Lately, there is much talk about using AI-generated content for SEO. What are the best tools? Is the content any good? What does Google think about AI content?
The interest is understandable. Creating content is the most labor-intensive part of SEO. If we could automate this process it would be a total game changer — we would see a dramatic reduction in time and money required to populate a website with content.
But are we there yet?
In this article, we’ll take a look at the current state of AI content and review some of the most popular AI writing tools.
The first thing to mention is that AI is not a new thing in terms of content creation. Are you using Grammarly? How about the Hemingway app? Google Translate? Have you ever checked your copy for plagiarism? All of these writing assistants use artificial intelligence, one way or another.
The jump from assisting to writing, however, is relatively new. The technology that’s made it possible is called GPT-3 — Generative Pre-trained Transformer 3. It’s a language-prediction model that uses deep learning to create human-like texts. It’s also the first and so far the only artificial intelligence to pass the Turing test.
The way the technology works is you provide some input, like a topic or a paragraph of copy, and the AI continues writing for you. It does not generate any original thoughts, but instead draws on other texts already written on this topic. Although it has to be said that the AI does not plagiarize other texts — it writes its own copy.
GPT-3 is used in all modern AI writing tools, to various degrees of success. It all comes down to the modifications and the user experience built around the technology. Different tools choose to use different inputs for the copy. Some of the tools edit the copy to make it more readable, while others insert keywords to optimize the copy for search.
AI writing is not yet sufficiently advanced to substitute human writers — most AI content is fairly nonsensical. But, it’s not without its applications in SEO. Here are the three most common cases for using AI content.
The biggest problem for many writers is to actually start writing. This is something that can be easily outsourced to an AI writing tool.
Inbar Yagur, a marketer from Anyword, says in her blog post that AI content tools help her deal with the fear of a blank page when she starts writing. She actually used AI to form an outline to the very article where she mentioned this.
She did not end up using the exact outline generated by the tool, but she took the ideas and paraphrased them to fit her article. In this case, improving upon AI copy could be easier than starting a copy from scratch.
In addition, Inbar notices that:
AI is not very good at writing more than a few paragraphs.
It can be hit-or-miss with short-form content too. For every good variation, you get one to two duds.
Manual polish is needed for even a good copy.
Let’s skip the part where we debate the usefulness of PBNs. The fact is, regardless of whether PBNs work or not, they are still widely used in SEO.
The problem with PBNs is that they are these massive networks of websites and pages and they have to be populated with minimally viable content that no one will ever read. Sounds like the perfect job for an AI writer.
Here’s the typical experience of someone using AI content for link building:
I've been using [AI content] for two autoblogs that I've set up [May 2021] and I'm letting the software autopost to those blogs each day until they either get deindexed (they are indexed at the moment), or start ranking. Most of the time it makes no sense, but I don't care. As filler content, this software-generated content is good. If you input broad keywords into the software... like "weight loss", or "dog training", or "virtual reality" or whatever you will get decent content that you can just slam on your money site as-is and without a problem. But when you try long tail keywords, or very specific keywords the software falls short. I don't use content generated by software on the sites that I care about, but I do use it on autoblogs and web 2.0s (for acquiring backlinks).
Although I doubt the content can be used as-is on your money site — personally, I haven’t been able to produce good enough copy with AI writing tools. But, for the purposes of spammy link-building, AI content could be just the thing.
If you yourself are knowledgeable about the topic, you can use AI writers to create generic copy, and then you can fill out missing details and edit the copy for readability. It ends up being a joint effort, where the AI is the writer and you are the editor.
Here is a user testimony that’s representative of this approach:
If it's edited and satisfies the search intent, Google will shove it to the top. No indexing issues with AI-generated content so far. [I use] Jarvis. It's a writing assistant. You must do like 30% of the workload.
You are not going to create outstanding content this way, because you are skipping the research part and relying on the tool and what you already know, but it’s a good compromise for when you need to create the copy quickly. It could also be a useful solution for copywriters creating copy for others.
