How I grew a site to 10k+ daily visits just using GPT-3 made SEO pages
Over the last year, I've been working on building out lots of different categories of pages for a site, to help them rank on SEO. I wanted to share a bit about how we did it, and the results, especially as I've seen a lot of negative talk about the quality of AI-written content lately (+ worries Google will penalise it).
The way it works is we'll define a type/structure/category of page, whatever you want to call it, and then find a big list of input data that we pass into GPT-3 to build out the pages' content.
For example, lets say you run a cooking blog and you want to build out pages on every type of cake in existence. You'll first go and compile a list of all the cakes you can find, and save those in a CSV. We'd then go and write some python scripts that iterate through each cake (row) in the file, hit the GPT-3 API, and generate a bunch of HTML about that cake. We do this in a few steps - first generating the headline, structure, and then filling in the gaps and putting it all together.
We save the output HTML of each row and write that as another file, which we then upload in the CMS (wordpress, webflow etc). The pages tend to rank pretty quickly if your domain is decent, and especially if you're using this method to produce more obscure/long-tail content.
Cakes maybe aren't the best example in this sense, as there's a decent amount of content already about each one. If you think about more niche types of keywords though (e.g. "how do I do {x} with {y}") with lots of long-tail permutations, then you can easily produce content on topics that wouldn't be economical with a human (meaning less competition).
Anyway, these are the results we've see with about 6 of these projects over the course of this year. Note not only the consistent growth, but also lack of any slowdown after Google's helpful content update (designed to penalise AI content)
I realise that there's some technical skill involved in putting together all the above, as well as some prompt engineering, so I've been working on a product that basically does all this for you. You just upload a big list of keywords, and shortly after get back a CSV with the HTML for each one. I'm launching it in a week or two if you're keen to check it out/get early access: https://byword.ai/
You're open enough to write url of the product you want to sell us but not open enough to site which you created using AI content. So we can verify your claim.
Sounds like a fake salesman story.
True but most niche site creators don't reveal their sites.
This is pretty common - it's rare to find people who are building AI content listing URLs as Google are liable to manually penalise them.
I would stay away from GPT-3 generated content as Google now considers it is content spam and penalizes it.
If you are wondering, it's pretty easy to detect: https://huggingface.co/openai-detector/
noticed it looks for misspelled words to determine if it is fake or real . At 55 tokens if I used "gettign it" it was 66% real but if the words are spelled right it is only 12% real
it is good start for the intention of detector but it has ways to go
If this has been done for demonstration purposes by 3 people in 2019, I guess Google has a way more accurate way to detect GPT-3 generated content today.
This comment was deleted 8 months ago.
Definitely a controversial tool, bouncing between incredibly useful and full of unethical potential. Nevertheless, a very interesting product! Congratulations for getting it started.
I'm already hooked and will give it a try. This is really useful for startups that want to start working on their SEO but don't have the budget to write a bunch of content.
Happenstance, a friend and I were discussing the same but our background of setting up SEO pages based on content creation has been a steep climb so far. Your pricing construct is extremely value oriented , so far your execution is just awesome.
Love the idea, keep doing it .
Thanks! Interesting to hear you were looking at the idea too.
If you don't mind, I'd love to know a little bit more what you mean by the pricing structure? I've noticed that every other tool out there works on a subscription basis, which had me a little nervous charging per article, so would love to understand what you made of it if you don't mind.
Have you looked at magic-copy.com? I'm trying this out to see if the articles will rank but seems like they're high quality and look like human written articles so I don't see why not. (I like Magic copy because it's no-code and adds a royalty free image automatically)
I would love to use this tool to create content for our site.
First and foremost, congratulations on developing this test. That is an outstanding accomplishment. It would be fantastic to get the structure as well. Waiting for it to use for my kayak fish finder page here: https://bestkayakfishfinder.com/
Thanks! Just to check I understand - you mean a list of the article subheadings (e.g. a comma separated string)? I should be able to build this in, if you are genuinely keen on using the tool
Hey, I have 2 websites where I would love to use your product. I can be one of your beta-paying users. I have ideas on keywords to be used as input and then your tool should be able to generate content.
Can you please ping me on Milan84(@) g****.com
Thanks! Would love to have you on the early access list if you wanted to head to https://byword.ai and sign up
That's a significant validation of what GPT-3 can do.
How do you monetize the website? Are visitors sticking around enough to make a newsletter, affiliate promotions, or other revenue streams work?
The brand from the mini case study above is actually B2B SaaS, in a fairly technical area. They've taken the view that the traffic potential is so huge, that even with low conversion rates it'll help them gain awareness/dominate their space.
Otherwise, I'd be fairly confident you could monetise decently by:
So it's not complete autopilot, but helps you explore so many more areas/keywords than you could by hand
First off, congrats. This is mighty achievement.
