The pipe dream is having LLMs to perform SEO content creation autonomously, and I did just that that and watched my ranking drop off the cliff and my organic traffic halved. This article is written entirely by me without any AI or help.
I'm going to tell you how I used AI and agentically fucked the SEO of nubela.co. Read on and I'll show you the SEO agent harness I built, so you can succeed and avoid the mistake as I made.
NinjaPear's SEO Agentic Dashboard
As a solo founder building NinjaPear to 1M ARR, the one thing I'm trying to do is to replace teams that I used to hire for Proxycurl. In this case, I'm trying to replace an SEO team with an agentic dashboard that does the following:
- Monitor existing keywords that I rank for and identify growth/loss trends. It does so by integrating into Google Search Console and grabbing data from that on a regular internal. The idea is that like the SEO agent that used to work for me when I was running Proxycurl, he'd report to me if a top performing keyword that we rank well has now dipped in ranking, and what we need to do to get it back up.
- Perform keyword research, and identify the gap with keywords that I do not yet rank for, but would be beneficial that I want to rank for.
- Write article plans. See what kind of articles currently rank which usually tell us what the audience clicks into, and perform web research and see how the SEO agent can write an article that is 10x better.
- Approve article plans. This step requires human in the loop AKA me. I will provide my human feedback, and refine the plan until it is good enough. Once it is good enough, I will approve it.
- Flesh out an article plan into a full-fledge article. The agent can write out full articles, or edit existing articles. On top of that, it can generate lead magnet artifacts such as sample code, libraries, like what it has published in our Github, as well as to generate PDF/XLSX documents or infographics. Whatever that is helpful. The goal is to genuinely product a 10x helpful article done with deep research.
- Publish and syndicate articles. As a proof of concept, I made it syndicate to Medium and Substack. I tried doing it for X but it was proving a little too difficult so I didn't invest more time into it.
What I did and repercussions.
I did two things after this agent was deployed.
- The nubela.co that this blog sits on, has a legacy of older articles although most of them are stale and have decayed. So I had the agent revamp the top performing articles with demo, github projects, etc. I think I did too much because after I updated them, their ranking dropped dramatically.
- For a bunch of keywords that the agent proposed that I need to rank for, I had the agent draft, write, publish and syndicate these articles to the blog.
Unfortunately, new articles did not rank, and existing articles that used to rank, stopped ranking.
But did it fail?
To be honest, I don't know. Because last I did something else. I began writing content without AI and engaging with people on Reddit, X and email authentically, and in the past 2 days, suddenly a bunch of these agentically written articles received a lot more traffic. From 300-400 visitors/day to 2.1K visitors a day on the 5th July!
I'm not sure if it is because of people searching for NinjaPear because I'm engaging publicly more and intriguing people, or if it is a matter of ranking improving.
As always, this could be a matter of flawed implementation. I still think there is space for AI generated content; for example, if you google "NinjaPear", you'll see a bunch of NinjaPear VS <company> and I just do not believe these are manually written articles. I believe they were AI generated en mass.
Regardless, I hope you find this helpful!