Google does not have a problem with AI content as such. What it does have a problem with is content that has no value to the user. Here’s Google’s current policy on automatically generated content:
From this policy we can see that the origin of the content is not a problem in itself. The problem is that automatically generated content is rarely of sufficient quality and is often used to manipulate the search algorithm. But, if you manage to create coherent, informative copy, there is no reason for Google to penalize it.
We have taken eight popular AI writing assistants that can generate long-form content and tried to produce some copy. Each tool was given the same (or almost the same, depending on the workflow) input to generate a blog post — we’ve asked them how to avoid stress at work. Then, we checked these articles for grammar mistakes, difficulty, keyword optimization, and plagiarism rate, and read the copy to see if it actually makes sense to a human.
We tried to interfere with the text generation process as little as possible, as our goal is not to create a super-cool article with the help of AI but to test what AI can do without any human assistance.
Here’s what we’ve got.
Jarvis is one of the most popular AI writing assistants. The description on the website says that all you have to do is open the long-form editor, ask the tool to write something, and enjoy the result within a couple of seconds. Let’s see if it’s really that easy.
Once you log in to the account, go to Templates > Long-form assistant.
Then, tell Jarvis what your article will be about by entering a brief description (not the title) and up to three keywords:
Then, you’ll need to come up with the title. You can write it on your own or you can ask Jarvis to come up with one for you.
The same with the intro paragraph:
Choose the length of the output text generated per session, put the cursor at the end of the last sentence of the intro, and click the Compose button.
Once Jarvis writes one part, click Compose again until you get the article of the desired length.
If you want the tool to dwell more on some idea in the middle of the text, just put the cursor at the end of the sentence you need to continue. If you want the tool not to pay attention to the passages above (i.e. we write a listicle, and each part describes some independent idea), separate these paragraphs with three asterisks.
To make the tool write a conclusion, type in the command Jarvis write a conclusion to the article about [our topic] and click Compose. Otherwise, the text may be endless.
I’ve produced an article of 1600 words. The more you write with Jarvis the farther away it goes from the initial idea. So, I felt like it was a good point to stop.
Jarvis used a solid outline to create a listicle of ten points on how to reduce stress at work. However, the copy within each section got increasingly nonsensical towards the end of the article. To the point where some sentences just repeated the word stress ten times over. All in all, there were a few good paragraphs and a solid outline — this would actually be helpful to jumpstart an article, but I would still have to do the majority of work myself.
I’ve also run the copy through the Hemingway app to check for readability and through the WebSite Auditor to check for search optimization. The Hemingway app came up with a score of 13 — the worst readability of all tested tools. The WebSite Auditor came up with an optimization score of 78% — the second lowest of all tested tools, mainly due to keyword stuffing:
Availability: Jarvis is priced at $99/mo and there is no free trial.
Frase is described as an AI-powered SEO content creation workflow. The peculiarity of Frase is that it uses its own AI called Frase NLG (Natural Language Generation), not GPT-3. The website descriptions say that you’ll forget about low-ranking content and optimization routine, as the texts generated with Frase are SEO-friendly by default. Let’s see if it’s true.
To activate the AI writing feature in the new document, switch to My content folder and click Write for me at the bottom of the workspace.
Then, enter the title and add some context as a description. Choose the length of the output text, set the creativity parameter, and click the Write for me button.
Just as we did with Jarvis, click Write for me until you get the desired amount of words.
The interesting thing about Frase is that the tool analyzes the top-ranked pages on your topic and shows you what an average good article on this topic looks like, from the number of words to the number of backlinks. The tool collects related searches and questions from different sources, analyzes keywords and their relevance, and uses them in the article it writes.
Frase also analyzed these pages’ structure and content — headers, sections, stats, and links.
I’ve managed to produce an article of almost 1200 words before the tool ran dry. Same as Jarvis, Frase went with a listicle and actually came up with a better outline than Jarvis did. The paragraphs also remained coherent all the way through. It did stumble a few times, like it messed up the numbers within the listicle and some paragraph turned into lists. But, overall, I was very impressed with the quality of the copy — with minimal effort it could be edited into a decent article.
