Good Product, Not a Business: A Clothes Box Monetization Postmortem

This post was originally written in Chinese. The original version is here.

$34.

That's the total affiliate commission I received after running the shopping mode in my AI Stylist app Clothes Box for 40 days. 6,000 users, 40,000+ product recommendations, 653 clicks, 12 net orders, $691 in GMV. My cut was $34.

Clothes Box is a mobile app that combines a digital wardrobe with an AI stylist. Users photograph their clothes to build a digital closet, and the AI stylist suggests daily outfits based on their schedule and weather using clothes they already own. Because we have ground truth on what's actually in the user's wardrobe, we can also make context-aware shopping recommendations: if the AI spots a gap in your wardrobe, it recommends a product to fill it and generates a virtual try-on image showing you wearing that new item with your existing clothes. (A more complete product description is in the appendix.)

Five months into building this product, this post is a postmortem on why the monetization strategy failed. Specifically: how I chose a business model that seemed smarter than subscriptions, what data it actually produced, and why the approach is structurally broken.

1. Why Not Charge a Subscription

Before building, I researched 40+ digital wardrobe apps on the App Store and spoke with 4-5 of their founders.

The picture was clear: everyone charges subscriptions at $5-10/month, but conversion rates are painfully low.

One founder's numbers stuck with me. Their product was fairly standard, a digital wardrobe with virtual try-on. Their co-founder had 1M followers on TikTok, and TikTok organic growth alone brought 200k downloads in 9 months. Even with this advantage, they kept their subscription price low at $2-4/month to test willingness to pay. The paid conversion rate barely crossed 1%.

1M followers of free traffic. A rock-bottom $2-4 price point. Still only 1% of users willing to pay.

The problem isn't pricing, and it isn't user acquisition. Users in this category simply don't see enough value in a wardrobe management tool to pay for it. If I went with subscriptions, I'd likely face the same outcome.

So I decided to try a different path.

2. Affiliate Commerce as the Business Model

My logic was this: if users won't pay for wardrobe management, make everything free to lower the acquisition barrier. Then leverage the wardrobe data to make precise clothing recommendations and earn affiliate commissions.

This was theoretically coherent. When talking to investors, I pitched it like this:

Because we know exactly what clothes users own, we can infer their style preferences and make shopping recommendations grounded in the context of their actual wardrobe.

Sounded compelling.

The core assumption: because I have complete wardrobe data that no other e-commerce platform has, my recommendation conversion rates should be far higher than generic ads. High conversion rates plus free-tool user acquisition should work better than trying to charge people who don't want to pay.

Time to test that assumption.

3. 40 Days of End-to-End Data

After the app launched in late March, I integrated Sovrn (a sub-affiliate network) as the product catalog for shopping mode. Shopping mode had two entry points: a conversational mode where the AI Stylist detects shopping intent during chat, extracts keywords, queries the Sovrn API for relevant products, reranks them, and presents recommendations; and a browse mode where users can scroll through recommendations like an e-commerce site, with products curated based on their wardrobe data and style preferences. In both modes, some products get an additional virtual try-on (VTO) image generated, showing the user wearing the new item paired with their existing clothes.

Here's the complete data from the 40 days shopping mode was live.

Core User Metrics

Users are distributed globally, but shopping mode only targeted US users (the affiliate network only covers US merchants). About 800 users were in the US. I feature-flagged shopping mode to 500 of them for testing.

Affiliate Funnel

Key derived metrics:

• Average order value: $691 / 12 = $57.6
• Commission per net order: $2.83
• Affiliate revenue per shopping mode user over 40 days: $34 / 500 = $0.068
• Annualized ARPU (per shopping mode user): $0.62/year, roughly $0.05/month

$0.05 per month. That's what each shopping mode user was worth.

4. Four Structural Problems

Looking back, this model has four structural problems.

Intent Mismatch

Users open Clothes Box to figure out "what should I wear today," not to buy new clothes.

This is the most fundamental issue. Out of 40,000 product recommendations, only 1.6% were clicked. 94 users clicked at least one product link, which looks decent (41% of 227), but spread over 40 days, that's an average of 7 clicks per person total. Given that users only open the app 1.6 times per week for a total of 5-11 minutes, the window for reaching users is inherently narrow.

Affiliate is fundamentally a volume game. It needs high-frequency, high-volume exposure to drive enough purchases. Users spend tens of minutes to hours per day on Instagram and TikTok. The ad exposure frequency on those platforms and a utility app opened 1.6 times per week aren't even in the same league.

