▶  SEO · S6 E146 · 50:19 · November 23, 2026

AI Automation vs. Content Quality: How to Scale AI Content with Tom Winter

Guest: Tom Winter, Co-founder and Chief Growth Officer of SEOwind
Hosts: Jon Clark & Constance Chen

Play episode
0:00 50:19
AI Automation vs. Content Quality: How to Scale AI Content with Tom Winter

ABOUT THIS EPISODE

Scaling AI content without losing quality starts with copying every step a human expert takes, then letting a system do the steps people skip.

How do you produce AI content at scale without the quality falling apart? Tom Winter, co-founder of SEOwind, says the answer is the opposite of what most teams do. Instead of simplifying the process for AI (“write an outline, write the article”), his team maps the full human workflow, roughly 30 steps with review loops, and automates all of it. He calls it human as a service (HaaS), and he argues that people, not models, are usually where quality breaks: humans skip steps on repetitive work like fact-checking the ~50 claims in a 2,000-word article.

Jon Clark and Constance Chen dig into how Tom runs content for 500+ clients, why giving AI more context doesn’t automatically improve output, how his team tested rewriting content around Anthropic’s new text watermarking without breaking facts or links, and why automotive is the hardest content niche he’s found.

In This Episode

  • What is human as a service, and why does it beat simplified AI prompts?
  • Fact-checking every claim in an article, including downstream impact
  • Removing “black boxes” from AI content workflows
  • Context engineering: how much context does AI actually need per step?
  • Turning non-developers into builders with Hermes Agent
  • AI watermarking, AI detectors, and content refresh vs. rewrite triage
  • Why automotive content is so hard to get factually right

A practical look at how to scale AI content production without trading away accuracy.

The Biggest Takeaway

As Tom emphasizes, the key to successful AI automation is to define what good looks like and make it measurable. Once you establish clear quality standards, you can build, test, and improve AI systems without losing sight of the results that matter.

Key Takeaways

  • Tom Winter, co-founder of SEOwind, argues that AI content quality breaks when teams simplify the process for AI. His human as a service (HaaS) approach copies every step a human expert takes, roughly a 30-step workflow with review loops, and automates all of it.
  • Tom says people are often where quality drops at scale, because humans skip steps on repetitive work. His team asks every week whether a human is still needed in each process, and he says the answer is usually no.
  • In a 2,000-word article, Tom estimates about every second sentence contains a data point, roughly 50 claims to check. His system fact-checks each one and also fixes the downstream impact when a corrected figure is referenced later in the article.
  • Tom Winter works to remove black boxes from AI workflows: data whose source he cannot trace is probably made up, because a model always has to answer. He gives the model data and tools through an API or MCP and uses it only to process them.
  • According to Tom, more context does not mean better output. He splits work into steps and gives the model only the knowledge each step needs, comparing it to answering a Harry Potter question from one paragraph instead of all seven books.
  • About eight months ago, Tom’s team turned non-developers into builders. They use Hermes Agent to check logs, retrieve data and write pull requests in natural language, and developers review the code, because the people closest to a problem build better tools.
  • Tom’s advice for any AI automation is to define what good looks like and make it measurable. With a clear target, he says his team can test 50 approaches instead of five, and some ideas he did not believe in turned out to win.

Connect with Tom Winter

Tom on LinkedIn
SEOwind

Chapters

  • 0:00

    Intro: Tom Winter of SEOwind

  • 2:07

    A Developer’s Approach to SEO

  • 3:39

    What Is Human as a Service (HaaS)?

  • 6:57

    Fact-Checking Every Claim in an Article

  • 9:56

    Non-Developers Building With Hermes Agent

  • 14:33

    AI Watermarking and AI Detectors

  • 20:27

    Content Refresh vs. Rewrite Triage

  • 22:36

    How Much Context Does AI Need?

  • 24:38

    Page Structure and SERP Scraping

  • 29:23

    Hiring Generalists and Editors

  • 33:41

    Conference Takes on AI Hallucinations

  • 36:45

    Testing Agentic Tools for Predictability

  • 38:16

    Why Automotive Is the Hardest Niche

  • 40:10

    Client Fit and AI Disclosure

  • 43:43

    Define What Good Looks Like

  • 47:10

    Rapid Fire

Read the full transcript

Transcript of S6 E146 of the Page 2 Podcast, “AI Automation vs. Content Quality: How to Scale AI Content with Tom Winter.” Lightly edited for readability. Click a timestamp to play from that point. Speaker names are not available for this transcript.

Marketers are under enormous pressure to produce more content, move faster, and do it all with fewer resources. AI promises to make that possible, but there’s a catch. Scaling content production is easy. Scaling quality is an entirely different challenge. Our guest today is Tom Winter, co-founder of SEOwind, a serial SaaS entrepreneur and developer who spent years building AI-powered systems designed to solve exactly that problem. He doesn’t believe humans necessarily belong in every step of the content production process. In fact, he argues that humans are often the reason quality breaks down at scale. His approach starts with something he calls human as a service, the idea that before you automate anything, you need to understand every step a human expert would take to do the job properly. And he’s putting that philosophy to work across more than 500 clients, building systems that can check individual claims, preserve context, and refresh content without sacrificing accuracy. We wanted to understand what that actually looks like in practice. We get into why Tom believes most AI content is fundamentally flawed.

