As a website owner, I’m always looking for new and better ways to write website content. After all, good content is what attracts people to my site. It's what helps websites rank in search engines. And it's what drives conversions. So when a new tool like GPT-3 was released, I was really excited to find out if it could generate quality useable content that could help me write copy for my websites faster.
In this post, I’ll highlight the GPT-3 does well, and also discuss some of the drawbacks so you can determine if using a GPT-3 content generator will work for you.
What is GPT-3?
GPT-3 is an AI system that can process and produce natural language. It has been trained on a large amount of data. This data includes everything from books to websites to articles. The purpose of GPT-3 is to generate text that is similar to the training data. For example, if you feed GPT-3 a website, it will generate text that looks like a website.
So, in theory, GPT-3 could generate almost any sort of text, from simple headlines to longer content like blog posts and articles.
But how good are the results?
The short answer is, "it depends."
How Can GPT-3 Be Used to Write Website Content?
In order to get a response from the GPT-3 system, we need to feed it some sort of prompt. The prompt is what the system uses to generate text. So, if you wanted a headline for your fishing website for example, you could basically just ask GPT-3 for a headline and it would do it’s best to complete your request in plain language.
This is what got me really excited about this technology. I could ask for any part of a website and it would, at best, generate something that I could use verbatim, or at worst, give me an idea to start with and take it from there.
GPT-3 can be used to generate pretty much any part of a website:
- Hero section headlines and intro copy
- Section headlines
- Feature lists
- Feature text
- Product descriptions
- Meta descriptions
- FAQs and answers
- About us sections
- Blog topic ideas and outlines
- Blog text
- Alt text for images
- Call to action
GPT-3 can do a surprisingly good job out of the box with simple prompts. But the better the prompt, the better the results.
Getting Better Results with Few Shot Learning
As impressive as the initial results were with simple prompts, I’m kind of picky about writing, so I wanted the results to be even better. Just a little more relevant, a little more clever and a little more accurate.
With GPT-3, additional information can be added to the prompt to improve the performance, sometimes dramatically. This additional information basically contains a few examples of what kind of result you’d like to see. This helps GPT-3 do a better job determining what kind of results you want to get.
This technique is call few shot learning. I’m saying, here’s a few examples (a few “shots”) of what I’d like to see. Let’s see if you can generate a result that has these characteristics.
When done well, the generated results will better match your intensions and will be quite a bit better than simple prompts.
The Down Side of Few Shot Learning
As far as getting better results go, there is only an upside for few shot learning. But there are a couple of downsides:
- it’s more expensive to run
- it takes time to figure out which examples are going to generate the best result
When using the GPT-3 engine, we pay according to the amount of text in the prompt plus the generated result. Longer prompts mean that it will cost more, so inherently few shot prompts will cost two to three times more than a simple prompt. But usually getting better results makes up for that cost in writing time saved, not to mention that you might have to run a simple prompt request multiple times to get a desired result anyway.
The time spent researching the best prompts to use is a one-time cost and ends up paying off in the long run.
WebWriter.ai Uses Few Shot Learning
When I was researching how to make the best use of GPT-3 for webwriter.ai, I made the decision early on to only use few shot prompts whenever possible. I wanted users to get the best possible results, even if that meant it would cost more per request.
I wanted webwriter.ai to be able to generate high quality useable text with as little input and as little frustration as possible. We have fine tuned our few shot prompts to be able to generate text for any section of your website with as little input from you as possible, saving you a lot of time and handing you ideas that you might never have thought of on your own.
Potential Drawbacks of Using GPT-3 to Generate Website Content
GPT-3 has the potential to revolutionize website content generation, but there are some potential drawbacks to consider.
One potential drawback is that the content generated by GPT-3 may not be completely accurate or reliable. GPT-3 is only as good as its training, and if there is conflicting information in the training data, it could result in factual errors.
Another drawback, although somewhat rare, is unintentional plagiarism. Because GPT-3 was partially trained on written works across the Internet, it is possible that some text may be taken verbatim from another source.
A third drawback, especially in the case of few shot prompts, is that you may sometimes get prompt “leakage.” This means that part of your prompt may end up in the result, even if it is completely unrelated to you initial request. If this happens, you can simply try again and re-run the request. More times than not, the system will not repeat that mistake.
In light of these possibilities, it is recommended that you review your generated content for any errors or instances of unintentional plagiarism before publishing.
Finally, GPT-3 content may not be as unique as content generated by a human writer. This could be an issue if you're looking for truly original content.
But despite these potential drawbacks, I have found that with a little oversight, GPT-3 is an invaluable tool to help me quickly write content for my websites.
So Really... How Good Are the Results?
With services like webwriter.ai that use the more expensive few shot prompt learning, the results are extremely impressive. I have saved and enormous amount of time writing headlines, feature lists, coming up with initial FAQs, and even more surprisingly, the FAQ answers. They are amazingly accurate in many cases, and I have been able to use a lot generated text verbatim without any editing.
For longer form content, like blogs and articles, GPT-3 results are very useable. If you are using it strictly as content generation for SEO purposes and don’t need to add your own ideas to the mix, I would say that you could use 90% of the text verbatim. You may need to edit transitions from one section to the next, but it is fairly minimal. You could easily generate a blog post in less than 20-30 minutes if this is your use case.
For other long for content that you want to integrate your own ideas and writing, GPT-3 can still save you a lot of time by generating an outline for your topic as will as text to get you started. What I have found is even if I don’t much of the text verbatim, it will spark my own ideas and make writing easier and faster. It actually jumpstarts my writing, so instead of writing exactly 0 blog posts, I’ll be able to crank one out in an hour or less.
The best way to find out how well GPT-3 works for you is to sign up for a webwriter.ai account. When you sign up, we’ll give you 3500 words for free. No credit card is required.