I was trying to use a wardrobe management tool to create purchase intent. But when users come to manage their wardrobe, they're thinking "what should I wear," not "what should I buy." Even the most precise recommendations have limited effect when shown to people who aren't in a shopping mindset.

What if users could form the habit of coming to Clothes Box when they want to buy clothes, instead of going to Google Shopping or brand websites?

That runs into the next problem.

Incomplete Catalog

This problem first surfaced in a conversation with a user. I asked her if she was satisfied with shopping mode and whether she'd come here to shop for clothes in the future. She said the recommendations were decent, but her favorite brands were missing. She specifically asked the AI to recommend Anthropologie items, and the AI came back with substitutes from Macy's.

It wasn't a recommendation algorithm problem. Sovrn's catalog simply doesn't carry Anthropologie.

To fix this, I later added Serper.dev's Google Shopping search as supplementary retrieval. Serper can cover brands and merchants that Sovrn can't, returning product titles, prices, images, and links. But Serper doesn't handle affiliate monetization, so I built an affiliate link router: if a merchant is covered by Sovrn, the link gets converted to an affiliate link; if not, the original link is preserved.

The result: to keep the user experience from falling apart, I had to recommend products I couldn't monetize. User experience and monetization were in direct conflict. This problem is unsolvable because many brands (especially those positioned as premium) consider affiliate commerce dilutive to their brand value and refuse to participate. No affiliate network can cover every brand.

Brian Sugar, former CEO of POPSUGAR, wrote a piece about AI shopping that nails this: "In fashion, if something is missing, everything feels broken."

The VTO Paradox

Virtual try-on (VTO) was the magic feature I offered users: instead of showing generic model photos that look nothing like the user, rendering an image of the user actually wearing the new item.

The data shows VTO works:

VTO boosted click rates by 2.3x.

But look at the economics:

VTO costs $1.40 per click and earns $0.052. Cost is 27x the revenue.

This is the paradox: the magic moment I give users is also my biggest cost center. The better VTO works and the more users engage with it, the more money I lose. But without VTO, my shopping recommendations are indistinguishable from any generic affiliate product listing.

The source of differentiation and the source of losses are the same thing.

Attribution Black Box

653 product link clicks. Only 16 orders attributed to me by Sovrn.

A 2.4% click-to-order rate looks normal for the affiliate industry (typically 2-5%). But I have no way of knowing how many actual conversions there were.

Here's the scenario: a user clicks my affiliate link, lands on the merchant's site, browses for a bit, and gets ready to check out (or leaves it in their cart for a few days). Then Honey or Capital One Shopping pops up with a coupon, overwriting the cookie. The commission for that transaction goes to Honey / Capital One Shopping. I get nothing.

Brian Sugar described this same dynamic: "A shopper discovers a dress on Daydream, clicks through to buy it, but Honey's browser extension hijacks the final commission. Daydream gets nothing for the discovery work. This scenario plays out millions of times daily."

Beyond my 16 attributed orders, there may have been more conversions that were intercepted. Or maybe 16 was the real number. I'll never know. This opacity is a structural flaw of the affiliate model: you can't optimize a funnel you can't fully see.

5. Affiliate vs. Subscription: The Math

Putting affiliate and subscription side by side.

First, the same 500 users (subscription estimated at 1% conversion, $7/month, 30% Apple cut):

At 500 users, they're nearly identical. Affiliate doesn't look bad.

But there's a critical difference: affiliate can only monetize US users (the affiliate network only covers US merchants), while subscriptions can monetize users globally. Scaled to the full user base:

Subscription revenue is 7x that of affiliate. The gap isn't in per-user value; it's that the monetizable user base differs by 7.5x.

I chose affiliate to avoid the "users won't pay" problem, but affiliate immediately excluded 85% of non-US users from monetization.

Even worse, the affiliate revenue above doesn't account for VTO costs. Including VTO's $400 over 40 days, the affiliate model's net revenue is negative $360. Even without VTO, relying only on organic clicks from non-VTO products, the 40-day commission would be roughly $19.

No matter how you run the numbers, the math doesn't work.

6. It's Not Just My Problem

After running this 40-day test, I came across the Brian Sugar article I've been citing throughout. He lived through the same thing, just at a much larger scale. POPSUGAR's subsidiary ShopStyle ran the exact same fashion discovery + affiliate model before being sold to Rakuten in 2017.

