A Straw Hat's AI Experiment: How Traditional Artisans Use Big Data to Sell to the World

A while back, I came across a rather interesting study about the straw hat industry in Zhejiang.

You might think — straw hats? Someone actually studies that?

Yes. And they studied it very seriously.

In Cixi, Zhejiang alone, there are over 50 straw hat manufacturers, generating an annual output value of 250 million yuan. The gold-thread straw hats from Changhe Town have reached nearly 70 countries and regions worldwide. This is no small-time operation.

But as these businesses grew, they ran into a problem: the domestic market had become saturated.

The production capacity was still there. The demand was gone. What now?

Sell overseas.

Selling Overseas Isn't That Simple

Cross-border e-commerce sounds wonderful in theory. But once you actually dive in, you find it's full of pitfalls.

The first pitfall: labor costs. The workflow for handling foreign trade orders is remarkably long — customer inquiries, design confirmation, prototyping, sample revision, contract signing, production, delivery — every step needs someone watching over it. As orders pile up, headcount has to grow right alongside. For a craft-based product like straw hats, there's almost no way around this labor overhead.

One study found that the back-and-forth of sample communication alone can eat up 30% to 60% of an entire order's cost.

You read that right. Half the cost, spent going back and forth confirming "what exactly does the hat you want look like?"

The second pitfall: not understanding the market. The veteran hat makers in these factories have been crafting hats their entire lives — their craftsmanship is flawless. But ask them "what colors are European and American consumers into this year?" or "which scenarios have the highest search volume for hats?" and they draw a blank. Small and micro enterprises have no capacity for market research. Production runs entirely on instinct and experience.

The result? Either you bet wrong and inventory piles up, or you miss the next hit product and watch helplessly as someone else cashes in.

Data Can Tell You What to Sell

This study did something very direct: it went to Alibaba 1688 and scraped all the straw hat–related data for the entire year of 2023.

A total of 3,607 product listings, each with 187 attribute fields. After cleaning, deduplication, and merging, 49 effective attributes remained.

Then the analysis began.

First finding: merchants offering customization earn more.

Merchants providing customization services — adding logos, manufacturing from designs, affixing rhinestones, and the like — accounted for 72%, with an average monthly sales figure of $1,198.68. And those who didn't offer customization? $676.30. Nearly double the difference. What's more, the group offering customization had more stable profits — less vulnerable to being dragged down by extreme outliers.

Second finding: hats that work across multiple scenarios sell better.

87.4% of straw hat products were suited to only a single scenario — work, leisure, or parties. But the data showed that hats appropriate for five or more lifestyle scenarios had notably higher average sales. The logic is simple: e-commerce search engines match by keywords. The more scenarios your hat can cover, the greater the chance it gets found.

The same held true for seasons. Hats suitable for multiple seasons were more popular. Straw hats tagged only as "winter" barely sold at all — in consumers' minds, straw hats mean coolness and style, not warmth.

What Do Overseas Consumers Actually Care About?

Domestic data only tells you about the supply side. To understand what overseas consumers are thinking, you need to look at reviews.

The research team scraped straw hat buyer reviews from Amazon, divided them into three price tiers — under $25, $25 to $50, and over $50 — and ran them through an LDA (Latent Dirichlet Allocation) topic model (an algorithm that automatically extracts topics from large volumes of text) to mine each tier separately.

The results were fascinating.

Buyers under $25 cared most about three things: whether the size fits their head, whether it's practical, and whether there's enough variety in style. This group buys hats for everyday use — sun protection on vacation, a small gift — cheap and functional does the job.

For buyers in the $25 to $50 range, a different picture emerged. They cared about whether the online photos matched the real thing, how good the customer service was, and how the after-sales experience held up. On color, solid tones were more popular because they're easier to match. Consumers in this price range aren't just buying a hat — they're buying an entire shopping experience.

Buyers over $50 treated straw hats as art. They focused on design uniqueness, the refinement of the craftsmanship, and even whether the materials were eco-friendly. For this group, a hat is a vehicle for expressing taste and lifestyle.

