[Journey] 1 million UVs/month in 12 months using AI generated content. Let's do it!

[doctorhm007]

It makes perfect sense. What really matters is their advertisers get performance, and since your traffic is from SERP's, regardless how you get it, the serp traffic is high, and gets conversions for their customers, so makes perfect sense that they would still payout.

Wow net 90? well they pay is nice just takes a long time to get it haha.

I'm stuck on H1 headlines, i have the content paraphraser pretty good. What apis or methods are you guys doing to generate a buzzfeedy title that keeps the correct keyword and makes it more "copy" and "clickbaity? I don't want same title as scrapped but just paraphrasing doesn't make it exciting. Any tips or ideas guys?

I keep all headlines as they are because they have weights on SEO so I don't like them to be paraphrased.

 

[Sartre]

It makes perfect sense. What really matters is their advertisers get performance, and since your traffic is from SERP's, regardless how you get it, the serp traffic is high, and gets conversions for their customers, so makes perfect sense that they would still payout.

Wow net 90? well they pay is nice just takes a long time to get it haha.

I'm stuck on H1 headlines, i have the content paraphraser pretty good. What apis or methods are you guys doing to generate a buzzfeedy title that keeps the correct keyword and makes it more "copy" and "clickbaity? I don't want same title as scrapped but just paraphrasing doesn't make it exciting. Any tips or ideas guys?

NET65.

I just add a 100 variants of [Expert reviewed!] etc in the end

 

[DamonSalvatore]

Hey @Sartre,

Is the content on Modern Bloke generated using your current paraphrasing model or previous ones?

Can you provide some pointers/suggestions in determining keyword difficulty programmatically?

 

[Sartre]

Hey @Sartre,

Is the content on Modern Bloke generated using your current paraphrasing model or previous ones?

Can you provide some pointers/suggestions in determining keyword difficulty programmatically?

it's my old method. We've actually come up with something amazing this week and I will be explaining it here over the next few days.

KW difficulty:

i created an algorithm that checks title/article relevancy to the keyword in top 10 SERPs, page/domain authority, and looks for user generated content on page 1. then weighs it all and calculates difficulty. I'd say it's about 50-100x better than Ahref's KD, efficiency-wise. They simply don't have the resources to calculate KD like this on such a big amount of keywords. I'm only calculating like 10k keywords at a time and it takes hours. And I've got pretty powerful hardware now.

 

[prvncqr]

it's my old method. We've actually come up with something amazing this week and I will be explaining it here over the next few days.

KW difficulty:

i created an algorithm that checks title/article relevancy to the keyword in top 10 SERPs, page/domain authority, and looks for user generated content on page 1. then weighs it all and calculates difficulty. I'd say it's about 50-100x better than Ahref's KD, efficiency-wise. They simply don't have the resources to calculate KD like this on such a big amount of keywords. I'm only calculating like 10k keywords at a time and it takes hours. And I've got pretty powerful hardware now.

did you purchase a new hardware or rent it?

 

[Uroperon]

i know someone get content from Q&A sites like stackoverflow, then translate to another language, and it works great, don't know how long it will last, but currently it works fine.

I have done it before, it works.

 

[DamonSalvatore]

it's my old method. We've actually come up with something amazing this week and I will be explaining it here over the next few days.

Looking forward to it. If it's possible, whenever you are explaining your new findings, can you also show us some content generated from your model along with original paragraphs.

KW difficulty:

i created an algorithm that checks title/article relevancy to the keyword in top 10 SERPs, page/domain authority, and looks for user generated content on page 1. then weighs it all and calculates difficulty. I'd say it's about 50-100x better than Ahref's KD, efficiency-wise. They simply don't have the resources to calculate KD like this on such a big amount of keywords. I'm only calculating like 10k keywords at a time and it takes hours. And I've got pretty powerful hardware now.

This helps a lot. This whole thread is very resourceful thank you for all the information you are sharing

 

[Sartre]

did you purchase a new hardware or rent it?

I've rented 2 threadrippers and a P100 in the last month.

 

[spectrejoe]

it's my old method. We've actually come up with something amazing this week and I will be explaining it here over the next few days.

KW difficulty:

i created an algorithm that checks title/article relevancy to the keyword in top 10 SERPs, page/domain authority, and looks for user generated content on page 1. then weighs it all and calculates difficulty. I'd say it's about 50-100x better than Ahref's KD, efficiency-wise. They simply don't have the resources to calculate KD like this on such a big amount of keywords. I'm only calculating like 10k keywords at a time and it takes hours. And I've got pretty powerful hardware now.

