In the Weeds: Print-on-Demand and the Listing Nobody Writes

Open any advice column about Etsy listings and you will be told, within about a paragraph, that the key is keywords. Stuff them in the title. Stuff them in the tags. Stuff them in the first 160 characters so the algorithm sees them before the human does.

This is roughly like being told that the key to a good restaurant menu is to list every ingredient on the front of the building.

There is a listing Etsy wants you to write, full of keywords and redundant phrases and the word "handmade" three times. And then there is the listing nobody writes, the one that actually sells — the one that reads like a human being who makes the thing is standing in front of you explaining, honestly and specifically, what the thing is and why somebody might want it. The second one is harder. The second one is what this piece is about.

I spent an afternoon in the weeds on this, running a composite scenario end to end, watching what AI could actually do for a print-on-demand shop, and more importantly watching what it could not. Here's the ledger.

About the composite. Maya is not a real illustrator. She is a composite I built from three real conversations with illustrators who run print-on-demand storefronts — two on Society6, one on Redbubble, one on a self-hosted Shopify with a POD fulfillment backend. Her portfolio, her backlog, her specific problems are assembled from those conversations. If she sounds familiar, it is because she is the shape of a problem a lot of working illustrators share, not because she is anyone in particular.

Maya's shop, on paper

Forty active designs. Mostly pen-and-ink botanical studies with a color wash, sometimes pushed into more graphic territory — a series of six state birds done as geometric shapes, a series of twelve mushrooms in a style she describes as "if Beatrix Potter had read Kraftwerk." She is a real illustrator doing real work and she is not an idiot about it.

Her shop has not been updated in eight months. Not the art — the listings. The text that accompanies the art. The titles. The descriptions. The tags. The things the algorithm and the human reader will see before they ever see the image, if they see the image at all. Her traffic has been slowly sliding down. Her conversion rate, when she looks, is about 0.8%, which is below where it should be for the niche, and she knows it.

Her question: Can I spend one afternoon with AI and actually fix this, or am I going to end up with forty listings of bland SEO-stuffed slop that I hate?

My question back: Depends entirely on what you tell it to do. Let's find out.

The listing nobody writes — and why

Before I handed anything to the AI, I wanted to establish what "good" actually looked like, because if I couldn't define it, the model was going to optimize for "what an Etsy listing usually looks like," which is exactly the failure mode I was trying to avoid.

Here is what a good listing is, in my opinion, having looked at a lot of them and having bought art from a few:

1. The first line names a specific thing. Not "beautiful wall art print" — "a house finch drawn from a bird feeder in late March, when the red on the head goes almost coral." The first line tells the buyer what they are about to look at, in language that tells them the maker was paying attention.

1. The second line gives them something to imagine. Where this belongs in a room. Who this is for. Not "perfect for any space" — "looks best in a kitchen next to something with wood grain, or above a reading chair in a room where the light is already a little complicated."

1. The middle tells them what they'd actually want to know. Paper stock, size, whether the colors will sit nicely next to their couch, how the print ships, whether they can ask for a different size. The boring-but-important part. Crucially, the boring part is written by a human who respects the buyer, not by a template.

1. The last line closes the loop. Not "add to cart now." Something the maker would actually say if she were handing it to them at a craft fair. "If you want this in a different size or on different paper, just message me — it's a small shop and I can make it work."

Total length: 180 to 300 words, depending on the piece. Tags and metadata: additional, but not jammed into the prose. Title: 60–80 characters, specific enough to be useful, not a keyword pile.

That's the target. Now let's watch the AI miss it.

Round one — the generic prompt

I started the way a lot of shop owners would start, because I wanted to see the failure mode at its purest. I gave Claude one of Maya's designs — the pen-and-ink of the Indigo Bunting, the one she's most proud of — and the most common prompt I see floating around in Etsy advice circles:

"Write an Etsy listing for this print. Make it SEO-friendly and optimize for search."

The draft came back immediately. 220 words. Every single one of them bad.

