We spent eight months teaching AI to write ecom copy. Here’s what we think.
Our studio writes ecommerce product copy. Lots of it. Like, tens of thousands of product descriptions a year, mostly for sports apparel brands. I have been a managing editor at Thread for almost six years, so I have had a whole lot of product copy cross my desk. We write about fishing poles, hiking boots, training tights, soccer cleats, coveralls, puffy coats, baby onesies, cycling onesies and one-piece snowsuits.
We have a rock-solid team of human writers here. We’re lucky to have among them Liz, who writes eloquently and authentically about global soccer products and culture; Elizabeth, who creates inspiring training and artful swim copy; Dave, who crafts outdoor copy with decades of experience trekking, skiing, mountain biking and more behind him; and Sarah, who sends in chatty, witty and current style copy. And many others in the mix, all of whom consistently hit a deceptively challenging target: short, meaningful, compelling copy that sells products.
It takes 300 characters (or less) to sell a product
To a consumer, product copy is just a short paragraph of text describing a shirt or pair of boots. The copy we write is always under 300 characters long. The task seems straightforward enough for AI, but as any brand or product copywriter knows, it’s not that simple.
A long time ago, I worked for HarperCollins Children’s Books. We’d get hundreds of submissions a year of very short manuscripts that were not publishable. Reading Goodnight Moon, it turns out, is a misleadingly easy lesson about writing a children’s picture book. It’s hard to get one sentence or less per page to add up to a childhood classic.
In most kinds of writing, shorter is nearly always harder. And the same is true of product copy. There are a lot of benchmarks to hit in a short amount of time.
Effective ecom copy always hits three goals
Thread’s decade-plus of product copywriting experience has taught us three must-haves for ecom copy:
- Accuracy: This means we have described the facts of the product correctly. This is obviously necessary for many reasons, not the least of which is that a consumer will return a product if what’s online doesn’t match what arrives in the mail. It’s also essential to build trust and earn repeat customers.
- Brand voice: This means we’ve aligned our copy with a brand’s overall guidelines for voice. This is also important for consumer loyalty. Brands build trust with their consumers when their voice is consistent wherever a consumer encounters them.
- Clarity: Clarity means it’s meaningful, easy to read and makes sense. Empty content — like technical jargon, marketing-type clichés or too much repetition — is off-putting to consumers.
Call it the “ABC Rule.” These three qualities (accuracy, brand voice, clarity) are a must for anything we deliver to a client. But when you’re writing at a massive scale, like we are, efficiency becomes a greater and greater challenge.
How does an AI writer compare to a human writer?
Anyone in any industry working at scale is looking at artificial intelligence, and we’re no different. The future of ecom copywriting, like so many production-focused industries, will likely involve AI in some fashion. As a studio that’s looking out for the needs of our clients and the future of our business, we’re very interested in these new technologies. And we’ve been following along with fascination.
AI is already writing more than you may realize. In recent years, AI has been gainfully employed as a journalist, and has taken a moderately successful turn as road trip scribe.
Across more than 10 years of writing ecom copy, we’ve learned that our formula (short length + accuracy/brand voice/clarity x quantity) works well, but it takes time. It adds up to many months for our team to write about thousands of products for a big client. And our writers typically create a lot of very similar copy each year. Like most companies, we’re forever asking ourselves, “Can we be faster and more efficient? Can we save our human writers time and save our clients money in the process?”
Enter AI. Or, as we’ve affectionately named our AI tool, Rosie.
The best thing about Rosie? She’s a (really) fast writer
We named Rosie after the robotic maid from the Jetsons. Rosie writes quickly. Very quickly. Rosie can write 300 characters, the average length of our product copy, in less than 20 seconds. That’s roughly 50 times faster than our human writers can. Imagine what that could mean for a brand that has hundreds or thousands of products to generate copy for.
Rosie can rewrite quickly, too. At the push of a button, she will generate all-new copy for a product, different from the copy she created previously.
