[Case study] AI programmatic SEO (post amount & rank correlation)

dohyung97022 · Jul 5, 2024

Purpose of this study
This is a test to check out how AI articles amount affect the CTR and Rankings of SEO.

I have created a AI blogging site and indexed 20000 articles at the start of june.

It is not doing too bad, but the rankings are quite low (40.9).
I think there is a correlation of how much content you post and how high you rank in keywords.
But this is a guess, I want to prove that this is true.

Also, quite a lot of people are reporting that indexing a lot of content leads to google penalizing SEO.

Test Cases
domain 1
5 AI posts per day

domain 2
10 AI posts per day

domain 3
30 AI posts per day

domain 4
50 AI posts per day

domain 5
100 AI posts per day

All articles have the same topic (Gaming).
All the articles have the same website html format and is created equal.

Some may post less, but that is cuz my code is shit.

dohyung97022 · Jul 5, 2024

Cases where people get penalized by google

Happy1221 said:
Can you please give a suggestion on this?

I started a website in the last week of January this year on an expired domain with some backlinks. I published about 200 AI articles in about 2 weeks to test if those rank. I was getting a few impressions. Some articles were ranking in the top 1-3 pages of the SERP. Later this happened.

praetserge said:
had absolutely the same when I decided to run a test on an expired domain.
First... impressions and clicks then nothing and eventually only like 10 impressions.

@Jaordas , have you had any problems with AI content? Still ranking as before? (I'm talking about your site from the other thread).

Jaordas said:
Bro, imma be honest with you, everyone and their mother wants to know what to do about this right now, ive seen over 50 different posts on this same exact thing
my site number 2 also ended up like this:

pepeneo · Jul 5, 2024

if you did this a couple of years ago you would be ranking and banking, but now as everybody spams ai articles the site will get penalized or deindexed after a couple of months

Fang Yu · Jul 5, 2024

This niche is competitive tbh

Jaordas · Jul 5, 2024

dohyung97022 said:
Purpose of this study
This is a test to check out how AI articles amount affect the CTR and Rankings of SEO.

I have created a AI blogging site and indexed 20000 articles at the start of june.
View attachment 357902

Also, quite a lot of people are reporting that indexing a lot of content leads to google penalizing SEO.

Wishing you Luck in the test.

questions: Are all articles aimed at keywords that should have a minimum number of searches (as seen from a tool) or are a lot of them just random with no particular ranking goal per article?
Im one of the people you used as an example of site algo penalized after heavy bulk publishing. It does not occur all the time but It is hard to know what exact mix of things on a domain can prevent it from getting picked up.

Here is one of my more recent algo penalty ones.

I still do these tests every now and then even though im no longer as active on BHW, because you can still take the data of all the keywords that were easiest to rank on your spam site and then make articles for it on your less spammy sites.

I hope you keep updating this post however, I want to see the results, also, if possible may you ad them to bing search console aswell to see how bing acts towards the posts for the 5 different domains.

dohyung97022 · Jul 6, 2024

pepeneo said:
if you did this a couple of years ago you would be ranking and banking, but now as everybody spams ai articles the site will get penalized or deindexed after a couple of months

Fang Yu said:
This niche is competitive tbh

Wish that was not the case.
Although I cannot say for sure, cuz my sites are relevantly new. So they may be penalized any day now.
This test may prove your point, or it may prove that it is a misconception. or it may give us a golden ratio.
Either way, it is a win.

Jaordas said:
questions: Are all articles aimed at keywords that should have a minimum number of searches (as seen from a tool) or are a lot of them just random with no particular ranking goal per article?

The keywords does not have a minimum number of searches. Pure random.
My AI script copies human written articles and replicates them from the web.
So, it does have a category (Gaming) but does not have aimed targeting for low rank keywords.

Jaordas said:
I still do these tests every now and then even though im no longer as active on BHW, because you can still take the data of all the keywords that were easiest to rank on your spam site and then make articles for it on your less spammy sites.

Btw, this is a great idea.
Check if low authority sites can rank on low rank keywords.
Replicate that on the high rank site.
Wanna try this one day.

Jaordas said:
if possible may you ad them to bing search console aswell to see how bing acts towards the posts for the 5 different domains.

On it right away.

dohyung97022 · Jul 11, 2024

Test Report 1

domain 1 (5 AI posts per day)

domain 2 (10 AI posts per day)

domain 3 (30 AI posts per day)

domain 4 (50 AI posts per day)

domain 5 (100 AI posts per day)

Flaws in my test

Keyword Difficulty
The main basis of this test is flawed.
The search position is decided with keyword difficulty.
My blogs are creating random AI articles that will rank in a random keyword.
I should have created all the AI articles with the same difficulty.

Some are lucky, some are not
If you see domain 4. The CTR is 11%.
That was because the AI has hit a lucky keyword.
Some "horror game free download" keyword was getting a lot of clicks.
This is pure random and is affecting the test results.

So read this test with a grain of salt.
But I will keep doing this test because I still want to know if they get banned.

Serpnest · Jul 11, 2024

Interesting! I'd like to see how Google responds to AI and programmatic SEO. I'll be following this.

azmi mhmed · Jul 13, 2024

Interested

dohyung97022 · Jul 18, 2024

Test Report 2

I have added 15 additional domains for this test.
So every set will have 3 domains each.

