Print-on-demand was the one stream I thought I understood before we started.
I'd seen the tutorials. I understood the business model. You design a product — a t-shirt, a mug, a tote bag — upload it to a platform like Printify, connect it to a storefront, and when someone orders it, the fulfillment company prints and ships directly to the customer. No inventory, no warehouse, no shipping logistics. You collect the margin between what the customer pays and what the platform charges for production and shipping.
It's clean in theory. In practice it's more complicated, and the places where it got complicated were not the places I expected.
The setup: two stores, two different positioning strategies.
We're testing two print-on-demand stores currently. KHD and EC are the internal designations — both are Printify-based, connected through standard e-commerce infrastructure, with AI-assisted design generation and product copy.
The positioning strategies differ. One leans into a specific aesthetic and community identity — the kind of store where the buyer isn't just buying a product but buying an affiliation with a worldview or subculture. The other is more trend-responsive — watching what's performing on platforms, generating designs in that direction, and seeing what sticks.
What we expected: the trend-responsive store to get faster early traction. What we got: the opposite.
The design problem I didn't anticipate.
AI image generation has gotten very good. I want to be clear about that before I describe where it fell short. For our directories, for the newsletter, for conceptual visuals — AI-generated imagery is genuinely useful and saves a real amount of time and money.
For print-on-demand, it's more complicated. The issue isn't quality in the abstract. It's quality at the file specifications print production requires. Print-on-demand platforms typically want high-resolution files — 300 DPI, specific color profiles, clean edges on designs that will be placed on a substrate with a particular texture and lighting.
AI image generators produce images optimized for screen rendering, not print production. The workflows to get from a Midjourney output to a print-ready file involve more steps, more judgment, and more technical knowledge than I had when we started. We've since figured out a workable pipeline, but it added time and friction that wasn't in my original estimate.
The traffic problem is the real problem.
Here's what the tutorials tend to gloss over: print-on-demand is a demand fulfillment business, not a demand generation business. The platform handles the production. You handle getting people to your storefront in the first place.
With no existing audience and no advertising budget allocated to these stores in the early testing phase, traffic was predictably slow. I understood this conceptually. I did not fully internalize what it feels like to have a well-built storefront with genuinely good products sitting at zero traffic for weeks while you figure out the acquisition question.
The organic search strategy takes time to build. The social media strategy requires consistency and platform-specific skills. The paid advertising strategy requires capital and testing cycles. None of this is a surprise in retrospect — all of it was more viscerally real than I anticipated.
What we're still testing.
The community-identity store is showing more organic engagement than the trend-responsive store, which suggests the positioning thesis there has something to it. When a design connects with a community, the sharing behavior is different than when someone just likes a t-shirt. That's a meaningful signal worth pursuing.
We've revised the design pipeline to handle the print-production specifications more systematically. What was taking four or five iterations now takes one or two.
We've also been more deliberate about connecting the POD stores to the broader content and newsletter operation. A design that emerges from the community documented in this newsletter has a built-in context and audience that a standalone store doesn't have. We're testing whether that integration changes the conversion dynamic.
I'd say we are not yet at the point where I'd call POD a validated stream. We are at the point where I'd say the thesis remains plausible but the path is longer than the tutorials suggested.
Which, honestly, is how most worthwhile things turn out.
This Week in AI: The generative image space has seen significant improvements in prompt-following accuracy over the past few months — models are getting better at producing exactly what you describe rather than an interesting interpretation of what you described. For applications like POD design where precision matters, this is meaningful progress. The print-production workflow issues I described above are a technical pipeline problem, not a model quality problem. Those are solvable with the right tools.
Want to follow along with both POD experiments as the data develops? Join The Upstream community — we go deeper in there than we do in the newsletter. Find the link at start.tenstreamslab.com.