It started with "Enhance your space with this stunning wall art print." It used the word "perfect" four times. It mentioned "home decor" in the second sentence. It claimed the print was "the perfect gift for bird lovers, nature enthusiasts, and anyone who appreciates fine art." It offered the print "for your living room, bedroom, office, or any space that needs a touch of natural beauty."

This is not a listing. This is the algorithm's hallucination of a listing. It is what happens when you ask a model to write in a genre that is already mostly filler. The model has read ten million of these. It knows exactly what the average one looks like. Asking it to write another average one is a task it can complete in three seconds, and the completion is worthless.

What went wrong, specifically — and it's worth naming this because every print-on-demand operator is going to hit it the first time they try this:

• The prompt asked for "SEO-friendly" and the model interpreted that as "keyword-stuffed." Those are not the same thing, but the model does not know that. You have to tell it.
• The prompt did not give the model anything specific about the actual print. So the model wrote about the generic category of "bird art print," not about this particular bird in this particular painting.
• The prompt did not define the audience. So the model wrote for everyone, which means it wrote for no one.
• The prompt did not give the model a voice to match. So it wrote in the default voice of the listing corpus, which is the voice of a tired marketing intern.

Four problems, one prompt. Let's fix them.

Round two — the specific prompt

For the second attempt, I had Maya sit down with me and describe the Indigo Bunting painting for sixty seconds. Not to write a listing. Just to describe it to me like I was a friend who had never seen it.

What she said, roughly: "It's from a bird I saw outside my studio in Ithaca in May of 2022. It was on a pokeweed stem and the light was going. The blue on an Indigo Bunting is not actually blue — it's refracted, like a butterfly wing — so in the painting I tried to show that by leaving the wing pale and letting the ink go darker on the body. It's smaller than most of my bird prints because I wanted it to feel like a moment, not a portrait. The paper I print it on has a slight warm tint so the blue reads even more strangely."

Everything in that sixty seconds is unusable by a generic prompt and essential to a good listing.

I wrote her a better prompt. Not a one-liner — a paragraph. Here, roughly, is what it said:

You are an honest, specific copywriter for a small art shop. Write a product listing for one print based on the notes below. The listing should be 200–260 words. The first sentence should name the subject of the print and one specific observation the artist made about it. The second sentence should suggest one specific place in a home where this would belong. The middle should explain paper, size, and shipping in plain language. The closing should invite the buyer to message the artist directly if they want a different size or paper. Do not use the words "stunning," "perfect," "enhance," "home decor," "any space," or "nature enthusiasts." Do not repeat keywords. Do not list the print as a "perfect gift." Write in the voice of someone who actually made the thing and is tired of seeing her work described in marketing language.

Then I pasted Maya's sixty-second description underneath.

The draft came back much better. Not perfect — we'll get to the failures — but recognizably a listing a buyer could trust. It started with the pokeweed stem. It mentioned the refracted-blue observation, slightly clumsily. It suggested the print "would sit well above a small table, somewhere the light is already interesting." It described the paper stock specifically. It closed with "this is a small shop — message me if you want a different size, I can usually make it work within a week."

Maya read it out loud and said, "That's 70% of a listing I would actually publish."

70% is not 100%. 70% is the number that matters in this piece.

Where the model was wrong, specifically

Here is what the specific prompt produced that still needed Maya's hand to fix. I am walking through this in detail because anyone running a POD shop needs to know exactly where the model's default draft will betray them.

The over-polish problem. The model wrote "the refracted blue of the Indigo Bunting's wing, which is not truly blue but a trick of light, like a butterfly." That's fine. It is also 30% more elegant than Maya's original observation, and in a way that removes Maya from her own sentence. She would never say "a trick of light, like a butterfly." She would say "it's the same thing that happens with butterflies, where the blue isn't really there — it's just how the feather bends the light." Clunkier. More specific. More her.

The fix: I added to the prompt, "Keep the observations slightly clunky. Do not smooth them out. The buyer should feel like the artist is explaining the painting, not a copywriter is summarizing it."