SET 1 - 5 AI articles per day

SET 2 - 10 AI articles per day

SET 3 - 30 AI articles per day

SET 4 - 50 AI articles per day

SET 5 - 100 AI articles per day

Thoughts

The additional 2 articles per every set would need some time to get running.
Although I think it is worth to add in more to the testset in order to make solid conclusions.

This test will take quite a lot of time.
I am not doing any backlinks, or keyword research on these domains.
but I am planning to do testing those topics also.

KingMasterEg · Jul 18, 2024

Nice Test , all the domains are new ?

dohyung97022 · Jul 18, 2024

king master said:
Nice Test , all the domains are new ?

Yeah. All the domains are new and bought cheap like .blog, .site .org

4440 · Jul 18, 2024

I wonder what's the cost per article in your setup?

dohyung97022 · Jul 18, 2024

4440 said:
I wonder what's the cost per article in your setup?

Great question mate.

I actually run my own AI model.
If I use vendor api, the bills go up very quickly, very fast.

I borrow RTX 3060 servers (costs 1.5$ per day) in vast.ai.
1 server can create 1500 articles per day. (about 1 per min)

So the cost would be 1.5$ / 1500 = 0.001$ per article.

4440 · Jul 18, 2024

dohyung97022 said:
Great question mate.

I actually run my own AI model.
If I use vendor api, the bills go up very quickly, very fast.

View attachment 361075 View attachment 361076
I borrow RTX 3060 servers (costs 1.5$ per day) in vast.ai.
1 server can create 1500 articles per day. (about 1 per min)

So the cost would be 1.5$ / 1500 = 0.001$ per article.

Thanks for the info.
Would you mind sharing what LLM you use on the 3060?

I don't need your exact model, some rough broad models like llaml 3 70B or Mistral xxB is enough.

AutoVideoFreak · Jul 18, 2024

dohyung97022 said:
So the cost would be 1.5$ / 1500 = 0.001$ per article.

It's the only reason it makes sense and I won't write a negative comment.

4440 · Jul 18, 2024

Jaordas said:
Wishing you Luck in the test.

questions: Are all articles aimed at keywords that should have a minimum number of searches (as seen from a tool) or are a lot of them just random with no particular ranking goal per article?
Im one of the people you used as an example of site algo penalized after heavy bulk publishing. It does not occur all the time but It is hard to know what exact mix of things on a domain can prevent it from getting picked up.

Here is one of my more recent algo penalty ones.

View attachment 357976

I still do these tests every now and then even though im no longer as active on BHW, because you can still take the data of all the keywords that were easiest to rank on your spam site and then make articles for it on your less spammy sites.

I hope you keep updating this post however, I want to see the results, also, if possible may you ad them to bing search console aswell to see how bing acts towards the posts for the 5 different domains.

wonder the cost per article in your project and what llm you are using, if you feel like sharing. thanks

dohyung97022 · Jul 18, 2024

4440 said:
Thanks for the info.
Would you mind sharing what LLM you use on the 3060?

I don't need your exact model, some rough broad models like llaml 3 70B or Mistral xxB is enough.

3060 has 12GB of VRAM.
Llama 3 70B or Mistral is too large of a model and takes too much vram.
Costs would rise drastically if you use heavy models. You would need a A100 for those models.

What I suggest
llama3 7B (generally great)
Gemma 9B (not tested)
phi3 mini (bad but may be great when fine tuned)

How I suggest to finetune
unsloth

4440 · Jul 18, 2024

dohyung97022 said:
3060 has 12GB of VRAM.
Llama 3 70B or Mistral is too large of a model and takes too much vram.
Costs would rise drastically if you use heavy models. You would need a A100 for those models.

What I suggest
llama3 7B (generally great)
Gemma 9B (not tested)
phi3 mini (bad but may be great when fine tuned)

How I suggest to finetune
unsloth

Thanks. I don't know if 3060 is able to run some quantized 14b or above, but the speed would be slow anyway.

The 7,8,9B models are also getting better every day. They may suffice your use case for bulk posting.

Thanks for all the info shared!

Valuable case study.

latneme7e · Jul 20, 2024

dohyung97022 said:
3060 has 12GB of VRAM.
Llama 3 70B or Mistral is too large of a model and takes too much vram.
Costs would rise drastically if you use heavy models. You would need a A100 for those models.

What I suggest
llama3 7B (generally great)
Gemma 9B (not tested)
phi3 mini (bad but may be great when fine tuned)

How I suggest to finetune
unsloth

Are you running your LLM on llama.cpp? I'm very happy with its performance.
When it comes to small models I think Mistral 7b is the best for storytelling. Ofc even it doesn't beat larger models.

I know this is just a case study but I recommend playing around with your prompts to get better articles. Add stuff like "in the style of XXX author" to get articles that don't sound generic.

With high parameter models you can get better quality that most of the clickbait trash articles that take the top spots in google search. It doesn't matter much for initial ranking but it matters for user retention and bounce rate which affect long-term ranking.

Same thing holds for formatting & images that you can also use AI for.

[Case study] AI programmatic SEO (post amount & rank correlation)

Regular Member

Regular Member

Power Member

Elite Member

Power Member

Regular Member

Regular Member

Junior Member

Newbie

Regular Member

Elite Member

Regular Member

Elite Member

Regular Member

Elite Member

Registered Member

Elite Member

Regular Member

Elite Member

Newbie

Main Menu

Marketplace

Making Money

BlackHat World