How his team approaches hallucinations and AI watermarking, and why giving AI more context doesn’t necessarily produce better results. We’re wrestling with many of these same questions at Moving Traffic Media. How much should we automate? Where does human expertise create the most value? And how do we make sure efficiency doesn’t come at the expense of the work we’re actually trying to improve? If you’re thinking through some of these same challenges, I think you’ll get a lot out of this conversation. The Page 2 Podcast is brought to you by Moving Traffic Media.

A full-service AEO agency specializing in SEO, answer engine optimization, and performance media. Learn more at movingtrafficmedia.com. All right, let’s get into it.

Welcome to another exciting episode of the Page 2 Podcast. I’m your host, Jon Clark, and I’m joined today by Constance Chen, the search director at Moving Traffic Media. Today we’re excited to welcome the co-founder of SEOwind and a serial SaaS entrepreneur who has spent years figuring out how to scale organic traffic using AI without sacrificing the quality, something very difficult. We’re excited to dig into all things automation today. Tom Winter, welcome to the show.

Hey Jon, hey Constance, really glad to be here and looking forward to the episode.

Awesome. In doing the research for this episode, there was multiple instances where you sort of self-described yourself as a developer, you know, marketer, even a salesperson. I thought what was really interesting was that you never use the SEO acronym. So just maybe talk about a little bit how you see that particular skill of fitting into your repertoire of interests and and focus.

That’s interesting. I didn’t notice that. Maybe I need to fix it. But definitely I’m a developer. Like so I think numbers, and if the numbers and data don’t show me something, for me that means it’s not true. In marketing, like when you’re working in marketing, being a developer helps a lot because I can automate most of the things and I can look at data, but to be honest, I’m not a fan of automating for the sake of automation because it makes no sense. If you don’t have a good process, you shouldn’t automate things. So for me, being a developer in this niche, especially like if I’m speaking at many conferences, helps me to understand if somebody talks bullshit or not. Because I actually can define like how it’s done or how it works. Of course, we can say everything how AI works, but definitely there’s a lot of nuances in how code works. So that helps me out to be better.

That’s a really interesting I guess take and attitude. I think this is a question that comes up quite often for our own team since we also like to build tools of our own. And there’s so much that you could build and do and in terms of automation. So how do you and your team decide what gets to be automated? And can you kind of describe that process to us?

Sure thing. So basically we start with our own things that we do on daily basis. If something is repetitive and we see that we can automate, definitely we are looking into automating it. But what we always start with, and I would encourage every single person to start with, something that I call HaaS. So we know all SaaS, which is software as a service. HaaS for me is human as a service. So basically we check how a human does it and we copy exactly the process of a human, like if there’s four steps, we do all four steps. What a human would take. Because what normally people do with AI, they go to AI and they simplify the process, thinking that AI can do the process for them, but it doesn’t work this way. In my opinion, like if we have a certain process and it works for us, we have a proven track record of getting the value that we want. Why are we skipping steps? So instead of writing an article by going to ChatGPT asking, write me an outline, write me an article based on this outline. Let’s do actually the whole research. Let’s do every single step that we do. We don’t do it as a human in one step or two steps. It’s actually a 30-step process that we go through, we review it and review it again, and so on and so forth. So I strongly encourage people to build systems to automate the things that we normally do as humans.

I wrote I wrote that down. HaaS, human as a service. That’s good. Can you dig into that a little bit? So we use a lot of SOPs, I assume, you know, pretty common for agencies or it should be. So when you’re thinking about that human element, is it sort of thinking through it, right? Is it as simple as taking each individual SOP and sort of breaking them down into an individual skill or an individual step in that process? And in doing so, right, you keep the human in the loop, as it were, because you’re you’re not skipping steps. You’re basically just building on the process that you have.

Yes, there is like this term like keeping the human in the loop. Like I don’t like it to be honest in most cases because I feel that humans are not good at repetitive tasks. And in most cases they’re I ask myself like in my team every single week, do we still need a human in this process? And most of the time the answer unfortunately is no we don’t, because the human can’t keep up with the process at the level and they start to skip steps, and skipping steps means lowering the quality. And we want the opposite. And in most processes, we copy exactly what a human would do. Of course, you have SOPs, but SOPs live on paper until you do them. And unfortunately, in most agencies that I worked with, there is an SOP and there is the way people work with an SOP. So they start, which is a problem, and it shouldn’t be this way.

So for me, when it comes to an AI system, and I I would distinguish AI system from generative AI that we’re using. So basically, if you’re going to ChatGPT, for me, this is a non-repetitive process because every single time, you prompt it different way. I like systems that are predictable. So when you have a predictable system, it doesn’t skip the process. It goes like through every single step that you want. And at some point, if you have the process which works good in an AI system, you can scale the process to make it impossible to repeat for a human. I’ll give you an example. One of the parts of the system that we have when we are writing content is fact-checking. For me, when I’m fact-checking the article, because people talk about hallucinations, I’m taking the article, let’s go with 2000-word article. I’m first highlighting every single sentence that we need to fact-check. For me, in a 2000-word article, that means every second second sentence like needs to be checked because there’s some kind of a data point. Then I’m starting to research every single one of them to say, like, yeah, it’s proven, it’s not, I need to find another way, like to rephrase it to get the data right. And if I correct something, then I’m checking for downstream impact because I if I have a data point and then I refer to this data point later on in the article, if I just correct the data point, that means like I’m not correcting the downstream impact.

