Since ChatGPT came out, I have probably created about 10 iterations, at least, of what I think is a proper infrastructure for an AI SEO consultant like myself, because new technology always shows us that there is a different way. As you know, new language models come out all the time, but we also know that their costs tend to go up the more powerful they are. With my development background and being a data nerd by default, I have always had this problem with AI automations: they might work 99 times, and in the 100th run, they might do something completely different. Now, there are different ways to work around this:
Absolutely checking for all kinds of errors.
Checking against a checklist and trying to make sure the system can figure things out by itself.
Something that was already part of some past iterations and is now the most essential part of my latest one: actually analyzing the case and saying what does not need AI at all.
If we are honest about a lot of the things we ask Claude, ChatGPT, or Perplexity to do these days, many of them don’t require an AI. You could just have an app created that does a lot of it, and then inject AI superpowers where it really matters.
What this allows you to do is to actually control a lot of the process, a large portion of it. I compare this to the written SOPs I had when we had a large team of SEO ninjas here at my agency, because a lot of the steps I can automate without ever touching a language model, just by having a tool like Claude Code create an application for that. I’ll give you a few examples:
fetching numbers from Google Analytics
getting lists of keywords a page ranks for
checking Google Drive for the client transcripts from the last 3 months
fetching from Slack the latest client conversations and internal project conversations we had
analyzing competitors’ HTML and comparing it to my client’s HTML
running performance tests
crawling websites, and so on
In an SOP, as we had it before, there might have been a step that said, “Now go to one of the senior SEO consultants and ask them to check your interpretation, or ask them to analyze this particular case based on all the data you have gathered.”
Now, this might be the step where I pick the best, smartest language model I have that can use all the information I have already automatically fetched and try to make sense of it.
The entire process is always going to be the same: the data fetching, the connections to the APIs, figuring out what to do if we are out of credits. All those kinds of things are always going to be the same.
It’s only where it really comes to making sense of some of the data, and it’s not a plain comparison (because even those could be done automatically), that I want a language model.
I also want the language model when it comes to creating a summary paragraph about all the findings, but I don’t need the language model to create a Google Doc. I don’t need it to format it. I don’t need it to put in all the different tables from the data I have fetched. I can do this automatically: no API credits needed, no language model needed.
But then there are the next action steps: this is again a point where I might have a language model suggest something that we then revise.
The reality is that language models are going to get more and more expensive. Right now, I believe we’re in the phase where the big technology platforms let us play with really expensive stuff. That’s going to get very, very expensive very quickly. The more we are able to work efficiently with AI and language models today and recognize where we can get the most leverage, the better prepared we are going to be to quickly adapt to new situations.
And it’s also about control. I don’t want to give up control over a process that I designed and developed, especially when it’s technology-based and I have already lined up all the steps.
Language models have made it very simple to create applications that run on your computer, and it’s even rather simple to host them on a server (where you can switch off your computer and they’re still going to be running).
The realization that you can create custom applications and inject language models only where it matters is not only going to save you tons of money and time, but also going to take out the frustration that comes with LLMs making up stuff here and there at random moments (and forcing you to keep creating more and more tests for potential failure cases, which is never going to end).
I’ve been creating quite a few custom applications for clients lately, very simple ones, very small ones, and have been able to blend my more than 15 years of past developer experience with today’s AI technology in software development. If you have any questions about this or ideas about what you want to do in your business, I’m happy to share my take on it.




Great take here, Gert. I love the idea of injecting AI, your language model, where it actually counts, where you might need to do messy interpretation, or have something try to make sense out of a bunch of data and turn it into prioritized action steps. I'm definitely going to keep your thinking in mind as I approach future problems, because I'm with you where at this point we get so much unbelievable power for pretty cheap with these subscriptions. But if the cost suddenly went up to what we probably would actually need to be paying, it would be exorbitant. So figuring out where to use and leverage AI to inject it is definitely a key thing. Thanks for sharing your insights.