E-Commerce Copywriting.

Product copy from images, at scale… from a dream to Dreem.ai. Introducing multi-dimensional automated copywriting.

Copy has a lot of benefits, even if a picture says a thousand words, the nuance and specific details that come from well-written copy make a huge difference. And not just for the specific product, but also for how a brand shows itself to the world.

There as plenty of AI tools that can help you write just about anything, but we realized that to best help out customers, those wouldn’t do. Short copy for e-commerce is something different than a blog post about avocado toast, or writing emails. This is both in terms of input, output(s), language and tone and much more.

Having always worked extremely closely with our customers and done endless research and experiments, we know we can provide value like no other in this. We offer niche solutions, and we are, if we may say so ourselves, really good at understanding that niche.

Putting an interface in front of a large language model (LLM) is not the hard part, and it was not the part that caught our interest. Going into development for our Copywriting tool, we became intrigued by the possibilities to capture nuance and provide content at scale while keeping the right tone of voice, knowing the customers, and delivering value in a way that is tailored to the need.

Manually writing copy at scale is a huge undertaking. Sometimes even to the level that it ends up not getting done. This hurts everything from SEO to sales to brand trust. By being able to deliver copy in any format, multiple languages, and for multiple platforms, we can help with finally catching up with that backlog of products, and make sure the newest products always have a beautiful and nuanced piece of copy.

Our approach

To make it easy for our users to quickly get copy written, we take the product images and use those as the basis for the product descriptions. Using AI, we can analyze details such as usage, categories, what makes it stand out, and much more. We can also pair this data with PIM data or other metadata and descriptions, as well as information from tools like our Garment Label scanner, to complete the job.

We decided to look at the realities of how copy works before diving in too deep. This resulted in a host of research that showed that the job is not that simple. One thing is trying to mimic the words of a brand tone, but cadence, contrast in headlines and text, rhythm, and an understanding of who will read it, who will buy it, and who will wear it all come into place—more so than the actual words in many cases.

We created a templating language that allows our users to define exactly how our customers want the copy structured. No more trying to wrangle an LLM to give you what you want. We made it easy to define even complex structures to stay on brand and make sure the buyer recognizes the pattern and can find what they need. It’s a flexible and open-templating language, so there are really no limits to how you want your copy written.

Using pre-filled features and image analysis to write copy following a predefined structure template

Initially, we started testing on training models to match the tone of voice needed, but there turned out to be a lot of challenges in delivering the quality we wanted.

• When doing traditional training, we need a lot of data. And to be honest, when we need thousands of pieces of content, they are not all to the desired level of quality. That would make the results average at best. We are not here to deliver average… we are not competing with “more of the same”. So in order to elevate the delivered product, we came up with a different solution.
• Training can be slow… very slow. Both computationally, but also in the effort at our customer’s side to gather training data, validate, etc. So not ideal.
• It gets difficult to be flexible. If our customers would rather have a specific LLM generating the copy, or a new and even better LLM comes out, we would potentially need to retrain all the models. This means slow updates and a lack of flexibility.
  
  

Using reference copy and templates to define all the outputs you need while matching tone of voice, cadence, and language dynamically

Instead, we opted to create a system of AI models and agents that allow us to get by with a templated structure, as well as 1-5 examples of the best copy for the output. That gets us around the issues of average and elevates the whole brand experience. So now it’s possible for our customers to spend some time crafting the perfect examples, and we will make sure to align to those! Much better!

What’s more, we found that it is valuable to deliver more than one output in many cases. For example, we always deliver alt text for the images used to write the copy. This helps with SEO and for visually impaired users.

Automatic alt text generation for all images for SEO optimization

But it doesn’t stop there. We have the possibility to also create multiple language versions, short and long versions, or maybe you want copy for your newsletter? We have you covered. There is a reason why we call it multidimensional.

Generated copy aligning to structure and tone of voice

AI models are changing and improving at a rapid pace. And we do not want to be the company that locks our customers to a specific model. So we have worked hard to decouple our system from specific LLMs and allow our customers to choose what they want to use. This also has the benefit of us being able to VERY quickly (usually within a day or two) enable new models as they come out and show benefits to the problem we are solving.

In order to provide the best copywriting, we mix all the available data. There is of course the image analysis, which plays a big role. But we can also work with property data from Creative Force, free text entries, or data from your PIM system. This gives us as much detail as possible, making the quality even better and more precise. If a new data source should come up, we can easily integrate that as well.