Using ChatGPT in eCommerce
A term that didn't exist not too long ago now dominates traditional and social media as well as conversations in the kitchen and at the lunch table: ChatGPT. And rightly so: The results that can be achieved with ChatGPT in the most diverse areas amaze everyone. In this article, we take a first careful look at possible applications in eCommerce.
What is ChatGPT?
Behind ChatGPT is the company OpenAI, in which Microsoft recently participated with an investment of 10 billion USD. OpenAI has developed a so-called "Large Language Model" (LLM) with GPT, which has recently appeared in version 4. ChatGPT is an application with which you can interact with this language model like in a conversation. ChatGPT also remembers what has already been said. It therefore very much resembles a chat with a real person, hence the name.
ChatGPT amazes people worldwide
And this language model is amazingly powerful. People who have never programmed before can suddenly develop apps for the iPhone. GPT in version 4 is no longer limited to language, but also masters images. A photo of the refrigerator is enough and ChatGPT tells you what can be cooked from it. A photo of a website, and ChatGPT programs the rest. An error occurs? The tool also fixes this if you ask it to. ChatGPT can write entire articles or books and illustrate them directly with the help of the image AI Dall-E. On request, it provides legal, tax or medical advice. And at the latest here it becomes dangerous, because sometimes ChatGPT lies like crazy, but more on that later.
An expert for language
ChatGPT can handle language very well. More precisely: languages. Because whether English, French or German, ChatGPT doesn't care. The model was trained multilingually. You can ask in German and ask for a Chinese answer. All of this is no problem for the model. And one more thing distinguishes ChatGPT: You can communicate with the model via an interface (API). And at the latest now every developer's heart beats faster, because this allows interesting applications.
ChatGPT in eCommerce
ChatGPT is therefore also interesting for eCommerce, because product texts, reviews and customer questions about products play a major role here. We would therefore like to briefly address some possible applications using the Amazon marketplace as an example.
Keyword Research
Amazon is a search engine and those who are not found cannot sell either. It is therefore crucial to include relevant search terms in the product listing, and ChatGPT can already support here. If you ask the tool quite generally which keywords should be considered for a "gaming monitor", it quickly delivers good results:
The tool also provides typical misspellings on request:
Designing Product Texts
Equipped with this information, it can then also start directly to design finished texts, as the following example shows, in which it is about formulating the so-called "bullet points" on Amazon.
The answer doesn't take long either:
Anyone who looks closely here quickly notices that ChatGPT took the keywords a bit too literally. These were inserted 1:1 and no attention was paid to upper and lower case or spelling errors.
But that's not a problem either: Ask ChatGPT to correct the bullet points accordingly and the result will be better:
Anyone who knows Amazon knows that these bullet points are still a bit too long. Therefore, we ask ChatGPT to shorten the text.
In the next step, we could also have this description translated into other languages with the help of ChatGPT. The implementation is trivial, so we refrain from a concrete example here.
Whatever language you choose: The result can already be seen and we haven't done any special prompt engineering. Prompt what?
Excursus: Prompt Engineering
The questions or requests that users make to ChatGPT are also called "prompts". And especially with ChatGPT, it is crucial that you formulate the task exactly to get the desired result.
Just imagine you had to brief an agency or freelancer in writing and only have one attempt. There are no follow-up questions. What is not included in your briefing, the agency or freelancer cannot take into account either.
A good prompt therefore includes at least the following elements:
- Define role: Specify what "job" ChatGPT has. It makes a difference whether you specify that you need a lawyer, marketing consultant or journalist
- Task: Describe the task exactly.
- Context: Give ChatGPT the necessary information it needs to work on the task
- Examples: If in doubt, provide examples of what the answer should look like
- Format: Specify exactly what you need. This can be a table, an article, an outline or just text. Also specify the length.
The first draft of a prompt is usually not the best. Look at the result and refine the request. Also point out errors to ChatGPT. ChatGPT will try to correct these in the next attempt. After a few rounds, your prompt will look much better. And last but not least, you can even ask ChatGPT to create a suitable prompt for your task. You might think that the cat is biting its own tail here, but let yourself be surprised or inspired.