This can really come in handy. Think of it this way: If a human writer makes an error that I, the editor, need to correct, then I have to tell them why I need a rewrite. This can be a sensitive situation, depending on various factors (like: how many other errors I’ve told them about that week, whether they got enough sleep last night, or if there’s been a jackhammer outside their window all morning). Rosie has no such reaction, or any reaction, ever. Because she’s a robot, not a person.
She isn't always accurate or relatable. Yet.
Rosie may be great at creating copy in the blink of an eye. But she’s not always nailing the “ABC Rule.” (Remember? Accuracy, brand voice, clarity.)
Accuracy? Well, sometimes she writes copy about a shoe when the product is actually a tee shirt. And if you ask her for a rewrite, she can’t respond to specific feedback, only try again (and it still might not be any good). Enough said.
And then there’s relatability (which is expressed through brand voice). Rosie’s lack of emotion can be viewed as a plus when it comes to process, but it’s a detriment to her copy quality.
Because AI can’t feel. It can’t experience. It can’t relate. It can't read a room and drop a well-timed joke. Rosie the product copywriter has never skied, run, swum, jumped awkwardly and muttered expletives while she pulled off a too-tight and too-sweaty sports bra, or suffered agonizing blisters after a hike in ill-fitting boots. Rosie has for sure never won or lost a marathon after years of training and dreaming, or dressed her tired and fussy kid in the morning, or craved the feeling of the sun on her skin during her bike ride home after a long day of working indoors.
Without these experiences, Rosie’s writing isn’t clever in a relatable way. Our AI can’t yet determine just what a consumer will latch onto about a product. I remember a perfect line one of our writers came up with to describe a tee: “Its fabric feels like the cool side of the pillow.” I have yet to see Rosie come up with anything close to that.
Rosie doesn’t know what the significance of an ultra-cool faded maroon color on a limited-edition sneaker is, or that its color and history are why it will sell out so quickly. She can’t know those things unless humans enter that exact info into a certain data field months before the AI “sees” it. And they probably won’t be able to describe it in a way she “gets.”
Takeaway: Rosie’s not nailing her assignments yet
When it comes to very simple products like apparel basics, Rosie’s copy often reads just fine. But we’re still working on the fact that normal-sounding copy often misses on accuracy. Even when we’ve supplied her with information about the product, she can still get product features incorrect.
In researching the state of AI writing development, I learned that AI has been noted as a divergent thinker, meaning roughly that it’s out-of-the-box and highly creative but often illogical. This is why it’s great at writing poetry and nearly won a literary award for novel-writing.
Unfortunately, creativity isn’t the leading skillset needed in large-scale production copywriting. Achieving a high level of accuracy is way more important. And Rosie’s struggle with accuracy is what’s currently standing in the way of her ecom writing career.
But she’ll improve with more training
One thing I learned: An AI writing tool requires a considerable investment of both time and energy. After it’s been coded by computer engineers, our team has to further refine and “train” it to write our way. And our ability to train the AI is limited by our ability to document our thinking about all the little decisions we often make without thinking.
It’s like an assignment my child’s second grade teacher gave her students: “Write down directions to make a pb+j, and I will follow them exactly.” Of course, the kids couldn’t write down exactly what they did at every stage, and the teacher was not able to make a sandwich from their incomplete instructions. Rosie’s like that teacher. She needs a lot of hand-holding to get to those 20-second product descriptions.
One way to train her is to keep “feeding” her approved product copy from our client, so she can learn what we’re looking for. But she needs to process lots and lots of copy to properly learn, which makes for a slow process.
Where will AI copywriting go from here?
Rosie has been part of the writing team for eight months now, and her writing still needs close editing and rewriting. She clearly still has some learning to do. But her progress is encouraging. She’s already a whiz at speed. And we’ve seen her accuracy and overall quality improve dramatically over the past months as we’ve given her feedback. This responsiveness and upward trajectory make us hopeful for her continued progress, and the future possibilities for product copywriting. And sometimes she gives us some real gems.* Who knows what she’ll be able to write five years from now?
So Rosie’s still got her job with Thread. We’re excited to see how she develops as a writer.
*“Every day is your playground. A playground that's packed with friends. No matter how far you travel, their laughter and storeys [sic] will follow you around wherever you go.” — Rosie, spring 2021