Think about it — these three groups care about completely different things. If you use the same product description to win over all of them, at least half your budget is wasted.

Finding the "Best Recipe" with Decision Trees

Up to this point, the analysis covered the impact of individual attributes. But consumers never buy a hat for one single reason — it's always a combination of factors stacking together.

The research team used a decision-tree model to run all these factors at once. They selected five variables — hat shape, style, suitable seasons, number of colors, and service rating — with sales volume as the target.

The model's conclusions:

Dome-shaped hats with no brim were the most popular. On style, personalized, vintage, and original designs sold far better than "no distinct style." The optimal number of colors was five or more. Seasonal positioning should target spring and summer.

And one finding that may be the most important of all: the merchant's service rating had a greater impact on sales than any product attribute.

In other words, no matter how well-made the hat is, if the shop's service rating doesn't hold up, it still won't sell. The model recommended pushing service ratings to around 0.9.

So You Know What to Sell — Then What?

At this point, big data has only told you the "surface problem" — what the market needs.

But what truly bottlenecks these businesses is the "root problem" — how do I produce it quickly?

Remember that number from earlier: the sample communication phase eats up 30% to 60% of the order cost. That's where profit leaks away.

The research team turned their attention to AI image generation.

Letting AI Draw Straw Hats

By 2023, generative AI image technology had become quite mature. Stable Diffusion, DALL-E, Midjourney — and in China, Baidu's ERNIE and Alibaba's Luban — could all generate images from text descriptions.

But using a general-purpose model directly comes with two problems.

First, copyright risk. These models are trained on publicly available internet images, which may include copyrighted material. Using them to generate commercial designs carries legal exposure.

Second, the generated images "don't look like your products." A general-purpose model can draw anything, but it doesn't understand your company's specific weaving techniques, material textures, or brand style. The customer takes one look and thinks, "this isn't what I wanted."

What to do?

The research team chose a Cixi company called Jusheng Straw Hat for their experiment. This company had accumulated a large volume of overseas orders and customer requirement records over the years — roughly 3,000 historical data entries, each corresponding to a final confirmed sample image.

They took Stable Diffusion as the base model and fine-tuned it using Jusheng's own historical data.

The specific technique is called LoRA (Low-Rank Adaptation). In plain terms: instead of training the entire model from scratch, you only adjust a small subset of its parameters. Training volume drops to one ten-thousandth of the original, VRAM usage is cut by two-thirds, and generation speed is unaffected.

How did it perform?

The model could generate straw hat design sketches directly from a customer's text requirements. A customer says "I want a beige flat-top straw hat with a black ribbon," and an image appears in seconds. Want to change the request? Adjust the color, tweak the brim width — also a matter of seconds. What would take a designer half a day to revise in the traditional process now has virtually zero wait time.

Even more critically, the generated images carried Jusheng's own product style — the weaving textures and material qualities all matched up. Because the training data came from this company's real orders.

A Replicable Framework

Putting it all together, the research team proposed a framework with three components:

First, market data. Scrape sales data and user reviews from e-commerce platforms, and use machine learning to uncover "what sells" and "what consumers care about."

Second, product development. Feed the key attributes mined from the data — hat shape, color, scenario, and so on — into the company's own AI model to directly generate design sketches and production plans.

Third, downstream adaptation. After new products go live, continue adjusting based on real-time platform feedback.

This framework isn't limited to straw hats. Swap in bamboo weaving, rattan chairs, or handmade ceramics, and the logic is identical: use data to find demand, use AI to compress the time cost at the front end of production.

Finally

What makes this study interesting to me isn't the cutting-edge technology it employs. LDA, decision trees, Stable Diffusion — none of these are new on their own.

What's interesting is that it stitched these technologies together to solve one very specific, very traditional problem — how to get a straw hat from a small workshop in Zhejiang to the other side of the planet, faster.

Traditional craftsmanship and AI sound like they belong to two different worlds. But in the straw hat factories of Cixi, they've already taken their places on the same production line.

Maybe before long, that handcrafted straw hat on your head — its very first sketch was drawn by AI.