He lists a string of names in the article: Polyvore, ShopStyle, Lyst, Wanelo, Fancy, Pose, Spring. All fashion discovery companies from the past decade-plus that tried the affiliate model. Most people have already forgotten them. The new wave includes Daydream ($50M seed round), Gensmo ($60M seed round), and Alta ($11M). The technology is better now with LLMs and virtual try-on, but the structural problems he identifies are identical to a decade ago: incomplete catalogs, hijacked attribution, and the asymmetric competition between standalone platforms and Google/Instagram for user attention.

His core thesis: "AI will not fix the business model." Better AI can't fix broken business model economics.

My first-hand data from Clothes Box validates this once more.

Closing

Shopping mode has been shut down.

The premise of a growth flywheel is that each acquired user generates enough value to cover the cost of acquisition. Since the monetization model doesn't work, user LTV can't support it, which means paid acquisition and advertising are off the table.

Clothes Box currently has a 4.9/5 rating on the App Store. Users genuinely like the product. It solves the "what's in my closet" and "what should I wear today" problem. I plan to keep maintaining it as a useful digital wardrobe and AI stylist tool.

But it's not a venture-scalable business.

I haven't decided on the exact next step. Possibly reverting to a subscription model to cover the app's operating costs and make a little money each month.

Good product, not a business. (The appendix has a side-by-side comparison of pre-launch estimates vs. actual results that shows just how far off the original projections were.)

Not the conclusion I was hoping for. But a negative conclusion is still a conclusion. Time to move on to the next thing.

Appendix

Clothes Box: Product Overview

The consumer experience around getting dressed and buying clothes is fragmented. People constantly complain that their closets are overflowing yet they have "nothing to wear." Meanwhile, US fashion e-commerce is extremely fragmented: you buy a sweater from J.Crew, jeans from Levi's, and these brands are data silos. The current shopping experience is stateless: e-commerce platforms don't know your personal style, let alone what's already in your closet.

Clothes Box was built to solve this problem. In short, it's an AI personal stylist that knows your wardrobe.

The product has three layers:

Layer one: digitization. Users photograph their clothes, and the app automatically processes images, removes backgrounds, and tags attributes (category, color, season, occasion), dramatically lowering the barrier to building a digital wardrobe.

Layer two: daily styling. The AI uses your calendar and weather to suggest outfit combinations from clothes you already own. It's not recommending you buy new things; it's telling you "today you can wear that black jacket from last year with those blue jeans."

Layer three: Shopping Mode. Because we have ground truth on the user's wardrobe, shopping recommendations can be context-aware. Users can chat with the AI to find clothes or browse recommendations like scrolling an e-commerce site. Recommendations aren't random; they're based on what you already own to suggest items that "complete an outfit" or match your style preferences.

Another highlight of shopping mode is how products are displayed. For the past decade, online shopping has meant looking at studio photos of models who look nothing like you, and mentally projecting how the clothes would look on your body, paired with your own pants. In Clothes Box, recommended products are rendered as images of you wearing the new item, styled with clothes already in your closet. What you see is what you get.

On the business model: all features are completely free. No subscription fees. Revenue comes from affiliate commissions when users purchase recommended clothing through Affiliate Network links.

This is what the article is about: why that business model didn't work.

A. Affiliate Funnel Detail

B. VTO Effectiveness Comparison

C. User Retention

D. Revenue Comparison: Two Models (Monthly Estimates)

E. Pre-Launch Estimate vs. Actual Results

Before building Clothes Box, I estimated per-user affiliate revenue using industry data. The reasoning went like this:

$10/year/user, or $0.83/month. For comparison, the subscription model: at 1% conversion, $7/month, minus Apple's 30% cut, the all-user ARPU is $0.59/year (1% × $7 × 12 × 70%). Affiliate at $10/year looked 17x better than subscription at $0.59/year. The choice seemed obvious.

The actual data:

The estimate was off by 16x.

The core error was the assumption that "30% of online apparel spend would go through Clothes Box." That assumption implies users spending $202/year through the platform, roughly $17/month on clothes. Doesn't sound like much, but it requires users to actually treat Clothes Box as a regular shopping destination, not just a wardrobe management tool.

The reality: users happily use Clothes Box to manage their wardrobe and get outfit suggestions, but when it's time to buy clothes, they still go to Google, brand websites, or links they find on Instagram. Having the user's wardrobe data didn't change their shopping habits.