And if it’s every second sentence, that means 50 phrases that I need to check. Which human would do that? I wouldn’t. After the third article, I would kill myself. I would quit the job. No way I would do that for the whole day.

Yeah. That makes sense. I I’ve heard you talk a little bit about using APIs more for, you know, your process and I don’t know, maybe even connecting to systems. Is there a difference maybe from working directly within, I know, Claude versus using the API? Are you seeing differences between those two types of connections?

Definitely I would always use Claude with API or MCP, however like you want. Like it there is no difference like from my perspective, like if you’re using API or MCP. But Claude doesn’t have the tools. It’s a general tool, like just like ChatGPT and all others. If you don’t provide them with the tool set that they need to resolve a certain specific problem, let’s go with SEO. It doesn’t know about keywords with high traffic volume.

So I’ve talked with many customers and I saw their prompts as like, provide me with keywords with the highest traffic volume. Like how it will tell you that, but where did it take it from? Who knows? Because it always has to answer. So what I’m trying to do whenever I’m building any kind of a system is to remove black boxes. Black box means like something that I don’t know where it came from. So if I don’t know where it came from, that means it’s a black box and it’s probably made up.

Why? Because AI needs to answer every single time that it’s questioned. So I’m trying to remove these black boxes. I want to know where it took the data from. Because when I’m using AI, I’m using it just like we use our brains. We don’t know everything. So what we do? We retrieve the data from somewhere and we use our brain to process it. And then we spit out something. So if we’re writing an article, we do the research, we find case studies, we find, I don’t know.

data from Wikipedia or whatever, we take it all in, we digest it, and we create an article. Simplified. We don’t write it from top of our head because we would hallucinate all over the place. So the same goes with AI. I use it as a processing engine just to process the data, but in a way that I want.

So I’ve also seen you post about working with Hermes Agent, and you said that your team has built some specialized agent profiles on top of that. Many of the agents have their own tools or different systems. What does each of those profiles look like? What is their job? How do you I guess implement them or use them?

So I’m a big fan of Hermes Agent. It’s one of the harnesses that you can get. And harness means basically it’s a system for using generative AI inside, but in a very specific way that is designed for a specific task. And Hermes Agent is one of my favorite agents because it’s very customizable and it learns with you, basically itself.

create skills based on the whole conversations, all the conversations that you’re having. So coming back to your question, a few months ago, I think about eight months ago, we decided that we need to change every single person’s position inside the company to become a developer. So all non-developers became developers, and not by learning how to code, but by learning how to interact with our system and with everything that we do.

And they’re doing it with a natural language, just like we’re talking right now. And they’re using Harness, which is in their case Hermes Agent, to help to first of all check everything inside the system, check all the logs, retrieve the data that they need to make decisions to see exactly what happened and why it happened, and then to write pull requests and to write code that would fix a certain problem.

And you don’t have to be a developer to do that. Actually, to be honest, non-developers that are working with our system are better than most of the developers working with our system. Because they’re hands-on the problem. It’s like if you’re working in a marketing agency, what you need to do to create a tool for yourself, you need to define it and make a brief for something. Then you need to go to the developer and play the game of echo. So basically, you’re telling something.

And they’re trying to understand what you just said. So this already will mess up the whole process because they don’t live, like they don’t do the things that you do. So then they build it, and within two weeks they come back to you and say, like, okay, I built exactly what you wanted. But it’s not the same thing. They tried to understand your problem, but they didn’t understand it fully because they don’t experience it. So we decided: let’s switch the process, let’s remove the time it takes for the developer to come back with something. Let’s build the things by the people that are non coders. And the developers just are just checking the code. That’s all. So this is how we work. And Hermes Agent definitely helps out with the with these things.

I think we’re we’re we’re finding something similar where you know you you you you sort of describe what’s in your head to a developer and they’re not working, for example, like every day with Google Search Console. So they have to understand like what are these metrics and how do I pull them and how do we make them work together? Just as a basic example. So it’s really interesting you’re sort of I don’t know, reinforcing our thoughts there. We’re it’s probably better for us to just build something initially and then sort of pass it to a developer for Q and A or I don’t know, maybe to make it more efficient.

Because you have to yeah, because you have to think about the job of a developer as a translator. So they translate what you think into code that should work like you’re doing. But to be honest, Hermes Agent with Claude Code, it can do the translation. Like really, it’s really good and it can do much, much more tests than a normal developer could have done like one year ago.

So in in the Hermes Agent example, if it’s sort of self-learning and I can’t remember how many clients you’ve said you’ve worked with. How do you keep the learning independent of each client? Does that make sense? So they’re not sort of, I don’t know, learning brand requirements from one client to the next. Are you keeping them in sort of like separate I don’t know, client folders or or projects, right? How do we how do you keep that sort of learning independent of of one another?

Sure, thing. So it’s over 500 different and clients that we work with. But this is the problem of building a knowledge base for each and separate client. So we can retrieve the data at a specific stage for anything that we’re doing for this specific client. And you have to keep the system very general, I would say, not niche specific.

But at specific stages you have to retrieve the data that you need to make specific decisions in the process. So basically creating a separate knowledge base for every single one of them.

I one question I have is as I know SEOwind, one of your I think features is that you scale up content production quite quickly. And watermarking was, you know, announced I think pretty recently with Claude. And I saw that you had posted that SEOwind has worked on a way to reduce watermarking, but also you were able to rewrite a lot of the articles while preserving facts and the quality of the content.