Analyzing Customer Feedback
Back to applications in eCommerce. Customers can leave product reviews, seller feedback and also questions on the listing on Amazon. ChatGPT is ideal for getting an overview of e.g. frequent points of criticism or typical questions. In the following example of our "gaming monitor", we gave ChatGPT a series of customer questions with the task of summarizing them.
Here too, ChatGPT does a complete job, as the answer shows:
This information can now be used to improve the listing accordingly, either manually or by asking ChatGPT to adjust the product description accordingly. Here you should just not forget to give ChatGPT the necessary information.
We mentioned at the beginning that ChatGPT can also be queried via an interface. The following example shows that only a few lines of code are needed (which, by the way, were also created with ChatGPT, of course), in which various positive and negative reviews of a carpet were analyzed (the reviews are not completely visible in the screenshot due to the missing line break).
This information can now also find its way into the product description.
Product Images
OpenAI has its own AI with DALL-E that can create images from texts. But there are already better models on the market. The "Midjourney" model is currently causing a stir, not least because it attributed a powerful down jacket to the Pope. Many thought the image was real because it also looks real.
How well Midjourney works in the current version 5 is shown by the following example:
This woman doesn't exist. The image was artificially created. The Midjourney prompt for this was provided by - how could it be otherwise - ChatGPT, after it was explained in a long prompt what makes a good Midjourney prompt:
Woman in yoga class, professional space ::8
Power yoga, energizing practice ::4
Focus, determination ::2
Sweating, athletic performance ::2
Supportive environment, group class ::1
While there have been tools for a while that can make backgrounds disappear on request, a series of tools are currently emerging that go exactly the opposite way: You have a cut-out image and want to place it in front of the Paris Eiffel Tower in the hand of an artificially generated model? No problem! A correspondingly trained AI is sufficient and you save the costs for the trip, the models, the photographer and more.
Problems with Using ChatGPT
As beautiful as all this sounds, there are also some problems in dealing with ChatGPT. The probably biggest problem in practice are the "hallucinations". The term refers to the fact that ChatGPT sometimes makes up facts.
This is because ChatGPT is a statistical language model in which word by word is determined based on probabilities which word should come next. And in doing so, untruths can also flow into the expected result.
If you ask ChatGPT in version 3.5 e.g. who wrote the book "Amazon Marketplace", two completely different authors are named instead of the true authors. In version 4, this error has already been (partially) fixed, but the simple example shows how quickly ChatGPT can be wrong.
Another challenge lies in dealing with internal company data. ChatGPT openly points out in its browser application that you should not use sensitive data with ChatGPT. Because you cannot know whether this data will not become part of the "general knowledge" of GPT in subsequent versions. Many companies are currently frantically writing corresponding instructions on whether and how ChatGPT may be used.
With images, there are currently discussions about who actually owns the rights to the artificially created images. Can these be used commercially? Often the images were trained based on copyrighted images. What influence does this have on newly created images?
Outlook
The speed at which ChatGPT is currently developing is enormous. So version 4 of GPT has recently appeared, which, as mentioned, now recognizes images in addition to texts and delivers significantly better answers than version 3 or 3.5.
In addition, version 4 can now also google and thus get current knowledge ad hoc. OpenAI has also launched a kind of "App Store" where developers can deposit so-called plugins to provide OpenAI with additional information if needed, such as current flight or hotel data. Thanks to corresponding plugins, ChatGPT has now also become a math genius, as it draws on Wolfram Alpha here.
ChatGPT can now program itself and also execute these programs directly. So-called "agents" can independently define multi-step activities and then execute them.
The enormous pace that ChatGPT is currently showing also calls critics to the scene, who rightly ask what social impact tools like ChatGPT will have? People who think one step further already see the beginnings of a real artificial intelligence that has access to all of humanity's knowledge and possibly develops a life of its own that can no longer be controlled by humans. Prominent people (including Elon Musk) even demand in an open letter a stop to current development.
I personally believe that Pandora's box is now open or - as ChatGPT would put it - "The confetti has left the cannon". ChatGPT offers many applications including in eCommerce. These should be used and the emerging risks controlled.
ChatGPT will become a tool that is no longer conceivable from everyday life. Based on ChatGPT, specialized applications for different use cases will emerge. It will be crucial that you know how to use ChatGPT for yourself. The prompt engineering described above as well as the skillful chaining of tasks are of the greatest importance.
You can find more information about using ChatGPT in connection with Amazon on our info page.