The "would sit well" problem. The model's default for "where this belongs in a home" is pure real-estate-listing language. "Would sit well above a small table." "Complements a quiet reading nook." "Brightens an entryway." All of these are slightly correct and completely interchangeable. You could swap them across any print in Maya's shop and nothing would break.

The fix: I had the model name a specific thing in the room — a specific color, a specific piece of furniture, a specific time of day. "This sits well next to something wooden, especially in the afternoon when the light is already a little yellow." Better. Not great. Maya rewrote it to: "Put this somewhere the light gets complicated in the afternoons. It does its best work between three and five." That is her sentence. The model could not write that sentence. It could only write toward it.

The shipping-and-paper problem. The model is actually fine at this. It will tell the buyer what the print is on, how big it is, how it ships, what the turnaround is. The catch is that it will write this in the same voice as the rest of the listing, and the rest of the listing is supposed to sound like Maya, and Maya does not actually talk about shipping with the same energy she talks about birds. The shipping paragraph should be a little drier, a little more matter-of-fact, a little more "here are the facts you need." The model defaulted to a kind of uniform warmth across the whole listing, which is actually wrong — a good listing changes temperature between the art part and the logistics part.

The fix: I told the model to write the shipping paragraph in a slightly flatter voice than the rest of the listing, "like the artist is answering a practical question honestly." It helped.

The closing CTA problem. This is the one I'd warn every POD operator about specifically. The model's default close on an art listing is "Thank you for supporting independent art!" It is, I am sorry, an abomination. It is a sentence I see at the bottom of 40% of Etsy art listings and it is doing nothing, because the buyer knows they are supporting independent art. They are on Etsy. They paid the Etsy price. They do not need to be reminded that their purchase is good for the artist.

What the closing is for is to tell the buyer one useful thing that will make them more likely to click buy. Usually that is: you can message me if you want a different version, and I am small enough that I will actually read your message and answer it. That is a real piece of information that real buyers use. "Thank you for supporting independent art" is not.

I told the model, explicitly, never to use that sentence or any variant of it. I told it the closing should offer one specific accommodation the buyer could take the shop up on. Fixed.

Round three — forty listings in an afternoon

Once the prompt was tuned, Maya and I ran all forty of her designs through it. Here's how that actually went, which is not quite as clean as the walkthroughs you'll see in the "I automated my shop" blog posts.

Pace. A good listing, with Maya's sixty-second description provided to the model, took about five minutes total — roughly one minute for her to describe the piece out loud and sketch her observation, two minutes for the model to draft, and two minutes for her to edit the draft into something she'd actually publish. Forty listings at five minutes each is about three and a half hours of real work, with breaks.

Batching. We tried running three listings in parallel. Bad idea. The model started blurring them — using the same "where this belongs" observation across two different prints, falling into a rhythm where the third listing was a slightly softer copy of the first. Batching is the fastest way to get from 70% to 55%. We went back to one at a time.

The dictation step. The single biggest time saver was not the model. It was the fact that Maya recorded herself describing each piece, on her phone, for about sixty seconds each. She did all forty in one forty-minute session, walking around her studio with the phone. Then we transcribed them and handed them one at a time to the prompt. The dictation was doing the creative work. The model was doing the formatting. That is the correct division of labor for this task, and it took us the whole afternoon to figure it out.

The seven listings that broke. Of the forty, seven came back in a way Maya hated — too flat, too generic, the model having latched onto an unremarkable phrase in her description and built the whole listing around it. For those seven, we re-ran with a slightly tighter prompt that quoted one specific sentence from her description back to the model as the required opening. All seven got usable on the second try.

The two listings she scrapped. Two of her older designs were not, on review, designs she loved anymore, and it became obvious once we sat down to describe them that she had nothing specific to say about them. The honest move was to retire them from the shop. AI did not cause this — sitting down with the shop for an afternoon caused this — but the AI-assisted process created the conditions where she finally looked hard enough to notice. Call this an indirect benefit.