So walk us through that process where your team, I imagine after that announcement, you had to strategize, you know, what to do about watermarking and thinking about this issue and then how you were able to create this process of how SEOwind works and it is able to bypass that.

Yes, that’s a great question. It was a fun project that we did on the side. First thing that I want to say, I don’t have anything against watermarking. Like in my opinion, I would keep it like as is because I don’t see any problem in the watermark inside the text that is created or enhanced by AI. Because we have to remember that if we use AI in any part of the process, if I as a human I write an article and I just use AI to go for the article and fix it like a last have a last glance. If it will rewrite it, it will add a watermark inside the text. It will not be distinguishable from a text that is fully written by AI. Both will have a watermark. But there is a problem on the market. Everybody is using AI detectors. I don’t believe that it’s any problem that people are using AI. Google, in my opinion, doesn’t see any problem in AI written content, they have a problem in AI slop, which is 99% of the texts that are created with AI. But it’s the same with humans. If you create a like basically spam, like you will not be indexed and you’ll be removed from Google. That’s normal. When Anthropic announced about two months ago the watermarking, they said like we’re adding a watermark to every single model that we have. At the moment I think they finished rolling out of this like with all the available models. But the problem that there is, there is no way to check if the watermark is there or not because they didn’t release API that will allow us to check the watermarking. So nobody knows exactly how they’re doing it or what is the watermark there or not. Probably it is because they announced it. But we’re waiting till second of December, I think, at least, because this is the date where they have to release, I think, they have to release the API and we all will check if removing the watermark works or not. So, because of that, that there is no API, we had to decide what to do to find a way to remove the watermark without knowing if we’re removing it or not. So the best way was actually create our own SynthID watermarking tool to watermark a test, then to be able to identify if we’re able to remove the watermark that we created.

Or not. So we created like our own watermarking system just for the sake of finding out are we able to remove it or not. But the problem with removing the watermark, it’s not only about rotating words inside the article, because most of the tools that I checked that are saying that they’re removing watermark, maybe they are removing watermark and they’re quite good with that, but they’re creating a lot of errors inside the article because they’re either removed like data.

Or they’re changing the data, they’re changing the article is structured. And to be honest, like it’s not valuable if they’re, for example, breaking half of the links inside the article. So we had to find a way to first find all the data points that we have inside the article and store them to see like if we’re keeping it them or not, and then find a way to rewrite the article.

To be honest, using a different model because I wrote it and we’re doing it with DeepSeek. So probably we’re not removing the watermark, but we’re changing the watermark from one to the other. But in the whole process, the key part is not to change the meaning of the article, not to break anything, and to keep both SEO and all the facts intact. So this was the whole problem that we had.

But we were able to do it and we didn’t release it for clients. Like for now, we’re still waiting for API to be released just to check fully if we’re doing that, if we’re removing the watermark or not. And then like to check it, we had to train our own version AI detector to be able to check if we’re removing or not. So that’s the cool part of being a developer that you can do things like that.

Yeah, definitely. I I was curious what your like do you have an opinion on AI detection tools? Cause I mean, for example, we’ve taken what we know is a human edited article and it’ll come back, you know, 90% AI. Like, is there value in these tools? Are they easily gamed? Like, what’s your do you have a threshold when you create an article and you run it through, I don’t know, maybe one of your own internal tools? Like, how do you think about these?

Actually like when I was testing them a couple of months ago, I feel that they got a lot, a lot better than they were a year ago. And the granularity of the text that you need to have to actually verify if there is if it’s AI written or not, got like a lot smaller. So for me at the moment it’s about one sentence level of tokens, like and they can verify.

In my opinion, Google can verify that. To be honest, probably they can even verify which model you used, or at least did you use Anthropic or OpenAI? But should we game it? I don’t think so. It’s really difficult to game it. I can by rewriting every single sentence like separately, and this is the only way I found works for me.

But at this level, if I write every single sentence separately, I will lose context. And it will look like a Frankenstein. So whenever I’m trying to game something, my most important approach is not to lose quality. If I’m losing quality, that means for me it makes no sense.

And I think that’s a really great you know segue into the next question. When when you have such a strong focus on quality, but you also have to scale up your content, like for example, how does your team manage these large content refreshes? Especially because we know freshness can be you know a big ranking factor with LLMs and search. Do you have a system for that and in terms of you know how you maintain freshness, especially after the initial publication of a lot of these articles.

we have a way to refresh the content, or actually we have even like further, because first we start with a triage, like just to check if something is meant for refreshness or rewrite. For us it’s a little bit different because refresh means like I will just like update a couple of things and polish it and it will be okay. Rewrite means like okay, I’m taking the most important parts from the old article, but to be honest, 80% like needs to be thrown out and we need to totally rewrite it, like find new data points, get new conclusions and so on and so forth. So but we automate the whole process. It takes a lot of time in the system, like about probably 15-20 minutes to go through the whole process. But we’re doing it automatically by approaching it again, exactly how we would do it as a human at scale, just to check all the things that we want about the article.

Verify them and then only then publish.

