise).
  • Cash payers = payment events in a completed status only.
  • Category labels on titles are rule-based heuristics for trend-reading, not a taxonomy product.

I’d love comments from folks running Gumroad-style code packs, course marketplaces, or any UGC site that peaked on SEO. What does your view→pay funnel look like in 2025–2026?