Again, I’ve checked the copy for readability and search optimization. The Hemingway app assigned a readability score of 8, which is slightly above the optimal score of 6, but a decent score among tested tools. WebSite Auditor’s score for search optimization is 92% — the best score of all tested tools, so there might be something to Frase’s focus on SEO:
Availability: Frase is priced at $39/mo and there is a free sandbox with very limited functionality.
Rytr is another pretty popular AI writing assistant aimed to help us create anything from YouTube video descriptions to social media bios. The tool has a blog writing feature, which we have tested in our experiment.
To start writing with Rytr, just click the Start Ryting button on the homepage.
In the workspace, select the language, the tone of your copy, and choose the Blog Writing Section option from the Choose use case drop-down menu. Then, enter the topic, and click Ryte for me.
The peculiarity of Rytr’s Blog Writing Section is that it can only write one section but not the complete article. So you’d need to come up with sections on your own and change the input any time you’re starting a new section of your post. Otherwise, the tool will just generate endless variants of the same section.
In our experiment, we gave Rytr the section titles generated by Jarvis.
Rytr produced a very coherent copy all the way through and it would actually require almost no editing if we wanted to publish it. The only problem is that each section of the article is written separately and is disconnected from the overall theme. For example, there is a section called Use positive affirmations, and the copy is highly relevant to the section name, but it says nothing about how it would help to avoid stress at work, which is our main topic. Although I’ve seen worse copy written by actual humans, so I guess it’s not a deal breaker.
As for readability and search optimizations, the scores are solid. Hemingway gives this copy a perfect readability score of 6, while the WebSite Auditor says the copy is 88% optimized — among the highest scores in this test.
Availability: Rytr is priced at $29/mo and a free allowance of 5000 characters per month.
AI Writer is an AI writing tool that promises to “generate articles from just a headline”, which will be 100% unique and SEO-friendly.
The only thing you need to have your article written is to give the tool an input maximum of 125 characters long, click Write Article, and wait until the tool performs all the research and writing job for you. The process may take up to 3 minutes.
Together with the newly-created article, AI Writer will show you the list of citation sources it used to take the ideas for your text, and the uniqueness rate.
You can also explore the keywords used in the article by switching to the Main Keywords folder.
AI Writer doesn’t seem to write as much as it simply copies a bunch of quotes from a few sources. The copy looked suspicious, so we tested this theory with a plagiarism checker and got a very low uniqueness score of 52%.
There is also no momentum in the article — it doesn’t seem to go anywhere and reads like a few dozen introductory paragraphs stitched together.
As to the readability and search optimization, the Hemingway app gives this copy a score of 12 — the copy is very hard to read, while the WebSite Auditor gives it an optimization score of 77% — a bit much on the keyword stuffing:
Availability: AI Writer is priced at $19/mo and there is one week of free trial.
Copysmith is one more AI-powered and pre-trained writing assistant that can create and rewrite copy for various purposes. The tool creators have recently introduced a beta version of the blog post writing feature, so let’s test it.
Login to your account, click the Create button in the right upper corner, name your file (it’s not a blog post title), select Blog Post from the menu, and click Generate Blog Post.
Now the tool will ask you to add a title, an intro passage, and three titles for blog sections. As for the intro, you can create one with Copysmith, too. Create a new file, and choose the Blog Intro feature from the menu. Then, enter your title, select the tone, and enjoy several versions of AI-generated intros.
Copy the chosen intro, paste it into the blog writing workspace, click Generate, and wait a few minutes to see the result.
I’ve managed to produce an article of approximately 900 words. Upon reading, the copy seems unintelligible. Individual sentences do make sense, and sometimes even form a string of coherent sentences. But the copy overall goes nowhere, there is no clear structure and no progression. Overall, I couldn’t make sense of the copy and wouldn’t be able to turn it into something I could publish.
Now, let’s check optimization scores. The Hemingway app shows a score of 11 — very poor readability. The WebSite Auditor shows a score of 92% — actually impressive, if the copy made any sense to a human:
As this copy also seemed a little suspicious, I’ve also checked it for plagiarism and the uniqueness score was 72% — far from ideal.
Availability: Copysmith is priced at $16/mo and there are three days of free trial.
Sassbook is an AI article writing and summarizing software that offers content creation routine automation for any marketing needs.