Quick word about the team behind this show. The Page 2 Podcast is produced by Moving Traffic Media, a full-service AEO agency that specializes in SEO, answer engine optimization, and performance media. In plain English, we help brands show up in traditional search and in AI answers when people ask ChatGPT, Claude, Perplexity, or Google about them. And we back that up with everything around it, from paid media and social to web development built for how AI agents actually browse. Joe DeVita and I founded Moving Traffic Media and we’re headquartered in New York.

with a second office in Miami. So if your brand isn’t showing up and AI answers the way it should, that’s what we do every day. You can find us at movingtrafficmedia.com. Okay, back to the show.

I was listening to the Digital Niche Agency podcast that you were on. It was a great conversation, but there was there was a part of it where you were talking about sort of like the context that goes into creating like a 2000-word article and you mentioned things like maybe starting with 150,000 words of context and data research and statistics. So I was wondering if you could sort of take us into what are all the inputs that you’re looking to I don’t know collect when you’re trying to create a 2000 or maybe 3000-word article? Is it as simple as just those things or are there a lot of other things that you’re thinking about as well?

it’s quite easy and difficult at the same time because like it’s basically again comes to a simple thing that I already said like three times. Do it as a human. Like, just what would you do as a human? What kind of data would you get at each step of the process? And this is exactly what you should do. You shouldn’t overwhelm AI with the amount of context for a single thing. And don’t put too many things into the process to do it at once. Like if if you want to eat an elephant, you have to eat them in pieces. And you have to divide the process into steps, and you have to provide the knowledge that it needs, and only the knowledge that it needs for a specific step to be performed. So if I ask you a question, which is trivial, probably like for Harry Potter fans, like what is the favorite candy of Harry Potter? You wouldn’t answer it quickly if I would provide you with seven books of Harry Potter.

You would probably answer it better if I just provided you a small snippet of the part of Harry Potter that talked about this specific thing. And you would probably be far better with answering this question if I just gave you this one paragraph, instead of seven books of Harry Potter. So the same goes with AI. We shouldn’t overwhelm him with data at a specific step.

And this comes to context engineering because you have to provide it with specific things for a specific step. So it knows and it’s used only for processing the data that you provide. Not retrieving on its own, but processing what you provide. And you have to find a way to provide it with this data, which is more complicated, I would say.

I also had a question around there there’s been so much, I don’t know, research and information around like a page’s structure, right? So like how does an LLM ingest content and retrieve it? When you’re creating your content or maybe even a level higher, what are you seeing from just a like a page architecture perspective, right? Like key takeaways or meta description, question answer. How are you thinking about building those things into the page templates that you’re creating or the content that you’re creating? Is that sort of top of mind or are you really just writing or creating content based on the inputs that you’re giving it?

Sure thing, great question. Basically how I think about content and any kind of a page is it has to be valuable for the end user. So for a human. If it’s good for a human, if it’s helpful for a human, valuable for a human, then it means that I’m doing the right work. If FAQ are actually FAQs are actually used, I would add them to the articles. The simplest way to check it is just to go like type in the keyword into Google.

Check top 10 SERP results, check if something is there or not. That means probably within the search intent, the FAQs are working or not. For me, the most valuable thing and something that we only think that humans can do is getting data and driving conclusions from that. So basically that means like mimicking experience. This is the experience because.

Not only are we looking at data points, but we’re trying to give our opinion what it means for us. And this is something that is really, really difficult from the perspective of a person that uses AI, because either it just like shows the data because it retrieves some kind of data, but driving conclusions and building a whole article that looks like an experienced person would write it. This is the difficult part. I think we’re going in this direction.

and normally we say this is the only thing that a human can do. In my opinion, not at the scale, because there’s too many data points that we can take into account, and there’s too many stages that we can build and loop everything so many times that a human just can’t do. But you have to get the knowledge and then drive conclusions.

I think one thing that’s is not it’s it’s very undervalued is just looking at what Google’s returning today and even clicking into those pages and seeing what the page structure looks like. Have you figured out an automation to or an automated way to do that? Like using, I don’t know, SerpApi to pull down titles and descriptions and then I don’t know, maybe going and visiting those pages independently, or sort of how to how does that piece work?

Definitely we’re trying to do this right now, but looking at today, Google poisoned the results. So they’re trying to provide to any SERP scraping mechanism, like results that are different from the ones that we see as humans, which becomes really tricky to go around until somebody will resolve it. I’m not building my own tool for fetching the data from Google because it will be just too much work.

So I’m using API of different providers. For example, as you mentioned, like SerpApi or DataForSEO and many more. But if we are not looking at this, that the data is poisoned because we have to live with something, we’re fetching the data from each and every single page in top 10 SERP results just to be able to verify what they’re saying, not to steal from them. Like we’re want to understand, as a human would do.

What is the search intent? We want to understand what are the key values that we should focus on. And this is the thing that we want to get out of it. But scraping part, it’s very, very difficult. And because you shouldn’t use AI for scraping because it’s not meant to do that. And most of the time, if you’re using going to Claude or ChatGPT asking, like, can you take this data from this page? You never know what they retrieved.

You trust that they went there. Even if they went there, you trust that they rendered the page, you trust that they actually took the data that you wanted from there. But there is a lot of ifs. Most of the time because they’re lazy, they would just verify what’s on the page by looking at the slide. And you don’t know if they did something or not. Even if they went there, they were they might scrape like 100,000 tokens of HTML, where there’s 10% of text. So again, it’s overwhelming for the model to retrieve the data. Or it might be JavaScript rendered. And they weren’t able to retrieve it. So understanding what kind of data are you adding to the process is the key because this is a black box that most of the people are not aware of. They’re asking AI to do something, but they don’t know.

if they did it or not. They trust that it did. I don’t trust because that means a lot of trouble for me.

