Everywhere you turn, people are talking about AI, especially generative AI. ChatGPT from OpenAI became the fastest growing application ever earlier this year when it reached 100 million users. And, after using the application for a few minutes it’s no wonder why. Its ability to generate human readable text for different purposes and in different styles is uncanny.
A blog post about goat rearing in the style of Hemingway? No problem. An essay on the rise and fall of Ancient Rome written by Hunter S. Thompson? Easy. A haiku about iron ore? Coming right up.
For eCommerce businesses, the use cases are myriad. Instead of hiring a copy writer to describe your product, turn that task over to ChatGPT for a fraction of the cost. Or, if you’re resource constrained, stop writing product descriptions yourself and free up time for all those other tasks you have.
Staying up late answering inbound questions from customers? Have ChatGPT draft something for you to review and edit.
What about formatting product information? Again, send ChatGPT a table of product specs and with the right prompts it will convert that into a bulleted list or whatever format you specify.
If you own an eCommerce business and have a task that requires generating text, any kind of text, then you should be using ChatGPT to reduce your time to completion or to cut costs. While this is a great development for owners of eCommerce businesses, it merely scratches the surface of what ChatGPT and AI can do for your business more broadly.
Wait, what else can GPT do for my eCommerce business?
To some extent, the wow factor of generative AI does a disservice to AI’s total potential to improve outcomes for eCommerce businesses. The wow factor of seeing the exact piece of text you need appear in response to a short prompt is, undoubtedly, a revolutionary development for eCommerce business owners. But the new breed of large AI models like ChatGPT for text and its counterpart DALL-E for images will create new opportunities for eCommerce businesses beyond generating new content.
One of those opportunities is the ability to analyze and manipulate unstructured data. Typical examples of unstructured data are videos, images, audio and prose (e.g., product reviews). These data are distinct from structured data which is usually stored in table with rows and columns. An example of structured data might be a table that shows revenue by product line for each month of a year. The years are the columns and the rows are the product lines. This is a structured piece of data you can easily query: “what were my sales for product X in June?”.
But unstructured data isn’t like this. Let’s say you run an eCommerce business that sells rugs. You have lots of unstructured data for your product: images that show the product in different settings, videos that do the same, reviews from customers, descriptions of the product. Unless you have laboriously provided structure to this data yourself, you can’t query it.
An example will help illustrate the problem. Recall that you sell rugs. Rugs are frequently placed on top of hardwood floors. Let’s say you wanted to compare sales performance for products with photos taken on oak vs. cherry vs. walnut floors.
Unless you recorded this information when you took the photos and stored it somewhere, this could be a laborious question to answer. Somebody would have to examine each photo and note the wood used. Even if you make this task manageable by sampling, it is going to cost a lot of time, money or both.
Let’s twist this problem slightly. Imagine you read a product review for one of your rugs that says something like this:
“I was really excited for this rug to arrive, it looked great in the photos! Unfortunately I will be returning it. My floors are made of oak, but the product photos are all taken on cherry wood floors. The rug simply doesn’t fit with the colors of my room so I will be returning it. I’m sad too because the quality is terrific and it is very reasonably priced.”
Here we have a lost sale due to a mismatch between what the customer expected and what they experienced. And it’s particularly unfortunate because aside from the color mismatch, the customer loved the product! A strong review that could have driven additional sales suddenly becomes a cautionary tale for other purchasers.
But before you go and spend money to take new product photos for this particular rug, you want to understand how big a problem this is. Is it just one review, in which case maybe you need not worry? Or are there more reviews like this that are depressing sales for a great product?
Again, this is another time consuming problem that requires ready a bunch of reviews. That’s expensive in either time, money or both.
But with the arrival of the large AI models like GPT, both of these problems can be solved much more quickly and cheaply.
eCommerce, meet vectors; vectors, meet eCommerce
To understand how AI will help eCommerce business owners take control of their unstructured data, it helps to dig into aspects of how these models work.
The most important concept is that of vectors and vector embeddings. In short, large AI models need a way of representing unstructured data in a format that can be read by a machine. A computer has no concept of “this dress is blue”. But what it does understand are numbers.
If you took any algebra in high school or college, you may already be familiar with vectors. But the simplest way to think about them is that a vector is just a line. Below is an example from NASA. Don’t worry too much about the words and just focus on the fact that a vector is a line:
One of the critical components of large AI models is the conversion of unstructured data into an extremely complicated vector, which we refer to as a vector embedding.
How the the AI model does this isn’t really important to us. What really matters is that when two piece of similar data are fed into a large AI model, the resulting vector embeddings will also be similar.
To drive this point home, consider uses of the word “crane” in the example below. Among other things, a crane can either be a bird or a piece of heavy machinery used in construction. Without getting too technical, the picture below shows vector embeddings that have been forced into two dimensions (i.e., an x and y dimension, like the image from NASA above). These embeddings are the product of various pieces of text that use the word “crane”.
We can see immediately that references to crane as a bird are towards the top of the image while those that refer to crane as machinery are towards the bottom. In other words, these two different types of meaning are far apart.
And this is the essential point arising from the fact that unstructured data which has a similar meaning will produce similar vector embeddings: you can use those embeddings a substitute for manually inspecting and tagging unstructured data.
Let’s return to our rug example from earlier. This is the review we were considering:
“I was really excited for this rug to arrive, it looked great in the photos! Unfortunately I will be returning it. My floors are made of oak, but the product photos are all taken on cherry wood floors. The rug simply doesn’t fit with the colors of my room so I will be returning it. I’m sad too because the quality is terrific and it is very reasonably priced.”
Recall that what we wanted to know was how many other reviews were similar to this one. If we convert our reviews into vector embeddings, we can now get an answer to that question without having to inspect each review because, as we discussed, similar reviews will produce similar embeddings.
We have already prepared one case study using reviews of women’s clothing, but we’re also going to explore some examples here. Consider this line from a product review:
there are no premarked holes for the areas the need nails and the screw holes for the drawers don't line up
This is an extremely specific complaint, and it’s likely something a seller can do something about. Change the design of the product for example, or include better instructions. But to determine the depth of this problem is difficult given existing tools. A seller would need to read all of the reviews to find out how many are complaining about premarked holes.
But with NNext’s search capabilities, which are power by GPT, this is easy! The search below captures instances of people complaining about how holes are drilled in their furniture purchases. Now we can see quickly how significant the problem is and determine what resources to put behind fixing it.
Now let’s look at another example, but this time we’ll explore electric appliances and we’ll reverse the sentiment: we’re trying to understand what motivates people to buy our product!
Consider this snippet of a review:
Electric cost savings over a full size 220 is worth the time.
So what we want to understand is how much do energy savings motivate our customers? Is saving them money something they care about? How many reviews mention that?
Once again, this is easy with NNext. We simply drop the search snippet into our search tool and leverage the power of embeddings to return relevant results:
What does this have to do with ChatGPT?
Simple, ChatGPT (or more specifically, the model ChatGPT is based on) can be used to generate embeddings on your data! OpenAI, the company behind ChatGPT, will let anyone create embeddings from their data for a small fee.
That means that any business, even a solo operator eCommerce business, can now conduct analytical queries on their unstructured data.
At NNext, we are building the tools needed to enable these queries quickly and, importantly, cost effectively. If you are an eCommerce business that needs help analyzing your unstructured data, get early access to NNext at app.nnext.ai!