Sassbook needs a short prompt that will be used as the beginning of the article (up to 30 words) and a title (optional) to start writing an article. You can also set up the level of creativity, the number of candidate generations, and the number of words per generation.
Once you fill out the fields, click the Complete button to start writing.
Click Continue to generate more text until you get the article of the desired volume.
After a bunch of generations, we have an article of 1026 words. We expected to see some tips on how to deal with stress, but the very first paragraph tells us about job sites. Then the text awkwardly dwells on remote work and downsides of offices and ends up with a description of some company hiring remote employees. The text itself is poorly written — many sentences are just nonsense, and the amount of pronouns is redundant.
Hemingway's difficulty score is 6 (well done!), while the WebSite Auditor’s keyword optimization score is 87%, still, some keyword stuffing has been detected.
Availability: Sassbook is priced at $32/mo and there is a free version with very limited functionality.
InferKit offers a web interface and an API for AI text generators and a text generation tool. Let’s see if it succeeds with blog post writing.
The tool will ask you to write a couple of words (or sentences, up to you) so it could continue the idea. You can also choose the number of characters per generation, the level of creativity (referred to as Sampling temperature), and the nucleus sampling score (they describe it as A probability threshold for discarding unlikely text in the sampling process). If you want each generation to be a complete extract, put a tick near Pause at end. After you’re done with settings, click the Generate Text button.
Click the button to create more text until the volume of the article is as you need.
InferKit gave us a complete article of 879 words and even came up with a text for the comments section and an author’s bio (we didn’t ask for this).
The article is a listicle with five points and the copy is coherent throughout. Although the outline is much more generic than some of the other tools were able to produce — I don’t feel like the topic is covered in full — we’ve barely scratched the surface. This would probably be a good tool to create throwaway content for link building.
Note: We have also tested various creativity levels of InferKit, and we don’t recommend changing the default settings. When the creativity level is low, the tool starts writing way too simple sentences. And if high, the text turns into some claptrap.
What about optimization? Hemingway's difficulty score is 6 (good job!), plagiarism — not detected. WA keyword optimization score is 89,5, with no keyword stuffing or keyword deficit.
Availability: InferKit is priced at $20/mo and there is a free version allowing 7500 characters per week.
ClosersCopy is a sales-oriented AI-powered copywriting platform trained on sales conversations and successful marketing campaigns. The tool has a Longform writing feature that is aimed to create sales pages, blog posts, and articles. Let’s have a closer look at what it can do.
To start an article, you will need a title and a short outline (optional) to help the tool navigate through the topic. Click the comet icon to activate the Longform tool, enter your input, select the creativity level, the output length, and the context, and click Write.
If you want to continue the section, copy the last newly-generated paragraph into the Longform input section and click Write. To start a new section, add the title of this section into the input field after the intro text, like this:
To generate a conclusion, replace the section title with the word Conclusion in the Longform input field.
The text itself sounds nice and smooth, still, the tool seems to have failed reading the intent — intro is the only part that dwells on stress at work. All the rest has nothing about work and describes the definition of stress and general stress management tips without mentioning the working environment.
Now let’s check the article by ClosersCopy with analytical tools. Hemingway’s difficulty score is 8. WebSite Auditor keyword optimization score is 94%.
Availability: ClosersCopy is priced at $21/mo and there is no free trial.
From what I’ve been able to produce with the tools above I don’t see them replacing human writers any time soon. Although some did pretty well (Rytr, Frase), all copy requires serious editing and perhaps further research to make it more unique. Those are writing aids at best, but no writers.
Users’ experience with GPT-3 tools and our experiment tell us the following:
AI-generated content can boost writers’ productivity by dealing with writer’s block;
GPT-3 tools can write good small pieces of text;
Google may index and rank AI-generated articles if they are of high quality;
All the AI-generated texts do need human editing before publishing;
Most tools fail when it comes to understanding the intention of the text;
Most AI tools need much guidance to produce something valuable;
AI tools tend to use general sentences and avoid details, so you’ll have to research on your own to add any stats or curious facts.
Have you tried GPT-3 for content creation? Share your experience in the comments.