Yeah, the we we have a couple episodes. One with Alex Moss. We talked about, before we started recording, is a lot about rendering and some of those challenges there. So definitely encourage people to listen to that one. I wanted to ask a question about hiring, or I don’t know, maybe even training, right? So you mentioned earlier around giving non-developers like more access to tools.

as you’re thinking about copywriters, there’s another podcast I was listening to where you said something to the effect of great copywriters are always going to have a job. And so how are you thinking about hiring from the perspective of AI? Are you hiring less? Are you interviewing differently? Like how how has your experience building these automations and sort of making things more efficient?

changed how you’re thinking about hiring over the last couple of years.

This is a difficult one. Because hiring right now, like it’s really, really difficult in our space. I had a bad to be honest, it wasn’t a bad hire like four months ago. We hired a developer. I I love the guy, like he’s an amazing person. And as you mentioned, like I’m a serial SaaS founder. I fired many people.

in my life, because I had many bad hires. But I had to let go this guy. He’s still working with us, but like he will not because I can’t teach him how to work in this process. He’s an amazing person. I love the guy. But he can’t up keep up with the speed that we’re working because so much changed in the development that I can’t teach him and I don’t understand.

what kind of things I need to teach them to be able to keep up with the whole thing. That’s why we’re switching into working with non-developers because we feel that they can understand the goal much, much better. They’re closer to the business. They understand exactly what’s not working. The only problem is the language. So Python. They have a problem with coding, but we can solve it. That’s not a problem.

So for me, we’re transitioning into hiring more people that are generalists, that understand how business works, that understand what goal is, and they can write a brief. Basically, can they can define what they want and how it has to look like. And this is the way we’re trying to solve it. So the closer the person is to the business.

And the better they understand the business, they understand the needs of your business or clients, I think these are the people that will be needed. And when it comes to the copywriters, I feel they can be great editors. Is copywriting still important? I don’t think so. There’s like AI systems, not generative AI, that can do it far better than I do. Look, I’m thinking about it from my perspective.

And I don’t feel there is one thing at scale that I’m better than a good AI system. I’m not able to compete with it, like at certain level. I’m good with judgment, I’m good with organizing the work, but am I better with coding? I’m not able to like compete with that.

I wanted to ask, we we we talk a lot about copywriters and copywriting on this on this podcast, but we rarely talk about editors. And arguably, like you said, it’s becoming the more important piece of the whole content production process. I guess what what skills make a good editor to you? Cause I guess it’s it’s different than what it used to be, right?

I think a good editor is an amazing journalist. So basically, if I would be a copywriter, I would learn how to interview the subject matter experts. So and try to get as much knowledge from them as possible. Because this is the value that we can add to the article. So it all happens before you even start writing. You have to get the knowledge that is not on the internet. You have to go to the client.

You have to go to your own company, interview your own CEO, CTO, whoever has the knowledge that is so amazing, opinions, whatever, gather it and provide it to the later stage of the process. And then when you get the article, maybe it’s good to have a look as a human. Is it needed? In my opinion, many times no. But if you want to make it better and you feel that you can add value on top, definitely do it.

And I think speaking of learning from experts, Jon and I just came back from a conference this past weekend. And we we see that you also attend quite a number of conferences. We saw that you attended one in Athens recently as well as brightonSEO. So it seems like you know one of the benefits of going through these conferences is being able to get together with those in the industry and kind of share what people are working on, especially thinking about in terms of solutions and you know different ways that they’re testing their theories with how how fast AI is changing. So, what’s one idea that you took home from either Athens or brightonSEO that really made you think about things or even consider maybe testing out something different?

It’s there’s so many good ideas as well as really bad ideas on the conferences.

but to be honest, for me one of the people like two people that I really like is Marco Giordano. And I forgot her name right now. Lazarina Stoy.

which are close to data science, and they show me how to use data in a way that is really difficult for a human to do. So anything connected to gathering the data and pulling the information from it, using models and old-fashioned machine learning techniques, definitely these these people are amazing. But on the other side, what I mentioned, it’s also you can hear a lot of pretty dangerous stuff on on conferences, in my opinion, especially in today’s world, because a lot of people on conferences are trying to prove that AI is not there. We humans rule and we’re so amazing. So you don’t have to do anything because AI hallucinates all over the place.

and it’s not there. And this starts to be really dangerous because they prove from the stage how AI hallucinates, but in my opinion, when I’m watching it, they’re only proving that it’s operator’s fault. So basically, the operator didn’t know how to use the system. It didn’t provide, for example, they are asking a question can you make a technical audit for SEO for this and that page? And then it retrieves the data, like it tries to do the audit.

But they’re not providing the tools. They’re not checking if they’d use the tools. Why are you asking for deterministic things to AI that is not deterministic? And you’re not providing all the tools that you would to a human. So if a human came to you, would you do the same? You’re just like asking them questions, taking from them computer and every tool that they’re normally using on a daily basis, and they’re asking and then telling them like, look, you can’t do SEO audit of a page. It doesn’t look like that. And this is dangerous because then the whole crowd is rooting and saying, like, yeah, like we’re good. I don’t need to change. I don’t need to learn. And I’m looking at them like, you’ll be the first ones to get fired. Like, try to embrace it. Try to be better. Find a way to provide the value. Otherwise, like you’ll be gone. And this is something that I don’t like on the conferences right now.