Nice! I assume you using Moz API for checking DA/PA?

I'm also going to look for footprints to find user generated content instead of only looking for reddit/quora/forum in link

 

[Archemike]

it's my old method. We've actually come up with something amazing this week and I will be explaining it here over the next few days.

KW difficulty:

i created an algorithm that checks title/article relevancy to the keyword in top 10 SERPs, page/domain authority, and looks for user generated content on page 1. then weighs it all and calculates difficulty. I'd say it's about 50-100x better than Ahref's KD, efficiency-wise. They simply don't have the resources to calculate KD like this on such a big amount of keywords. I'm only calculating like 10k keywords at a time and it takes hours. And I've got pretty powerful hardware now.

look forward to it!

Torch now can use M1 GPU with torch device is mps, thats exciting for anyone on silicone, generations are so much faster now.

 

[Sartre]

look forward to it!

Torch now can use M1 GPU with torch device is mps, thats exciting for anyone on silicone, generations are so much faster now.

how fast is it compared to a P100 or something like that? Not too fast I assume.

 

[Archemike]

how fast is it compared to a P100 or something like that? Not too fast I assume.

not nearly as fast maybe 20% . How is P100 vs RTX 3090? What do you think about running cross just a bounch of google collabs/kaggles/ or other gpu python instances int he cloud? Some have free gpus now.

 

[Sartre]

not nearly as fast maybe 20% . How is P100 vs RTX 3090? What do you think about running cross just a bounch of google collabs/kaggles/ or other gpu python instances int he cloud? Some have free gpus now.

P100 is pretty old maybe 3x slower? P100 - 9 TFLOPS (single precision) - 3568 Cuda Cores

 

[Archemike]

Yeah I see p100 in a lot of g collabs and other jupiters as the listed free gpu

 

[Sartre]

Yeah I see p100 in a lot of g collabs and other jupiters as the listed free gpu

Yeah but the free collabs are really bad, you get disconnections all the time, can't run them in the background, have to play with saving checkpoints all the time, running multiple google accounts. big hassle. They are also shared.

 

[Tiovater]

An inspiring journey here, great work OP!

 

[Excell]

me and my partner built a NLP model technically more similar to old-style NLP like this: https://ai.googleblog.com/2021/12/a-fast-wordpiece-tokenization-system.html than to ML. It's more of a synonimizer than a full rewriter. For white hat sites I'm using a human editor because I don't want to take any risks for websites that are already on Mediavine, but it speeds up the writing process 5-fold. For spam sites I'm just posting whatever it produces. We will see how this works out in the long run.

This is a new spam site started a month ago, so far the most successful launch of an automated site:
View attachment 201969
View attachment 201970
US-centric niche.

Taking 2-4 weeks to improve the code and then going full-scale with these. I just hired another programmer temporarily to help out.

What ad network are you using now for these spammier sites? How did the spammier sites do since back then?

Is the RPM for the spammier sites lower?

 

[nioarm]

sir satre, how do you detect ugc sites using python ? is there a library that can do that? or you collect ugc sites or forums manually and enter it into the database ?

 

[nextlvlig]

For scraping, beautifulsoup text parsing, and requests is cpu or ram more important? And for spacy/nlp stuff is cpu or ram more important? Confused on the difference between i/o and cpu bound.

 

[nreispqwi]

I initially made a setup but was not happy with the results of the scraping (how to get relevant paragraphs) so went back to the start trying semi manually to try and figure out how to automate. I am still kind of stuck here.

I tried the paa question and answer way and paraphrased just the answers, to avoid the problem, but then it looked like a 'paa site'. I would like to have articles which look like 'normal' articles but unsure how to scrape content there.

So currently for me I am making articles by manually going to the serps for main kw phrase, then opening up the top ten. I then take some decent paragraphs from the results by looking through each article manually for relevant stuff and place them into the article template. I repeat the process for a few more paa question searches until it has over 1k words.

Then I run it through pegasus and will manually edit the result so it makes sense. This seems to be a similar process to the semi automated way the other seos use with commercial ai tools.

This produces good articles I am happy to post, rather than the semi coherent stuff the bots came up with doing it totally automated, however it still takes a good chunk of my time and of course the goal is automation.

What are some tips to be able to have a higher degree of accuracy via bot to automate this process of choosing paragraphs that will make sense in the article as a whole as this is the main sticking point currently in terms of automation. There is so much variance in articles and content I am unsure what logic/nlp to use in order to get reliable results in what the bot would pull.

I know I could pull the h2s and 1-3 paragraphs beneath from articles, but still not sure how to parse for coherent content.