At the end of the afternoon, she had 38 published listings, each of which read like a small paragraph from a real person about a real thing, each of which was distinctive enough that a buyer scrolling through could have picked a favorite, each of which closed with an invitation she actually meant.

The traffic, and the honest caveat

I cannot tell you how much the new listings moved her conversion rate, because this is a composite and the conversion rate is not a real number. What I can tell you, from the real operators I built the composite from, is that the pattern is consistent: specificity in listing copy produces a real lift in conversion, usually in the 0.3–0.8 percentage-point range on shops in Maya's size bracket. That sounds small. On 40 designs with a few thousand views a month per design, it is not small.

But here is the caveat the optimistic version of this piece would skip: the lift comes from the specificity, not from the fact that AI wrote the listing. The AI is a writing accelerator for specificity you already have. If you sit down and cannot say anything specific about your own work, the AI cannot help you. It will produce the generic version, faster, at scale, with slightly better grammar.

If you have been running a POD shop for a while and the work no longer excites you enough to describe it out loud, that is not a listing problem. That is a separate problem the AI cannot touch, and I want to be honest that I have watched it happen to real illustrators, and the fix is not a prompt.

Tools in the workflow, specifically

Three things I used during the afternoon, each of which earned its place:

• 🛍️Print-on-Demand Listing Copy — the prompt this piece is built around. Not the generic one. The tuned one, with the banned words and the voice constraints. I'd treat this as the starting point, not the finished product. You will still tune it for your voice. That is a feature.
• 🎨The Freelance Art Director — for the two designs where Maya wasn't sure if they still belonged in the shop. Not for the listings themselves, but for the honest conversation about whether a piece is earning its place in the catalog. That conversation is uncomfortable. The soul is good at it without being mean.
• ✍️The Plain-Spoken Copy Editor — for the second pass, the "does this sentence actually say anything" pass. I ran the final listings through this soul and it caught three places where the model had smuggled a filler sentence back in. Catching those is what that soul exists for.

Those three, plus the dictation, plus the afternoon, plus the forty designs. That was the whole stack.

The three principles that survive the session

If I compress everything the afternoon taught me into three things I'd tell another illustrator in Maya's position, it's these.

One. You cannot prompt your way out of having nothing to say. The best listings come from the sixty-second description. If you cannot do that out loud, the model is not going to do it for you. Do the dictation step. It is the single most important part of the workflow.

Two. The model's default is slightly worse than yours, and that's okay if you are willing to push back. This is the pattern that shows up again and again with AI-assisted creative work. The first draft strips specificity. The second draft, after you have pushed back with "keep it clunky, keep it specific, name the pokeweed stem," is better. The third draft is usually the one. If you're not willing to run three drafts, don't bother running one.

Three. The listing is the smallest unit of respect you offer your buyer. A listing full of keyword mush tells the buyer you do not expect them to read. A listing that names a specific observation about a specific bird on a specific evening tells the buyer you trust them to care about the same things you care about. The second listing converts better not because it is "optimized" but because it is a small act of respect, and respect is expensive to fake and cheap to do honestly.

Eight months of not touching her shop, forty designs, one afternoon. Thirty-eight listings Maya would stand behind, two retired pieces she would not. The indigo bunting still on the pokeweed stem, still in the window between three and five. The phone in her pocket with forty sixty-second recordings on it, most of which she is going to delete now because their job is done. A shop that reads, when you scroll through it, like somebody made it.

The listing nobody writes is not nobody's fault. It is what happens when the task looks boring enough that everybody — including the AI — wants to do it fast. Slow it down. Do the dictation. Use the model for the formatting, not the looking. The looking is still yours.

That is the whole practice, and it is one afternoon of work, and if you run a print-on-demand shop and you have not done it in a while, I would go do it now.

You can find more working-illustrator tools in the catalog at a-gnt. Start with the two souls above. They will earn their place in an afternoon.