And I think you know, that speaks to a lot of just I think the change the the fast changes that are happening constantly. And I think just at this last conference, I think one theme that we’ve noticed is the people talking about obviously the agentic web and and the rise of people using them that I think bot traffic now makes up more than half of traffic. So I think have you tested any of these agentic tools? Like for example Muse I know Muse has been pretty popular.

Grok Bot and and others.

So definitely there are amazing tools like as your concierge for normal human beings. For me, Hermes is the one that I try to use because I understand more what it does. It’s an open source agent that shows me exactly the process of thinking and everything that it went through. And for me, the most important part is predictability of the results. So if I ask it once and it gives me an answer.

I’m actually running it 100 times and seeing if I get the same answer if I run it with the same data. It’s kind of like with a human. If I go to a human expert and I provide him with a data set, let’s go with my page and please audit this page for SEO. I would expect humans to come up with similar ideas, 80% similar ideas, 10 times that I would ask with the same data set. When it comes to AI, I expect the same.

I don’t want it to answer a different with a different answer every single time. If there is no predictability of results, that means I’m gambling.

Yeah, that makes a lot of sense. I wonder the I mean, you arguably have probably the most experience working with sort of automation in terms of content and working with all sorts of industries. In fact, I think you have most industries listed on your website. Is there an industry that is more like more difficult than another? Like I or maybe to ask it differently, is there an industry where this is much, much harder than say, I don’t know, SaaS or e-commerce? Have you found across industries that there’s some discrepancies in terms of how you can apply content? At this scale?

So for me, the hardest and my beloved industry, I love this industry, is automotive. Like this is the hardest one that can be on the planet. I don’t I haven’t found a harder one. Why? Because models of cars change every six months, more or less. The data is all over the place, about the same cars, like you can’t distinguish the model from 2026, from 2027. And there’s multiple pages that are trying to write about it.

and they make up data. Like so basically they provide data that is made up, like but because they use AI in the most ridiculous way possible. And there is a lot of made-up data. And it’s about horsepower, fuel economy, interior design, and everything else. So we had to build like a whole system both for retrieving the data and fact-checking to be able to do it with automotive. And we knew that we did it really, really good once we saw okay, we’re 90 something percent accurate when it comes to the data. And it’s so freaking hard to do it in automotive. Because as I said, the the the brands are all over the place. Sometimes we even were able to find on automotive manufacturer pages that they have mistakes there because they had like different data on different pages.

About the same thing, which is really something that is funny from my perspective, but this is the hardest niche that I found.

I guess that makes sense because there’s so many different combinations, right? You can go in and customize your own car. And so there isn’t a standard model. They’re they’re all so different. But actually it raises a good question. So when you, you know, obviously you’re creating content. I assume people want to see growth from that content, right? But there are things outside of just pure content creation that impact that visibility, technical, you know, is it a fully JavaScript website? Things like that. So when you start working with a client, are you doing any sorts of audits like that to say, hey, here’s what we think we can do, but there are limitations, or are there situations where you just have to say, sorry, this isn’t a good fit for us because the foundational pieces just aren’t in a good place.

Good part of my job is that I’m working mostly with agencies where I’m providing white label content creation. So basically, we’re their operations team when it comes to content. If they were working with contractors earlier and they were working at scale and they couldn’t reach a certain quality, because like some of the writers were really amazing, but most of them were like mediocre or really bad. And they couldn’t scale it.

We’re taking it over and we’re able to scale it with keeping the quality or even exceeding the quality. So I don’t have to make these decisions. I don’t have to go into technical SEO because we’re just providing content. Simple as that. Very rarely, about 10% of the time, we’re working directly with the customer, and then exactly we’re going through problems like that. And then we have to work with their development team. Sometimes we’re taking over the development.

Because it’s just easier from our perspective to correct these things on our side. And it’s much, much simpler and faster. So then we have to do that exactly. Many times we said to the clients, look, it will not work out because you broke so many things on the website that no matter what we will do, it will not bring you results. I would get the money because you would pay me, but it makes no sense from your perspective. It’s not worth it from your side.

And I already said it like a couple of times in the last two years, to a couple of clients saying them like no, it makes no sense. I can deliver, I will exceed your expectations probably when it comes to quality. It will not change a bit.

Interesting.

Maybe a follow-up to that. What’s your point of view on agency disclosures to clients around like, hey, we’re using some sort of AI automation for content? I imagine you must come up against this occasionally, right? With the work that you do.

Fortunately, I’m leaving this decision to the agencies. We have NDAs and I’m not telling them any, like I’m not sharing this with their clients. But I would be transparent, like if I was an agency, with the way I’m working. But you have to take into account that most of the agencies they’re using contractors to build content. I said it on one of the conferences that in most cases, these contractors are using AI. The management was just not CC on the email. And this is a problem because there’s no process. And if there is no process, there is no predictability. And people are using their own prompts in really way the AI shouldn’t be used. That means they have these black boxes where they’re retrieving the data and they don’t have control over where the data came from. Which causes even more problems because this data is not fact-checked, it’s hallucinated all over the place. And the client, even the agency, is not aware because they’re trying to close their eyes and think, I don’t know what they’re thinking, to be honest, but trust that somebody is not using AI and they are using AI. But we’re just not aware of it.