A few tactical questions
Did you build backlinks to this site?
Did you publish all the pages at once? If so, how did google crawl the site (assuming there are a lot of pages)
Thanks! For your questions:
Awesome. Rooting for your win. What's the plan to monetize?
Affiliate? Ads?
Thanks! The site in the original post was a client's B2B SaaS, and goal there was just to dominate the space/build awareness.
(Goal for me is to productise what I did with them!)
Yeah I've made something similar as well.
But my approach is less on bulk articles (since GPT-3 does generate some fairly garbage data sometimes or things that are factually incorrect), and focus a lot more features on allowing user to pick the topic associated with keywords they want to write, and then letting them edit the article with AI-assisted editing if they want to inject their own voice / make the article more unique.
My thinking is AI isn't mature enough to 100% automate it just yet. But I think AI is still a great tool for folks to maybe save 90-95% of the time as far as drafting, coming up with an idea, editing content.
Other words: Instead of hours writing from scratch (or the other side of the pendulum which is 0% writing), I think a good tradeoff is just a hybrid workload where the AI writes most of it and the human might take 5-10 mins of reading/editing. And then just export the HTML/markdown.
I like the approach and congratulate you on your success.
However, what I am missing is what is coming after the ranking. I stopped caring too much about ranking a site, since my main focus is converting the traffic that I get. And my (limited) experience with AI-generated content so far is that you can rank easily, but it does not necessarily help you with converting all that traffic into paying customers.
How's your experience with that? Do you just add a CTA at the bottom of each page, or in the text, or not all because you do ...someting else?
I'd love to hear more about how that AI-supported ranking translates into conversions.
Hey, you're 100% right that this can be a weakness of this sort of content. My main thoughts are:
I love your last point.
That confirms what I was suspecting.
Still, this is a great approach for affiliates, but for my B2B SaaS? Not so sure.
I have a site with 1k+ pages and what really made the difference was not the content, but the backlinks to individual pages.
Yep, that's very fair. For what it's worth, the brand in the original post is B2B financial SaaS. They're taking a long-term view on trying to dominate/build awareness in their space, but this definitely doesn't lead to quick $.
Congratulation for making this test and sharing your success story!
One question about GTP-3: is it possible to generate a structure with it? I thought you can only generate the next paragraph.
What is the difference when you generate the headline and generate the content itself?
Hey, great question! It's pretty tough to get the structure in too, and this is where prompt engineering comes in. All the output I build has standard SEO structures (H1/2/3s, ordered and unordered lists) and is formatted in HTML. Getting to the point where you can output this is a bit tricky, which is really where the advantage comes in.
Hi mackgrenfell.
I love reading your story & the idea. I am torn between thinking it is fantastic or evil. Fantastic as it shows the value of AI; evil as it makes the efforts of writing stories pointless...
I will probably give it a try. From my past experience it can take a few go to tune the only keyword. Would you have any tips on this topic? Also what is the policy if one is not satisfied with the article generated by the keywords?
Hey, I agree there's definitely a conflict! The areas where I see this being most useful are really in the long-tail space, where it's particularly tedious/uneconomical for humans to write content. I don't see this stuff replacing good quality writers/storytellers any time soon.
What do you mean by tune the only keyword?
And re: satisfaction, if there are any errors or unusable content I'd just refund straightaway. If the content isn't quite as expected, but still good quality - it's a bit more complex. I'm planning to give people a bunch of free articles that they generate around any keywords in the app, and then once they're happy with format/quality they can go ahead and start building bulk projects (upload a list of keywords) as a paid feature.
I was meaning to write the 'right' keywords - but the phone keyboard had a mind of his own...
I was trying to refer to the problem you seem to have already thought about - it takes a bit of practice to find a set of keywords that generates a satisfactory outcome. So being able to practice a bit for free is/would be very helpful indeed...
Ah yep, you're right, there may be a bit of playing around to find the right keyword but it shouldn't be too hard. The general gist of it is to be a bit more explanatory in your input. E.g. 'how to bake strawberry pie' rather than just 'strawberry pie'.
Tried it out, have signed up for early access too. Is the main way to aim to differentiate your offering from copy.ai that you are looking to help people scale up to a huge number of articles?
Thanks - I guess it's two points:
Not necessary for a user to want to build a huge number of articles, but does support that.
Have you used copy.ai? It does sound a little similar. There's no harm in that, I run a business in an established space myself, but they might have some UI models and stuff you can borrow.
I think they've struggled to make users stick around because the quality is unreliable, but it was built on an older version of GPT so maybe they've improved it now!
I like the sample article you sent me :) It did repeat itself multiple times throughout, but I imagine that's hard to avoid with some keywords though.
This comment was deleted 2 years ago.