At at a number of conferences, you know, they have like a bio page for for the speakers. And I think there was one where you you ended it with sort of like loves to be wrong. And I’ve heard you mention like, you know, challenge me, like ask, you know, debate me, right? Is there anything that has come up where you had to change your mind? You know, through either someone debating a a topic with you or I don’t know, maybe just how quickly this industry is moving, right? Is there anything you’ve come up against that’s forced you to change your mind?

Many times. I love to be wrong. Like I’m one of the people when I’m actually telling people, look, if you will tell me, Tom, great job, what I would do, I would just tap myself on the on the back and just like move on. But when you actually tell me, like, look, you can do it in a better way, I love that because I can become a better person. I know it’s super weird to tell somebody like you’re wrong, but it helps me to build better things.

But from the conferences, most of the people are not able to get like tell this feedback because it doesn’t sound good. It’s not politically correct to do it. So mostly I’m trying to do that with my employees and they know that they can do it. And I really appreciate it. And I also like to test things on my own just to be wrong, because being wrong means for me I’m one step closer to being right.

And one thing that I can tell you as a developer, two years ago, I was kind of predicting the winner. Whenever I was trying to solve a certain thing, I was predicting a winner. Although I was calling it testing, I was picking five solutions that I can test within a week because this is how much time I had to test something. But I had to choose these five things. So I was predicting one of these will be the winner. And at the moment, what I can do with the systems that we have.

I can actually test 50 things. So I can actually test things that I don’t believe in and check if they will be working better than what we have right now. Because AI and AI system, Hermes and Claude Code and everything else, can do that, can pivot 50 times to search for a winner. And this is an actual search because if I can define what good looks like, so if I can measure what I’m aiming at, I’m able to test things that I don’t believe in.

And in the last few months I saw that some of these things actually won. They were better than something that I predicted to win, even if I predicted a couple of solutions that probably these will win.

I I would definitely say one thing that’s been very prevalent is just the volume of tests and data analysis that’s coming out, which is really awesome. That’s cool that you’re able to apply that to your own sort of line thinking and and come up with some ideas. But I I know we’re yeah.

If I can add one thing to that, if you will leave this podcast with one simple sentence that you one thing that you should know is define what good looks like. Easy as that. If you’re able to define what good looks like and measure it, you’re able to test things because you know what you’re aiming for. When I’m going, like when we start working with a client, the first question that I ask them.

Do you know, can you give me a definition of what good content looks like for you? Nobody responds. So I’m telling them look, send me one good copy of content that you sent to a client. I can reverse engineer it. Some of them, like really 10%, will send me something. But if you are able to define what good looks like, you’re able to test, you are able to search for a solution.

Yeah, that’s really good advice. Well, Tom, I know we’re bumping up against our our time with you. Do you have a couple minutes for some rapid fires?

Sure thing All right. This may not be a rapid fire, but I was curious so you’ve scaled to over a hundred countries. You’re also bootstrapped. Which of those two things do you think was more difficult?

I think bootstrapping is more difficult in the sense that you’re alone in that. Like there’s a cult of getting money from investors. I don’t like it because that means being trapped. And I feel I I want to have a smaller cake, but mine and with all the possibilities of pivots that I want.

I would really recommend don’t look for money, do it yourself. If you want to get some money for your business, just sell your products or services. Easy as that.

Warsaw or London?

I prefer Warsaw. I live there like in Warsaw right at the moment. Poland is booming. Like it’s really growing really fast. And it’s the safest, one of the safest places to live in in Europe. So I really love it.

I wanted to ask about content briefs. We probably should have spent some more time on that, but when you’re creating a content brief, do you prefer full narratives or bullet point format?

Both. Like to be honest, I try to get as much data as possible in the content brief because I want to have a buffet of data. I’m even going to a level of defining how do you spell certain things inside the article. So if I’m working with automotive, like I know exactly BYD.

will be spelled with capitals and I’m adding it to the content brief. So because there’s so many mistakes that people do. And definitely spelling is not one of them that I want to do.

And then what is the most underrated tool you would say that free tool that you do you love to use?

I would say Hermes that we talked about, Hermes Agent, start using it on daily basis. This will be your agent that will help you out on daily basis with all the tasks that you can do.

Dur during the research, I actually wrote that down as like a a task for myself for the agency to figure out how we can integrate that a little bit more.

Well, Tom, this has been amazing. Thank you so much for for getting in the weeds with us and talking through your process. Let our listeners know where they can find you online. Are you gonna be speaking anywhere coming up?

Not until the end of the year, but then in March and in April, I’m going to Asia. One is SEO Summit in Saigon. The other one is SEO Summit Phuket in Thailand. Then probably a couple of conferences in Europe. But mainly you can find me on LinkedIn. And please, if you want to talk about anything, just find me on LinkedIn, ask me questions. I want to give back as much as possible to the industry.

the industry helped me so much. So if I can help you out also with things, please let me know. I don’t buy it, I will not charge you if if you just ask some questions and I can can help you out like without spending too much time. Don’t be afraid of me.

Perfect. We’ll definitely link to those in the show notes. But Tom, thanks again for for joining us on the Page 2 Podcast. This has been a lot of fun. For those listening, if you enjoyed the show, please remember to subscribe, rate, and review. We’ll see you next week. Bye bye.

Thanks a lot. Cheers.